diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_tensor([4192, 4192])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_tensor([4192, 4192])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45c978d697da10a5b1c05a26b63a26bd18e6fde1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_tensor([4192, 4192])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera remains still, focused on a pair of calloused, rough-skinned, aged hands intently engaged in a woodcarving process. A warm, directional light from above and side precisely illuminates the work area, casting three-dimensional shadows across the wrinkles of the knuckles and the deep lines of the palm. One hand firmly holds a block of light-colored, fine-grained basswood, while the other grips a carving knife, its handle sleek from years of use and its blade gleaming with a cool, metallic sheen. With steady, deliberate pressure from the wrist, the sharp tip of the knife delicately slices into the wood, shaving off thin, naturally curly chips. These chips hover briefly in the air, like light snowflakes, before slowly drifting down to the workbench, blurred into a blur of light in the background. The entire scene, imbued with a profound tranquility, is a silent celebration of concentration and skill. The warm yellow primary color creates a subtle contrast with the cool glow of the knife's blade, highlighting the power and delicacy of the creative process.", + "seed": 17251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_4192_seed_17251_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": -0.06210144441574812, + "vlm_logprob": -0.06210144441574812, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06210144441574812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06162458285689354 + ], + "opposite_logprob_tokens": [ + -2.936624526977539 + ], + "logprob_reward": -0.06162458285689354, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06162458285689354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04895751178264618, + 0.2618524730205536, + 0.0508420392870903, + 0.013488197699189186, + 0.06862787902355194, + 0.023689117282629013, + 0.009243875741958618, + 0.1957719624042511, + 0.022384563460946083, + 0.3051423728466034 + ], + "credit_map": [ + 0.005555296316742897, + 0.003514678683131933, + 0.006305025890469551, + 0.006007535383105278, + 0.007244812324643135, + 0.007190065458416939, + 0.0076031493954360485, + 0.0055369543842971325, + 0.014598490670323372, + 0.0067337858490645885, + 0.0726824626326561, + 0.01759704202413559, + 0.01129397377371788, + 0.00740240840241313, + 0.09644467383623123, + 0.03509964793920517, + 0.004529948346316814, + 0.004154543858021498, + 0.006466616876423359, + 0.008899198845028877, + 0.006606488022953272, + 0.006749223452061415, + 0.007201103959232569, + 0.006234917789697647, + 0.0012709852308034897, + 0.0010323850437998772, + 0.002071244874969125, + 0.0021082591265439987, + 0.0020686655770987272, + 0.0013594662304967642, + 0.0017928510205820203, + 0.001784340594895184, + 0.003762344364076853, + 0.0015201278729364276, + 0.011632834561169147, + 0.003931454382836819, + 0.0026947029400616884, + 0.0034315078519284725, + 0.02614433877170086, + 0.01551057305186987, + 0.0031620291993021965, + 0.0017689908854663372, + 0.0037081402260810137, + 0.003854389302432537, + 0.002612713724374771, + 0.0027793054468929768, + 0.0032969298772513866, + 0.002506620716303587, + 0.0007497055339626968, + 0.0006454724934883416, + 0.00156496895942837, + 0.0018360407557338476, + 0.0008957760292105377, + 0.001039805356413126, + 0.0014556886162608862, + 0.0010564185213297606, + 0.013330470770597458, + 0.00527934031561017, + 0.037108756601810455, + 0.019857779145240784, + 0.007646895479410887, + 0.006188934203237295, + 0.07386835664510727, + 0.03249143809080124, + 0.0025384356267750263, + 0.0014093080535531044, + 0.002811322920024395, + 0.004401427693665028, + 0.002221093513071537, + 0.002061113715171814, + 0.0036840555258095264, + 0.003257809206843376, + 0.03190004825592041, + 0.0084554273635149, + 0.011201060377061367, + 0.013404282741248608, + 0.007608320564031601, + 0.017406871542334557, + 0.0931205302476883, + 0.1220458522439003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.06162458285689354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03817487135529518 + ], + "opposite_logprob_tokens": [ + -3.288174867630005 + ], + "logprob_reward": -0.03817487135529518, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03817487135529518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043587662279605865, + 0.18031300604343414, + 0.027425864711403847, + 0.014758672565221786, + 0.11482122540473938, + 0.026332704350352287, + 0.018287435173988342, + 0.20383605360984802, + 0.015923887491226196, + 0.35471346974372864 + ], + "credit_map": [ + 0.0051768203265964985, + 0.002705126069486141, + 0.005795627366751432, + 0.004683469422161579, + 0.005098109599202871, + 0.0072733028791844845, + 0.006797960959374905, + 0.0060572498477995396, + 0.010484801605343819, + 0.005621661432087421, + 0.04702319577336311, + 0.012641980312764645, + 0.008196882903575897, + 0.007188120391219854, + 0.06505715847015381, + 0.024099227041006088, + 0.0035065598785877228, + 0.002565541071817279, + 0.003030121559277177, + 0.00382035318762064, + 0.00277476548217237, + 0.0036711820866912603, + 0.004408678505569696, + 0.0036486650351434946, + 0.0015126049984246492, + 0.0014235478593036532, + 0.0021572678815573454, + 0.0022680675610899925, + 0.0020647267811000347, + 0.0013927845284342766, + 0.0022374256514012814, + 0.001702248933725059, + 0.006705485284328461, + 0.0034389058127999306, + 0.02925206534564495, + 0.007495671045035124, + 0.005111716687679291, + 0.005760660395026207, + 0.040259718894958496, + 0.0167970173060894, + 0.0023319816682487726, + 0.0013563716784119606, + 0.0034904845524579287, + 0.006253939587622881, + 0.0031555613968521357, + 0.002617454156279564, + 0.004353934898972511, + 0.0027729778084903955, + 0.0017436135094612837, + 0.001318750437349081, + 0.002619295148178935, + 0.0036031496711075306, + 0.0017881023231893778, + 0.0024522822350263596, + 0.0026675728149712086, + 0.0020946708973497152, + 0.013625905849039555, + 0.00560031458735466, + 0.04091091454029083, + 0.019711459055542946, + 0.008203256875276566, + 0.00715209124609828, + 0.07575813680887222, + 0.03287399560213089, + 0.0014944106806069613, + 0.0009918282739818096, + 0.002207762561738491, + 0.0032184156589210033, + 0.00127732555847615, + 0.0015785275027155876, + 0.0027029230259358883, + 0.0024526952765882015, + 0.03715995326638222, + 0.009858004748821259, + 0.014741073362529278, + 0.015619732439517975, + 0.008931194432079792, + 0.023131761699914932, + 0.10650597512722015, + 0.13876579701900482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.03817487135529518 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11302636563777924 + ], + "opposite_logprob_tokens": [ + -2.2380263805389404 + ], + "logprob_reward": -0.11302636563777924, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11302636563777924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06290417909622192, + 0.39651620388031006, + 0.08281851559877396, + 0.03147345036268234, + 0.19132964313030243, + 0.053651776164770126, + 0.0324823260307312, + 0.08300095051527023, + 0.042288973927497864, + 0.02353396639227867 + ], + "credit_map": [ + 0.00790451094508171, + 0.003917159978300333, + 0.010890345089137554, + 0.009351339191198349, + 0.006080655846744776, + 0.007392627187073231, + 0.008843068033456802, + 0.008524477481842041, + 0.02181961014866829, + 0.009554330259561539, + 0.1121099591255188, + 0.026456259191036224, + 0.017043760046362877, + 0.01129460334777832, + 0.14513786137104034, + 0.0530998520553112, + 0.013356505893170834, + 0.007149684242904186, + 0.01045930664986372, + 0.008741496130824089, + 0.00805955845862627, + 0.01370630320161581, + 0.011624874547123909, + 0.009720788337290287, + 0.003923007752746344, + 0.0020499888341873884, + 0.00492221862077713, + 0.00662051560357213, + 0.002464807126671076, + 0.0022476594895124435, + 0.005814745090901852, + 0.00343050598166883, + 0.012446852400898933, + 0.006550072226673365, + 0.058753252029418945, + 0.012982966378331184, + 0.009679947048425674, + 0.01222089771181345, + 0.05905480310320854, + 0.019640862941741943, + 0.00611794413998723, + 0.004820225294679403, + 0.005364099983125925, + 0.009887968190014362, + 0.005160022526979446, + 0.007911122404038906, + 0.0077043925411999226, + 0.00668600108474493, + 0.0032154901418834925, + 0.002397386357188225, + 0.0043867407366633415, + 0.006478065624833107, + 0.0033157384023070335, + 0.0034311343915760517, + 0.004729814361780882, + 0.004527956247329712, + 0.006417400669306517, + 0.0038797541055828333, + 0.019203657284379005, + 0.007619384676218033, + 0.0044154454953968525, + 0.0032300911843776703, + 0.027878904715180397, + 0.010356312617659569, + 0.006191220134496689, + 0.003261654172092676, + 0.003721656510606408, + 0.00915272906422615, + 0.002838608343154192, + 0.004290455486625433, + 0.006746524944901466, + 0.006086125038564205, + 0.002373587805777788, + 0.0013216804945841432, + 0.0027386266738176346, + 0.0021495679393410683, + 0.002008230658248067, + 0.00487752677872777, + 0.004711555317044258, + 0.0033531920053064823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.11302636563777924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890422984957695 + ], + "opposite_logprob_tokens": [ + -2.4640421867370605 + ], + "logprob_reward": -0.0890422984957695, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890422984957695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02807135321199894, + 0.2993580400943756, + 0.01633644849061966, + 0.017257610335946083, + 0.14124098420143127, + 0.035908181220293045, + 0.013058103621006012, + 0.17795103788375854, + 0.029375864192843437, + 0.24144241213798523 + ], + "credit_map": [ + 0.0034607204142957926, + 0.0016233834903687239, + 0.003390206489712, + 0.0033243002835661173, + 0.0035310990642756224, + 0.003950457088649273, + 0.004107869230210781, + 0.004683318547904491, + 0.016491921618580818, + 0.006882360205054283, + 0.08162820339202881, + 0.020446419715881348, + 0.012346222065389156, + 0.008005823940038681, + 0.11143467575311661, + 0.04212242737412453, + 0.0018662795191630721, + 0.0010685212910175323, + 0.002004694426432252, + 0.0031052236445248127, + 0.001356435241177678, + 0.0022427462972700596, + 0.0024779869709163904, + 0.002214561216533184, + 0.002221957314759493, + 0.0011997520923614502, + 0.0019327654736116529, + 0.0017423804383724928, + 0.0034706369042396545, + 0.001989018404856324, + 0.0020369107369333506, + 0.002664190484210849, + 0.0073416465893387794, + 0.002914462238550186, + 0.029369240626692772, + 0.008469666354358196, + 0.005735116079449654, + 0.008639112114906311, + 0.05239613726735115, + 0.02637559548020363, + 0.004385293461382389, + 0.002780273323878646, + 0.003488838905468583, + 0.004663574974983931, + 0.004858703818172216, + 0.006524918135255575, + 0.004153703339397907, + 0.005052878521382809, + 0.0014395744074136019, + 0.0012755850329995155, + 0.0018623817013576627, + 0.001880783005617559, + 0.0012445012107491493, + 0.0019432571716606617, + 0.0017766287783160806, + 0.0016353920800611377, + 0.013194190338253975, + 0.005135238636285067, + 0.027071520686149597, + 0.01977142132818699, + 0.007220023311674595, + 0.00732520641759038, + 0.06626299768686295, + 0.031970422714948654, + 0.0023807722609490156, + 0.0023642706219106913, + 0.004269844386726618, + 0.006354950834065676, + 0.002698733704164624, + 0.0025212590117007494, + 0.005022492725402117, + 0.0037635406479239464, + 0.026562081649899483, + 0.006416650023311377, + 0.00872352346777916, + 0.011095618829131126, + 0.00654692342504859, + 0.01877029612660408, + 0.06721901148557663, + 0.09610830992460251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.0890422984957695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863910373300314 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.00863910373300314, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863910373300314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07535135746002197, + 0.41227400302886963, + 0.03108823299407959, + 0.025297552347183228, + 0.12388468533754349, + 0.021501116454601288, + 0.019373351708054543, + 0.0907478928565979, + 0.08896977454423904, + 0.11151203513145447 + ], + "credit_map": [ + 0.011133081279695034, + 0.004408308304846287, + 0.009588006883859634, + 0.0098073435947299, + 0.007959776557981968, + 0.010759996250271797, + 0.01121892686933279, + 0.01047592144459486, + 0.022307755425572395, + 0.009432381018996239, + 0.08551479876041412, + 0.03156227990984917, + 0.015374468639492989, + 0.010900279507040977, + 0.1623631864786148, + 0.07481883466243744, + 0.004949487745761871, + 0.001835490926168859, + 0.003808886045590043, + 0.0047329566441476345, + 0.004064705688506365, + 0.003957494627684355, + 0.004393226467072964, + 0.003345984499901533, + 0.003168562427163124, + 0.0019859049934893847, + 0.0037704166024923325, + 0.004052239470183849, + 0.00199159630574286, + 0.0021089287474751472, + 0.004329492803663015, + 0.0038904095999896526, + 0.009847412817180157, + 0.005411264020949602, + 0.03589702025055885, + 0.010625223629176617, + 0.008093442767858505, + 0.012242362834513187, + 0.026293957605957985, + 0.01547401025891304, + 0.002129094675183296, + 0.0015869333874434233, + 0.003571766661480069, + 0.004725087434053421, + 0.002028718125075102, + 0.0020444949623197317, + 0.0031156372278928757, + 0.0022993842139840126, + 0.0014704267960041761, + 0.0013334896648302674, + 0.003031745320186019, + 0.0038388539105653763, + 0.0017585093155503273, + 0.0020792405121028423, + 0.0032311826944351196, + 0.00262990384362638, + 0.00636706268414855, + 0.002199214417487383, + 0.014134402386844158, + 0.009747719392180443, + 0.004144898150116205, + 0.0033657618332654238, + 0.03488139808177948, + 0.01590743660926819, + 0.00800543837249279, + 0.006375479977577925, + 0.010686091147363186, + 0.02244120091199875, + 0.008144709281623363, + 0.008190515451133251, + 0.013301304541528225, + 0.011825035326182842, + 0.006431382615119219, + 0.002399162854999304, + 0.009176854975521564, + 0.0054717897437512875, + 0.004504480864852667, + 0.010026155970990658, + 0.030642280355095863, + 0.04285992681980133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.00863910373300314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on a pair of calloused, rough-skinned, aged hands intently engaged in a woodcarving process. A warm, directional light from above and side precisely illuminates the work area, casting three-dimensional shadows across the wrinkles of the knuckles and the deep lines of the palm. One hand firmly holds a block of light-colored, fine-grained basswood, while the other grips a carving knife, its handle sleek from years of use and its blade gleaming with a cool, metallic sheen. With steady, deliberate pressure from the wrist, the sharp tip of the knife delicately slices into the wood, shaving off thin, naturally curly chips. These chips hover briefly in the air, like light snowflakes, before slowly drifting down to the workbench, blurred into a blur of light in the background. The entire scene, imbued with a profound tranquility, is a silent celebration of concentration and skill. The warm yellow primary color creates a subtle contrast with the cool glow of the knife's blade, highlighting the power and delicacy of the creative process.", + "seed": 17252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_4192_seed_17252_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": -0.039006621763110164, + "vlm_logprob": -0.039006621763110164, + "vlm_yes_no_margin": 3.949999976158142, + "vlm_token_credit": -0.039006621763110164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008878634311258793 + ], + "opposite_logprob_tokens": [ + -5.633878707885742 + ], + "logprob_reward": -0.008878634311258793, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.008878634311258793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23711295425891876, + 0.1253044456243515, + 0.0484650656580925, + 0.08863192796707153, + 0.1547120213508606, + 0.06541290879249573, + 0.04278459772467613, + 0.028995409607887268, + 0.16675586998462677, + 0.04182474687695503 + ], + "credit_map": [ + 0.01990455947816372, + 0.03086313046514988, + 0.020205724984407425, + 0.04095228761434555, + 0.025945881381630898, + 0.02882327139377594, + 0.03542022034525871, + 0.03499787673354149, + 0.013159447349607944, + 0.018946973606944084, + 0.012787425890564919, + 0.02721714787185192, + 0.010822249576449394, + 0.008805104531347752, + 0.018803445622324944, + 0.014762644656002522, + 0.006173224188387394, + 0.005557034630328417, + 0.005994873587042093, + 0.00861102994531393, + 0.005967606790363789, + 0.004158945754170418, + 0.008041786961257458, + 0.003960563801229, + 0.010780408047139645, + 0.010012825019657612, + 0.009888377040624619, + 0.016881387680768967, + 0.008425097912549973, + 0.007386897224932909, + 0.014516144059598446, + 0.010740786790847778, + 0.015259654261171818, + 0.01725480891764164, + 0.019681701436638832, + 0.02970304526388645, + 0.01546710729598999, + 0.011278733611106873, + 0.027536503970623016, + 0.01853046379983425, + 0.006205049809068441, + 0.009198897518217564, + 0.006966230925172567, + 0.00997735932469368, + 0.006210145074874163, + 0.006634162738919258, + 0.012207119725644588, + 0.008013948798179626, + 0.004922066815197468, + 0.00458652013912797, + 0.005051902029663324, + 0.00918867439031601, + 0.003673219121992588, + 0.0044870381243526936, + 0.00596063444390893, + 0.004914543125778437, + 0.0040077888406813145, + 0.0037400659639388323, + 0.002509163459762931, + 0.004575018770992756, + 0.002997436560690403, + 0.003984143026173115, + 0.003878798568621278, + 0.0033029948826879263, + 0.025814266875386238, + 0.023975204676389694, + 0.015326196327805519, + 0.027141042053699493, + 0.016037607565522194, + 0.013430333696305752, + 0.028487378731369972, + 0.01654384285211563, + 0.005346287507563829, + 0.005624355282634497, + 0.004972814116626978, + 0.006391764152795076, + 0.0045450604520738125, + 0.0026924642734229565, + 0.006643742322921753, + 0.005608258303254843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.008878634311258793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020618975162506104 + ], + "opposite_logprob_tokens": [ + -3.8956189155578613 + ], + "logprob_reward": -0.020618975162506104, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020618975162506104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2263544201850891, + 0.13139043748378754, + 0.10915029048919678, + 0.1095978319644928, + 0.0938810333609581, + 0.0897565484046936, + 0.05155361816287041, + 0.019506007432937622, + 0.11962256580591202, + 0.049187321215867996 + ], + "credit_map": [ + 0.0256792064756155, + 0.025757664814591408, + 0.01864852011203766, + 0.03712250292301178, + 0.021928725764155388, + 0.022566204890608788, + 0.04184446483850479, + 0.0328071191906929, + 0.017444806173443794, + 0.016609476879239082, + 0.013935709372162819, + 0.020642701536417007, + 0.013794958591461182, + 0.011248808354139328, + 0.021234584972262383, + 0.0164793748408556, + 0.014716205187141895, + 0.013520943000912666, + 0.012377314269542694, + 0.01664472185075283, + 0.013936663046479225, + 0.011493338271975517, + 0.017680464312434196, + 0.008780626580119133, + 0.013421815820038319, + 0.011855975724756718, + 0.011833778582513332, + 0.020127855241298676, + 0.01066139992326498, + 0.009814576245844364, + 0.015440431423485279, + 0.016441987827420235, + 0.0098916320130229, + 0.010148108005523682, + 0.01193399541079998, + 0.01830124482512474, + 0.009602822363376617, + 0.006840793415904045, + 0.015670171007514, + 0.011492257006466389, + 0.010513144545257092, + 0.010295279324054718, + 0.011254043318331242, + 0.01720716618001461, + 0.009145158343017101, + 0.008570081554353237, + 0.01337180845439434, + 0.009399852715432644, + 0.00543800788000226, + 0.005762497428804636, + 0.006243240088224411, + 0.011001917533576488, + 0.004257748834788799, + 0.005114846862852573, + 0.008394925855100155, + 0.005340426228940487, + 0.002610775176435709, + 0.0029771639965474606, + 0.0015738426009193063, + 0.0023169093765318394, + 0.002996044233441353, + 0.0020484626293182373, + 0.0030748029239475727, + 0.0019080056808888912, + 0.018148081377148628, + 0.01611398719251156, + 0.011760200373828411, + 0.019502488896250725, + 0.012383921071887016, + 0.009896901436150074, + 0.02083817683160305, + 0.01097879745066166, + 0.008213963359594345, + 0.006981258746236563, + 0.004120017867535353, + 0.005774051882326603, + 0.004521217662841082, + 0.0033730079885572195, + 0.00817030854523182, + 0.008033490739762783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.020618975162506104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710295617580414 + ], + "opposite_logprob_tokens": [ + -2.127102851867676 + ], + "logprob_reward": -0.12710295617580414, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12710295617580414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15915735065937042, + 0.12027512490749359, + 0.08266966789960861, + 0.10148017853498459, + 0.16413885354995728, + 0.13032567501068115, + 0.029025159776210785, + 0.030097533017396927, + 0.14610958099365234, + 0.03672083839774132 + ], + "credit_map": [ + 0.017402298748493195, + 0.019154349341988564, + 0.011507993564009666, + 0.02512933500111103, + 0.01932019740343094, + 0.01624710112810135, + 0.018101997673511505, + 0.032294079661369324, + 0.014100204221904278, + 0.014318394474685192, + 0.013152709230780602, + 0.02038622833788395, + 0.01322916243225336, + 0.010853265412151814, + 0.019166195765137672, + 0.015068965032696724, + 0.009571515955030918, + 0.013074172660708427, + 0.006488293409347534, + 0.012157195247709751, + 0.009640342555940151, + 0.014728189446032047, + 0.008032621815800667, + 0.008977340534329414, + 0.012905781157314777, + 0.011408082209527493, + 0.011281497776508331, + 0.017738578841090202, + 0.010935618542134762, + 0.010549117811024189, + 0.011672012507915497, + 0.01498948410153389, + 0.019461452960968018, + 0.017549872398376465, + 0.02158176153898239, + 0.03166857361793518, + 0.015956318005919456, + 0.012769068591296673, + 0.026327695697546005, + 0.018824098631739616, + 0.011271917261183262, + 0.016846004873514175, + 0.015510674566030502, + 0.023278137668967247, + 0.012879022397100925, + 0.011842358857393265, + 0.023252593353390694, + 0.015444966033101082, + 0.002818578388541937, + 0.0032150240149348974, + 0.003318624570965767, + 0.005968278739601374, + 0.002581565175205469, + 0.0031890678219497204, + 0.004907447379082441, + 0.003026572521775961, + 0.0038373167626559734, + 0.004328109789639711, + 0.002538251457735896, + 0.004271256737411022, + 0.00419000955298543, + 0.003414371283724904, + 0.004062640015035868, + 0.003455578116700053, + 0.023977162316441536, + 0.023118101060390472, + 0.01296868734061718, + 0.02131565473973751, + 0.014816325157880783, + 0.011652330867946148, + 0.024133862927556038, + 0.014127444475889206, + 0.004917792044579983, + 0.005355160683393478, + 0.003009797539561987, + 0.005514688789844513, + 0.003721602726727724, + 0.0034851476084440947, + 0.005327971652150154, + 0.005388676188886166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.12710295617580414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979450300335884 + ], + "opposite_logprob_tokens": [ + -3.529794454574585 + ], + "logprob_reward": -0.02979450300335884, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979450300335884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21922233700752258, + 0.10040797293186188, + 0.12181795388460159, + 0.06968856602907181, + 0.04477256163954735, + 0.2704130709171295, + 0.042732905596494675, + 0.029347117990255356, + 0.06350567936897278, + 0.03809189423918724 + ], + "credit_map": [ + 0.01959434151649475, + 0.025600578635931015, + 0.018428979441523552, + 0.03802220895886421, + 0.025473622605204582, + 0.033362988382577896, + 0.02983349934220314, + 0.028906114399433136, + 0.00892768707126379, + 0.016687920317053795, + 0.005550122354179621, + 0.013169861398637295, + 0.011738305911421776, + 0.01478629745543003, + 0.01215825043618679, + 0.017389517277479172, + 0.014160833321511745, + 0.014702281914651394, + 0.013012937270104885, + 0.0218499805778265, + 0.012147638946771622, + 0.01786518096923828, + 0.017333412542939186, + 0.010745678097009659, + 0.007200383115559816, + 0.007556393276900053, + 0.005329356994479895, + 0.011026278138160706, + 0.006982840597629547, + 0.006566343363374472, + 0.015686536207795143, + 0.009340430609881878, + 0.004909928888082504, + 0.0056930542923510075, + 0.003939582500606775, + 0.008514964953064919, + 0.004586445167660713, + 0.006653050892055035, + 0.006254986859858036, + 0.004220540169626474, + 0.03175283595919609, + 0.028507115319371223, + 0.02509244531393051, + 0.04580000042915344, + 0.028015922755002975, + 0.048618752509355545, + 0.024086935445666313, + 0.038539014756679535, + 0.004794569220393896, + 0.0058462913148105145, + 0.004001174587756395, + 0.007942301221191883, + 0.004443368408828974, + 0.00432333629578352, + 0.006180987227708101, + 0.005200876854360104, + 0.0028517141472548246, + 0.003674929030239582, + 0.002939635654911399, + 0.004186433274298906, + 0.0028017463628202677, + 0.00271999160759151, + 0.004829250741750002, + 0.005343414377421141, + 0.007827944122254848, + 0.00932834018021822, + 0.0054484824649989605, + 0.008730337955057621, + 0.006862996611744165, + 0.005766655318439007, + 0.01199034322053194, + 0.00755057530477643, + 0.003866223618388176, + 0.004596637096256018, + 0.0029639299027621746, + 0.005039192736148834, + 0.005861165001988411, + 0.00382980820722878, + 0.007101536262780428, + 0.004833401180803776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.02979450300335884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2098332941532135, + 0.03824267536401749, + 0.030499964952468872, + 0.12984266877174377, + 0.10887539386749268, + 0.07603602111339569, + 0.03590492531657219, + 0.11556585878133774, + 0.22335782647132874, + 0.031841378659009933 + ], + "credit_map": [ + 0.021721526980400085, + 0.026764845475554466, + 0.020050549879670143, + 0.037057992070913315, + 0.022051513195037842, + 0.023673653602600098, + 0.028243135660886765, + 0.030270056799054146, + 0.0031983221415430307, + 0.0045206863433122635, + 0.003766315756365657, + 0.006535198539495468, + 0.0035983757115900517, + 0.005359983537346125, + 0.0051706391386687756, + 0.006093147210776806, + 0.0036407571751624346, + 0.004146294202655554, + 0.003118176944553852, + 0.003807469503954053, + 0.0033082261215895414, + 0.003551597474142909, + 0.004990415181964636, + 0.003937026485800743, + 0.014258665964007378, + 0.015873372554779053, + 0.014710958115756512, + 0.023465557023882866, + 0.012436803430318832, + 0.009530911222100258, + 0.018235789611935616, + 0.02133060060441494, + 0.010888745076954365, + 0.01273201871663332, + 0.010889515280723572, + 0.017940353602170944, + 0.010311299934983253, + 0.008559808135032654, + 0.023275505751371384, + 0.014278137125074863, + 0.008088498376309872, + 0.008954002521932125, + 0.006910168565809727, + 0.014965045265853405, + 0.007920697331428528, + 0.012023267336189747, + 0.008420553058385849, + 0.008753784000873566, + 0.0039267754182219505, + 0.004346049390733242, + 0.004609441384673119, + 0.008017764426767826, + 0.0037771023344248533, + 0.0032794007565826178, + 0.004150846973061562, + 0.003797543467953801, + 0.01144182588905096, + 0.012494509108364582, + 0.012487935833632946, + 0.020313309505581856, + 0.011675609275698662, + 0.011561136692762375, + 0.02079913392663002, + 0.014792395755648613, + 0.028588753193616867, + 0.029524216428399086, + 0.0220665093511343, + 0.03748786449432373, + 0.01994256116449833, + 0.022392069920897484, + 0.04535256326198578, + 0.018003273755311966, + 0.003382198279723525, + 0.00376726477406919, + 0.0030959921423345804, + 0.004705376923084259, + 0.002878866158425808, + 0.003375269239768386, + 0.006789149716496468, + 0.00384726095944643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.008638040162622929 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on a pair of calloused, rough-skinned, aged hands intently engaged in a woodcarving process. A warm, directional light from above and side precisely illuminates the work area, casting three-dimensional shadows across the wrinkles of the knuckles and the deep lines of the palm. One hand firmly holds a block of light-colored, fine-grained basswood, while the other grips a carving knife, its handle sleek from years of use and its blade gleaming with a cool, metallic sheen. With steady, deliberate pressure from the wrist, the sharp tip of the knife delicately slices into the wood, shaving off thin, naturally curly chips. These chips hover briefly in the air, like light snowflakes, before slowly drifting down to the workbench, blurred into a blur of light in the background. The entire scene, imbued with a profound tranquility, is a silent celebration of concentration and skill. The warm yellow primary color creates a subtle contrast with the cool glow of the knife's blade, highlighting the power and delicacy of the creative process.", + "seed": 17253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_4192_seed_17253_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": -0.028452198207378387, + "vlm_logprob": -0.028452198207378387, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.028452198207378387, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02522290125489235 + ], + "opposite_logprob_tokens": [ + -3.9002230167388916 + ], + "logprob_reward": -0.02522290125489235, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02522290125489235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08199609071016312, + 0.05636625736951828, + 0.03929942101240158, + 0.04301979020237923, + 0.5462498068809509, + 0.12502844631671906, + 0.03032298944890499, + 0.04025653004646301, + 0.013410969637334347, + 0.024049680680036545 + ], + "credit_map": [ + 0.008113723248243332, + 0.004541339352726936, + 0.018925679847598076, + 0.0051109399646520615, + 0.005959818605333567, + 0.008839565329253674, + 0.005764050409197807, + 0.02474098466336727, + 0.009443923830986023, + 0.012279024347662926, + 0.005646444857120514, + 0.0037334549706429243, + 0.004032365046441555, + 0.005729017313569784, + 0.005716351326555014, + 0.00978567823767662, + 0.006552814040333033, + 0.00879549328237772, + 0.004583609290421009, + 0.004415981471538544, + 0.002022409811615944, + 0.003881812794134021, + 0.0030982487369328737, + 0.005949057172983885, + 0.006166824605315924, + 0.006773531436920166, + 0.005077683832496405, + 0.005318712908774614, + 0.0037206371780484915, + 0.004782934673130512, + 0.005055710207670927, + 0.006123761180788279, + 0.05891622602939606, + 0.2375318706035614, + 0.07269356399774551, + 0.022351445630192757, + 0.03312966972589493, + 0.03768864646553993, + 0.03588774427771568, + 0.048050668090581894, + 0.010687016882002354, + 0.04423220083117485, + 0.019997188821434975, + 0.00866709090769291, + 0.01015957910567522, + 0.010549493134021759, + 0.00966673530638218, + 0.01106914784759283, + 0.0032323882915079594, + 0.002741029020398855, + 0.003991773352026939, + 0.0031592436134815216, + 0.003970480058342218, + 0.005598148796707392, + 0.003135845996439457, + 0.004494083113968372, + 0.00550844706594944, + 0.008804673328995705, + 0.00410600658506155, + 0.0031635044142603874, + 0.006052708253264427, + 0.004404894076287746, + 0.003651595441624522, + 0.004564701579511166, + 0.0015813603531569242, + 0.0010999924270436168, + 0.0014401799999177456, + 0.001670323545113206, + 0.0024859923869371414, + 0.0014819729840382934, + 0.0014602536102756858, + 0.0021908951457589865, + 0.005025647580623627, + 0.002136825816705823, + 0.0028189204167574644, + 0.002557395724579692, + 0.0031882147304713726, + 0.001983358757570386, + 0.0029977429658174515, + 0.0033415756188333035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02522290125489235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226562343537807 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.014226562343537807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226562343537807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10889172554016113, + 0.13782525062561035, + 0.12660691142082214, + 0.17034640908241272, + 0.07350220531225204, + 0.07791101932525635, + 0.05165191367268562, + 0.10672086477279663, + 0.0868365615606308, + 0.05970717594027519 + ], + "credit_map": [ + 0.00873042456805706, + 0.013078327290713787, + 0.019873622804880142, + 0.010473924688994884, + 0.009957262314856052, + 0.010775898583233356, + 0.011402547359466553, + 0.024599717929959297, + 0.013556399382650852, + 0.03618906810879707, + 0.012346661649644375, + 0.009161043912172318, + 0.014202294871211052, + 0.013263645581901073, + 0.011560899205505848, + 0.027545245364308357, + 0.013841027393937111, + 0.02886442095041275, + 0.01362602785229683, + 0.006502616684883833, + 0.013645715080201626, + 0.010345520451664925, + 0.011495007202029228, + 0.02828657440841198, + 0.013904435560107231, + 0.050762858241796494, + 0.01623874343931675, + 0.008899759501218796, + 0.016365572810173035, + 0.011821993626654148, + 0.011247138492763042, + 0.04110591858625412, + 0.006845130585134029, + 0.023989880457520485, + 0.006935310084372759, + 0.0039056031964719296, + 0.004992024973034859, + 0.004681190010160208, + 0.006067286245524883, + 0.016085781157016754, + 0.00786615815013647, + 0.02670670486986637, + 0.013309935107827187, + 0.006797825451940298, + 0.004913657438009977, + 0.006583321839570999, + 0.006248787045478821, + 0.005484625697135925, + 0.005316682625561953, + 0.005074756219983101, + 0.007034360431134701, + 0.007639685180038214, + 0.006800705101341009, + 0.007626824546605349, + 0.006706457119435072, + 0.005452439188957214, + 0.012740097008645535, + 0.02553205005824566, + 0.01358895655721426, + 0.012223812751471996, + 0.01632034033536911, + 0.009460804983973503, + 0.0075966049917042255, + 0.009258199483156204, + 0.008530093356966972, + 0.0101793073117733, + 0.01018170639872551, + 0.013189258985221386, + 0.018858419731259346, + 0.008744153194129467, + 0.008591650985181332, + 0.008561969734728336, + 0.007401638198643923, + 0.004581769462674856, + 0.009169534780085087, + 0.01192069984972477, + 0.0061445399187505245, + 0.005218948237597942, + 0.008010228164494038, + 0.007259818725287914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.014226562343537807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430925153195858 + ], + "opposite_logprob_tokens": [ + -3.1680924892425537 + ], + "logprob_reward": -0.0430925153195858, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430925153195858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11412222683429718, + 0.057460006326436996, + 0.2991851568222046, + 0.07766719907522202, + 0.16839395463466644, + 0.02597706764936447, + 0.038872130215168, + 0.08034049719572067, + 0.10167612135410309, + 0.03630562126636505 + ], + "credit_map": [ + 0.010843600146472454, + 0.011536721140146255, + 0.022643020376563072, + 0.00738870445638895, + 0.010237670503556728, + 0.013392865657806396, + 0.010279107838869095, + 0.02780054323375225, + 0.0063899802044034, + 0.006660460028797388, + 0.005086837336421013, + 0.0056530628353357315, + 0.00818862859159708, + 0.006381731480360031, + 0.01241227239370346, + 0.006687036715447903, + 0.03053976036608219, + 0.0723726898431778, + 0.02842259779572487, + 0.017665142193436623, + 0.03694140166044235, + 0.027261344715952873, + 0.024145860224962234, + 0.061836354434490204, + 0.006427066400647163, + 0.013670925982296467, + 0.005372846033424139, + 0.004130801651626825, + 0.020528413355350494, + 0.00923759676516056, + 0.010892318561673164, + 0.0074072289280593395, + 0.017817169427871704, + 0.07006160914897919, + 0.02495403029024601, + 0.007991163991391659, + 0.011770226992666721, + 0.011803585104644299, + 0.012045553885400295, + 0.011950615793466568, + 0.00227328366599977, + 0.008177476935088634, + 0.0033671227283775806, + 0.00197027251124382, + 0.0024697724729776382, + 0.002574739046394825, + 0.0027900899294763803, + 0.0023543089628219604, + 0.004806141834706068, + 0.0029667005874216557, + 0.005211186595261097, + 0.00456799054518342, + 0.005095904227346182, + 0.006383543834090233, + 0.0048166788183152676, + 0.005023980978876352, + 0.00733792083337903, + 0.011985303834080696, + 0.00884050503373146, + 0.01172135304659605, + 0.016176637262105942, + 0.006667463108897209, + 0.010441095568239689, + 0.007170214783400297, + 0.008921725675463676, + 0.01274960394948721, + 0.010841449722647667, + 0.01116687897592783, + 0.027167757973074913, + 0.01177907083183527, + 0.009248809888958931, + 0.009800822474062443, + 0.006197992246598005, + 0.004411014262586832, + 0.004193434491753578, + 0.005489199422299862, + 0.004433308262377977, + 0.0038351083640009165, + 0.003699241206049919, + 0.004046323709189892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0430925153195858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864592105150223 + ], + "opposite_logprob_tokens": [ + -3.0486459732055664 + ], + "logprob_reward": -0.04864592105150223, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864592105150223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12468632310628891, + 0.13668309152126312, + 0.06717630475759506, + 0.047052931040525436, + 0.4129825234413147, + 0.030962403863668442, + 0.04021163657307625, + 0.04146458953619003, + 0.053731583058834076, + 0.04504869878292084 + ], + "credit_map": [ + 0.013914838433265686, + 0.02389492653310299, + 0.01810751110315323, + 0.011844162829220295, + 0.009710876271128654, + 0.01531297154724598, + 0.016865402460098267, + 0.015035632997751236, + 0.016620084643363953, + 0.030588559806346893, + 0.01590382494032383, + 0.008187680505216122, + 0.008211671374738216, + 0.01459254790097475, + 0.012346422299742699, + 0.030232282355427742, + 0.0104305325075984, + 0.012643816880881786, + 0.006598317064344883, + 0.006779919378459454, + 0.007413370534777641, + 0.007732669822871685, + 0.0070712873712182045, + 0.008506381884217262, + 0.005581600125879049, + 0.005822912789881229, + 0.0054861693643033504, + 0.005410042125731707, + 0.005772085394710302, + 0.004919667262583971, + 0.007197931408882141, + 0.006862524431198835, + 0.04459301009774208, + 0.1761714220046997, + 0.056251317262649536, + 0.018740486353635788, + 0.028804071247577667, + 0.028796635568141937, + 0.026924191042780876, + 0.03270137682557106, + 0.003415846498683095, + 0.005360438022762537, + 0.0038173168431967497, + 0.0031753098592162132, + 0.0044242809526622295, + 0.004003241192549467, + 0.003575880080461502, + 0.0031900920439511538, + 0.004504395183175802, + 0.0035617980174720287, + 0.004806807264685631, + 0.006088198162615299, + 0.005589944776147604, + 0.004893734119832516, + 0.006373497191816568, + 0.004393258597701788, + 0.007159282919019461, + 0.005094898398965597, + 0.004003182519227266, + 0.0036518177948892117, + 0.006554207298904657, + 0.004906160291284323, + 0.005003622733056545, + 0.005091419909149408, + 0.00978924985975027, + 0.006407070904970169, + 0.005939386319369078, + 0.006674916483461857, + 0.008189289830625057, + 0.005949248094111681, + 0.00424816133454442, + 0.006534261628985405, + 0.006180234253406525, + 0.007718510925769806, + 0.004307845141738653, + 0.005301944445818663, + 0.008359147235751152, + 0.005010643973946571, + 0.0040354048833251, + 0.004134970251470804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.04864592105150223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3063998520374298, + 0.11032263189554214, + 0.06215497851371765, + 0.09502008557319641, + 0.14954005181789398, + 0.07478176802396774, + 0.03991454467177391, + 0.04612260311841965, + 0.0433727502822876, + 0.07237070053815842 + ], + "credit_map": [ + 0.023237494751811028, + 0.01752666011452675, + 0.07886073738336563, + 0.015438597649335861, + 0.016581149771809578, + 0.030865110456943512, + 0.018136082217097282, + 0.10575404018163681, + 0.010583572089672089, + 0.019731080159544945, + 0.012736847624182701, + 0.008129920810461044, + 0.010463886894285679, + 0.013508700765669346, + 0.016764219850301743, + 0.018404414877295494, + 0.009297043085098267, + 0.00579424574971199, + 0.008337815292179585, + 0.006436680443584919, + 0.007108884397894144, + 0.00686702411621809, + 0.0061784666031599045, + 0.012134821154177189, + 0.009323520585894585, + 0.02782781422138214, + 0.006757418159395456, + 0.005316845141351223, + 0.012927887961268425, + 0.008884431794285774, + 0.007840455509722233, + 0.016141710802912712, + 0.015576642006635666, + 0.06411730498075485, + 0.019559403881430626, + 0.007594646420329809, + 0.008679797872900963, + 0.010637910105288029, + 0.010098537430167198, + 0.013275821693241596, + 0.007001994643360376, + 0.02459510788321495, + 0.011581500992178917, + 0.007528951857239008, + 0.005357511341571808, + 0.006556689739227295, + 0.006467313040047884, + 0.005692698061466217, + 0.004389455541968346, + 0.0034708837047219276, + 0.0048685697838664055, + 0.004205414094030857, + 0.007720425724983215, + 0.0052656205371022224, + 0.004590155556797981, + 0.005404019262641668, + 0.003937473986297846, + 0.004424628801643848, + 0.005442170891910791, + 0.008170795626938343, + 0.009481155313551426, + 0.004363149870187044, + 0.006143510341644287, + 0.004159719683229923, + 0.004652156960219145, + 0.003992535639554262, + 0.004161627031862736, + 0.0043939449824392796, + 0.01242339052259922, + 0.004691810812801123, + 0.004301608074456453, + 0.004755675792694092, + 0.007736231666058302, + 0.008943929336965084, + 0.008782769553363323, + 0.012009757570922375, + 0.012055649422109127, + 0.009678700938820839, + 0.006941578816622496, + 0.006222090683877468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.011073091067373753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera remains still, focused on a pair of calloused, rough-skinned, aged hands intently engaged in a woodcarving process. A warm, directional light from above and side precisely illuminates the work area, casting three-dimensional shadows across the wrinkles of the knuckles and the deep lines of the palm. One hand firmly holds a block of light-colored, fine-grained basswood, while the other grips a carving knife, its handle sleek from years of use and its blade gleaming with a cool, metallic sheen. With steady, deliberate pressure from the wrist, the sharp tip of the knife delicately slices into the wood, shaving off thin, naturally curly chips. These chips hover briefly in the air, like light snowflakes, before slowly drifting down to the workbench, blurred into a blur of light in the background. The entire scene, imbued with a profound tranquility, is a silent celebration of concentration and skill. The warm yellow primary color creates a subtle contrast with the cool glow of the knife's blade, highlighting the power and delicacy of the creative process.", + "seed": 17254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_0/index_4192_seed_17254_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": -0.03292366899549961, + "vlm_logprob": -0.03292366899549961, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03292366899549961, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026523014530539513 + ], + "opposite_logprob_tokens": [ + -3.9015231132507324 + ], + "logprob_reward": -0.026523014530539513, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026523014530539513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961079090833664, + 0.16977235674858093, + 0.10743030160665512, + 0.0339927114546299, + 0.11167167872190475, + 0.07959083467721939, + 0.028852077201008797, + 0.038753293454647064, + 0.04867727681994438, + 0.1851515918970108 + ], + "credit_map": [ + 0.015797194093465805, + 0.020365899428725243, + 0.015268069691956043, + 0.03045147843658924, + 0.022732015699148178, + 0.04332255944609642, + 0.022998766973614693, + 0.02517193928360939, + 0.016152435913681984, + 0.033536192029714584, + 0.016224965453147888, + 0.02081468515098095, + 0.01951693184673786, + 0.020626714453101158, + 0.017611127346754074, + 0.025289304554462433, + 0.009497882798314095, + 0.018212970346212387, + 0.011323429644107819, + 0.015745176002383232, + 0.007488156668841839, + 0.01359498593956232, + 0.011767558753490448, + 0.019800135865807533, + 0.002902246080338955, + 0.003929512109607458, + 0.0034726280719041824, + 0.007629454135894775, + 0.0029248585924506187, + 0.003771944670006633, + 0.0041481382213532925, + 0.005213929805904627, + 0.012040474452078342, + 0.012081146240234375, + 0.010554743930697441, + 0.016498951241374016, + 0.01472489070147276, + 0.012937293387949467, + 0.012298108078539371, + 0.020536063238978386, + 0.008903390727937222, + 0.008596835657954216, + 0.00540442718192935, + 0.008187170140445232, + 0.0161790419369936, + 0.011197363026440144, + 0.009243970736861229, + 0.011878629215061665, + 0.0030279012862592936, + 0.00306805782020092, + 0.0030078731942921877, + 0.005499090533703566, + 0.003035344183444977, + 0.0034774483647197485, + 0.003101647598668933, + 0.004634713288396597, + 0.004351114388555288, + 0.004361263010650873, + 0.0036959885619580746, + 0.006882859859615564, + 0.005281539633870125, + 0.0045683966018259525, + 0.004206126090139151, + 0.005406005773693323, + 0.003953263163566589, + 0.00539052952080965, + 0.0035196468234062195, + 0.009008044376969337, + 0.006937836296856403, + 0.0049195317551493645, + 0.007245399057865143, + 0.007703030947595835, + 0.01718808151781559, + 0.005974440835416317, + 0.009566097520291805, + 0.0074192071333527565, + 0.014150094240903854, + 0.08170374482870102, + 0.005927551072090864, + 0.04322238266468048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.026523014530539513 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020650971680879593 + ], + "opposite_logprob_tokens": [ + -3.895650863647461 + ], + "logprob_reward": -0.020650971680879593, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020650971680879593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1653776466846466, + 0.05229055508971214, + 0.18136076629161835, + 0.04204787686467171, + 0.06062012538313866, + 0.03809848427772522, + 0.02895909920334816, + 0.04741447791457176, + 0.04781455546617508, + 0.33601635694503784 + ], + "credit_map": [ + 0.014777034521102905, + 0.022593848407268524, + 0.012110579758882523, + 0.01677887514233589, + 0.02505718730390072, + 0.0418090745806694, + 0.014586878940463066, + 0.01766417734324932, + 0.005372349172830582, + 0.006825799588114023, + 0.005524252541363239, + 0.0075646061450243, + 0.006077125668525696, + 0.0077186631970107555, + 0.006837396416813135, + 0.006370368413627148, + 0.020720956847071648, + 0.01579451747238636, + 0.026892103254795074, + 0.029932862147688866, + 0.015452295541763306, + 0.024038510397076607, + 0.025252345949411392, + 0.023277197033166885, + 0.003287183353677392, + 0.009680813178420067, + 0.0037114694714546204, + 0.006338309030979872, + 0.004772385116666555, + 0.0046148598194122314, + 0.004118025302886963, + 0.005524836480617523, + 0.007552187889814377, + 0.00497428048402071, + 0.0058122677728533745, + 0.006866826675832272, + 0.008418025448918343, + 0.011525076813995838, + 0.006873455364257097, + 0.008598006330430508, + 0.003125817282125354, + 0.0038444402161985636, + 0.002254511695355177, + 0.004195415880531073, + 0.006495193112641573, + 0.006470121443271637, + 0.003954182844609022, + 0.007758805528283119, + 0.0029624905437231064, + 0.005792466923594475, + 0.002018362283706665, + 0.004458954557776451, + 0.003144933143630624, + 0.002895725890994072, + 0.0035865572281181812, + 0.004099608398973942, + 0.006332986056804657, + 0.005103143397718668, + 0.003568071173503995, + 0.005150233395397663, + 0.006219075061380863, + 0.010981089435517788, + 0.004539587534964085, + 0.005520296283066273, + 0.0032672686502337456, + 0.005763505585491657, + 0.0031615886837244034, + 0.007384224329143763, + 0.009273163974285126, + 0.005198314320296049, + 0.007464361842721701, + 0.006302133202552795, + 0.03394017741084099, + 0.010367383249104023, + 0.01907689869403839, + 0.012602955102920532, + 0.026959983631968498, + 0.15084442496299744, + 0.010651730932295322, + 0.07157283276319504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.020650971680879593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15330319106578827, + 0.1447492390871048, + 0.1195453405380249, + 0.028976188972592354, + 0.06554529070854187, + 0.02916874922811985, + 0.04593022167682648, + 0.08376245200634003, + 0.0834546834230423, + 0.24556466937065125 + ], + "credit_map": [ + 0.011610831134021282, + 0.032432641834020615, + 0.013710765168070793, + 0.018827242776751518, + 0.023140428587794304, + 0.017053402960300446, + 0.01494518481194973, + 0.021582700312137604, + 0.012465978972613811, + 0.03130657970905304, + 0.013131015002727509, + 0.017137622460722923, + 0.01658228226006031, + 0.018116794526576996, + 0.014244835823774338, + 0.021764112636446953, + 0.010404916480183601, + 0.02778540551662445, + 0.01084927562624216, + 0.012707439251244068, + 0.0119382468983531, + 0.012994230724871159, + 0.008644252084195614, + 0.02422155626118183, + 0.002535406034439802, + 0.006167464889585972, + 0.002434495836496353, + 0.005283750593662262, + 0.003019144758582115, + 0.0029995024669915438, + 0.0029879144858568907, + 0.003548507345840335, + 0.007853378541767597, + 0.006134625989943743, + 0.00642482889816165, + 0.010529367253184319, + 0.006806870456784964, + 0.007642618380486965, + 0.008257864974439144, + 0.011895732954144478, + 0.001288452884182334, + 0.0024021046701818705, + 0.0013183723203837872, + 0.0018357676453888416, + 0.002153001492843032, + 0.009478764608502388, + 0.0013681708369404078, + 0.009324111975729465, + 0.004581233486533165, + 0.008469498716294765, + 0.004126470535993576, + 0.008228939957916737, + 0.004689262248575687, + 0.005554802715778351, + 0.004797446075826883, + 0.0054825684055686, + 0.014062752947211266, + 0.01021917350590229, + 0.005289462860673666, + 0.009582682512700558, + 0.009534267708659172, + 0.017877714708447456, + 0.007886990904808044, + 0.009309405460953712, + 0.00631156750023365, + 0.014903056435286999, + 0.006308514624834061, + 0.010982134379446507, + 0.014944793656468391, + 0.009190783835947514, + 0.010665767826139927, + 0.010148057714104652, + 0.02827027440071106, + 0.009216152131557465, + 0.0132530452683568, + 0.009065574035048485, + 0.01940125972032547, + 0.10937798023223877, + 0.008442646823823452, + 0.048537708818912506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903401345014572 + ], + "opposite_logprob_tokens": [ + -2.464034080505371 + ], + "logprob_reward": -0.08903401345014572, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903401345014572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17742858827114105, + 0.056643564254045486, + 0.11741482466459274, + 0.0589887797832489, + 0.1691247671842575, + 0.05620924383401871, + 0.07652489840984344, + 0.04502271115779877, + 0.13989320397377014, + 0.10274949669837952 + ], + "credit_map": [ + 0.018296729773283005, + 0.015691641718149185, + 0.011896690353751183, + 0.0190709438174963, + 0.027636710554361343, + 0.04725584015250206, + 0.017566226422786713, + 0.02001378871500492, + 0.0064662969671189785, + 0.010364077985286713, + 0.004706156440079212, + 0.006075243465602398, + 0.0077486722730100155, + 0.007926873862743378, + 0.006488364655524492, + 0.006867877673357725, + 0.011462309397757053, + 0.027123134583234787, + 0.015115740709006786, + 0.013542389497160912, + 0.0112299844622612, + 0.014040714129805565, + 0.012078993953764439, + 0.012821543961763382, + 0.004814434796571732, + 0.01446562260389328, + 0.004558413289487362, + 0.008732056245207787, + 0.0074340300634503365, + 0.008190883323550224, + 0.004412471316754818, + 0.006380865350365639, + 0.0173936877399683, + 0.018955888226628304, + 0.017852844670414925, + 0.02788449078798294, + 0.02459249645471573, + 0.022825337946414948, + 0.017397094517946243, + 0.022222908213734627, + 0.0033411721233278513, + 0.005426399409770966, + 0.0032282460015267134, + 0.0037723996210843325, + 0.006715320982038975, + 0.018798312172293663, + 0.003689070465043187, + 0.01123831607401371, + 0.006446527317166328, + 0.011460171081125736, + 0.007553842384368181, + 0.014527512714266777, + 0.008644592948257923, + 0.0071592554450035095, + 0.00896670762449503, + 0.011766284704208374, + 0.00591541500762105, + 0.004764006473124027, + 0.003059366485103965, + 0.00453033484518528, + 0.00687723932787776, + 0.008624871261417866, + 0.004377160221338272, + 0.00687431450933218, + 0.01407804898917675, + 0.019392363727092743, + 0.009281637147068977, + 0.016999460756778717, + 0.020466972142457962, + 0.01782005839049816, + 0.015004360117018223, + 0.02685028687119484, + 0.006939418148249388, + 0.005236006807535887, + 0.007540835998952389, + 0.0056846425868570805, + 0.008220570161938667, + 0.040930889546871185, + 0.004396992269903421, + 0.023800132796168327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.08903401345014572 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 the main subjects a pair of aged, rough-skinned hands engaged in woodcarving?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is one hand holding a block of light-colored, fine-grained basswood while the other holds a carving knife?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the knife blade gleaming with a cool, metallic sheen while the handle appears sleek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are thin, naturally curly wood chips shown hovering in the air and drifting down to the workbench?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned in a still close-up shot with warm, directional lighting from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419494766741991 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.01419494766741991, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419494766741991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06239969655871391, + 0.0573190376162529, + 0.03886186704039574, + 0.040804918855428696, + 0.09999258816242218, + 0.11376594752073288, + 0.03439587354660034, + 0.02605881541967392, + 0.06672848016023636, + 0.4596727788448334 + ], + "credit_map": [ + 0.006007409654557705, + 0.012156697921454906, + 0.00526966480538249, + 0.006817962042987347, + 0.01103140227496624, + 0.007236070930957794, + 0.00427867379039526, + 0.009601814672350883, + 0.005625032354146242, + 0.010902530513703823, + 0.004608483985066414, + 0.005057753995060921, + 0.010060030966997147, + 0.00941787101328373, + 0.004508215934038162, + 0.007139122113585472, + 0.0033506695181131363, + 0.007267187815159559, + 0.004231051541864872, + 0.004992163740098476, + 0.004551473539322615, + 0.004491806961596012, + 0.005302837584167719, + 0.004674672614783049, + 0.003375800559297204, + 0.006520946975797415, + 0.0034613474272191525, + 0.006510993931442499, + 0.004772324115037918, + 0.006208817474544048, + 0.0046379826962947845, + 0.005316704977303743, + 0.00965280644595623, + 0.011488054879009724, + 0.009762008674442768, + 0.014037775807082653, + 0.01286469865590334, + 0.015095890499651432, + 0.009760404005646706, + 0.017330950126051903, + 0.008680134080350399, + 0.007553538307547569, + 0.004519471433013678, + 0.006723626516759396, + 0.006761173717677593, + 0.03572497516870499, + 0.005505260080099106, + 0.03829777240753174, + 0.003056186717003584, + 0.005528518930077553, + 0.003188683418557048, + 0.0057861763052642345, + 0.004661619197577238, + 0.0030809936579316854, + 0.00411900645121932, + 0.004974688403308392, + 0.002572892466560006, + 0.004448932595551014, + 0.0022852737456560135, + 0.004422064870595932, + 0.0027429626788944006, + 0.0030738578643649817, + 0.003044264856725931, + 0.0034685670398175716, + 0.005229467060416937, + 0.01066496316343546, + 0.004835972562432289, + 0.00992860272526741, + 0.011486715637147427, + 0.006908579729497433, + 0.008783955127000809, + 0.008890220895409584, + 0.055562738329172134, + 0.012265303172171116, + 0.023076755926012993, + 0.014043636620044708, + 0.03672396391630173, + 0.2100549042224884, + 0.014546148478984833, + 0.09339934587478638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.01419494766741991 + } + ], + "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/0000022/group_1/index_tensor([1588, 1588])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_tensor([1588, 1588])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c30cc1ce337999ee1fa955f247f4b68c7d0a09d5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_tensor([1588, 1588])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic overhead shot, the camera slowly tilting upward, begins at the glass and steel spire of a modern skyscraper. Against a gloomy, stormy sky, a massive, geometrically shaped alien mothership silently and majestically tears through the thick, rolling gray storm clouds. The vessel's sleek, dark metal hull gleams with an unearthly, iridescent sheen in the dim skylight. Its surface is covered in intricate, self-luminous blue patterns, like a giant circuit board, its structure seemingly undergoing minute, silent reconfigurations as it moves. With a slow, inexorable momentum, it majestically sweeps past the skyscraper's spire, its massive shadow instantly engulfing the cityscape below. As the vessel's energy field distorts the air as it passes, creating subtle ripples in the reflections on the spire's mirrored glass facade. The scene is imbued with an epic scale and a breathtaking, awe-inspiring stillness, creating the feeling of a grand and tense science fiction film.", + "seed": 28762, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28762_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": -0.07361453445628285, + "vlm_logprob": -0.07361453445628285, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.07361453445628285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.30513256788253784 + ], + "opposite_logprob_tokens": [ + -1.4301326274871826 + ], + "logprob_reward": -0.30513256788253784, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.30513256788253784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832439571619034, + 0.19628365337848663, + 0.12484478205442429, + 0.07995917648077011, + 0.047183532267808914, + 0.23260141909122467, + 0.06693775206804276, + 0.031202923506498337, + 0.05663835629820824, + 0.06602400541305542 + ], + "credit_map": [ + 0.01734786108136177, + 0.011708928272128105, + 0.015128730796277523, + 0.011294144205749035, + 0.011971698142588139, + 0.009842172265052795, + 0.012642833404242992, + 0.008388019166886806, + 0.026449399068951607, + 0.025014963001012802, + 0.03772962838411331, + 0.024747412651777267, + 0.0185710322111845, + 0.015611478127539158, + 0.026695912703871727, + 0.021463818848133087, + 0.007831783965229988, + 0.009674595668911934, + 0.01748264953494072, + 0.013870852068066597, + 0.019980836659669876, + 0.026902033016085625, + 0.017698299139738083, + 0.011403721757233143, + 0.007242217659950256, + 0.01030178926885128, + 0.012032371945679188, + 0.00908293854445219, + 0.011495504528284073, + 0.00874330010265112, + 0.008279282599687576, + 0.012781761586666107, + 0.007639447692781687, + 0.006935399025678635, + 0.004432633984833956, + 0.005825209431350231, + 0.006478287745267153, + 0.005667438730597496, + 0.004657390993088484, + 0.005547722335904837, + 0.019278714433312416, + 0.025960516184568405, + 0.03942955657839775, + 0.0183365810662508, + 0.0306325051933527, + 0.03581308200955391, + 0.016339505091309547, + 0.04681093618273735, + 0.01321212388575077, + 0.010353068821132183, + 0.007776289246976376, + 0.006279523950070143, + 0.007163431029766798, + 0.008939427323639393, + 0.005206253845244646, + 0.008007628843188286, + 0.004790150094777346, + 0.004954389296472073, + 0.0038363521452993155, + 0.0038685123436152935, + 0.0034761307761073112, + 0.003687245072796941, + 0.0034945732913911343, + 0.00309556582942605, + 0.007553671021014452, + 0.007042508106678724, + 0.010985537432134151, + 0.006906399969011545, + 0.00497707212343812, + 0.006696682423353195, + 0.005560088437050581, + 0.006916392128914595, + 0.009038964286446571, + 0.010528681799769402, + 0.007059955038130283, + 0.007493768818676472, + 0.00666534248739481, + 0.008102972991764545, + 0.0068779680877923965, + 0.010256349109113216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.30513256788253784 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651899941265583 + ], + "opposite_logprob_tokens": [ + -4.8826518058776855 + ], + "logprob_reward": -0.007651899941265583, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651899941265583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1367369145154953, + 0.1583203375339508, + 0.20861230790615082, + 0.08946210145950317, + 0.050825413316488266, + 0.10450205951929092, + 0.06831106543540955, + 0.09958496689796448, + 0.06159159541130066, + 0.022053183987736702 + ], + "credit_map": [ + 0.014701852574944496, + 0.019179692491889, + 0.02062973938882351, + 0.014897982589900494, + 0.021056154742836952, + 0.016036834567785263, + 0.01850668340921402, + 0.011727988719940186, + 0.016574610024690628, + 0.01743912696838379, + 0.021641988307237625, + 0.02231982909142971, + 0.015870941802859306, + 0.01934238336980343, + 0.0288588497787714, + 0.016272621229290962, + 0.01346498541533947, + 0.013667406514286995, + 0.028477519750595093, + 0.025648001581430435, + 0.03107021376490593, + 0.039693042635917664, + 0.037563521414995193, + 0.019027644768357277, + 0.008870045654475689, + 0.011812061071395874, + 0.010059859603643417, + 0.008105931803584099, + 0.016859499737620354, + 0.019423868507146835, + 0.005984123796224594, + 0.008346719667315483, + 0.006129966117441654, + 0.0079368120059371, + 0.005939751863479614, + 0.006496005225926638, + 0.008726018480956554, + 0.006311607547104359, + 0.003574509872123599, + 0.005710748489946127, + 0.007651439867913723, + 0.013589010573923588, + 0.019935939460992813, + 0.008105338551104069, + 0.011806356720626354, + 0.015003841370344162, + 0.008387092500925064, + 0.02002304419875145, + 0.007252131588757038, + 0.006863932590931654, + 0.011184843257069588, + 0.00814258772879839, + 0.008315670304000378, + 0.010194703005254269, + 0.008219383656978607, + 0.008137817494571209, + 0.013731065206229687, + 0.016044987365603447, + 0.011038489639759064, + 0.011868899688124657, + 0.009108076803386211, + 0.015194362960755825, + 0.012136843986809254, + 0.010462249629199505, + 0.009696335531771183, + 0.0074415551498532295, + 0.009643083438277245, + 0.008459602482616901, + 0.004889816511422396, + 0.009648489765822887, + 0.005785695277154446, + 0.006027022376656532, + 0.0021947682835161686, + 0.002919401740655303, + 0.004150174092501402, + 0.002445932710543275, + 0.0025172510650008917, + 0.0023757426533848047, + 0.002898221369832754, + 0.0025516925379633904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.007651899941265583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022050845436751842 + ], + "opposite_logprob_tokens": [ + -6.127204895019531 + ], + "logprob_reward": -0.0022050845436751842, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022050845436751842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14401257038116455, + 0.09755229204893112, + 0.05656107887625694, + 0.06293497234582901, + 0.05000870302319527, + 0.07335302233695984, + 0.06836158782243729, + 0.0752546414732933, + 0.24220943450927734, + 0.12975165247917175 + ], + "credit_map": [ + 0.013327977620065212, + 0.019760873168706894, + 0.021330850198864937, + 0.0126344608142972, + 0.017592355608940125, + 0.025445155799388885, + 0.019930191338062286, + 0.013990722596645355, + 0.00889301486313343, + 0.01126796379685402, + 0.018838558346033096, + 0.0131351538002491, + 0.010350984521210194, + 0.014095660299062729, + 0.009921180084347725, + 0.01104978658258915, + 0.004144764505326748, + 0.0040110996924340725, + 0.012457209639251232, + 0.005332270171493292, + 0.007165992632508278, + 0.011348958127200603, + 0.0062753609381616116, + 0.005825426895171404, + 0.006551168859004974, + 0.008362138643860817, + 0.009413627907633781, + 0.005600358825176954, + 0.01039111241698265, + 0.011377627030014992, + 0.004310049116611481, + 0.006928896065801382, + 0.008416790515184402, + 0.0073662917129695415, + 0.007015320006757975, + 0.005294506438076496, + 0.006510953418910503, + 0.005811899434775114, + 0.0033226213417947292, + 0.006270325277000666, + 0.008505773730576038, + 0.009885655716061592, + 0.014435834251344204, + 0.006156550720334053, + 0.008319972082972527, + 0.00996827706694603, + 0.005749993026256561, + 0.010330975987017155, + 0.011186454445123672, + 0.010222792625427246, + 0.007961676456034184, + 0.006177467759698629, + 0.008150015957653522, + 0.009655458852648735, + 0.006837902124971151, + 0.008169824257493019, + 0.010715815238654613, + 0.012221765704452991, + 0.009962054900825024, + 0.009875922463834286, + 0.0076822564005851746, + 0.010244828648865223, + 0.0072593227960169315, + 0.007292683236300945, + 0.031170077621936798, + 0.027502594515681267, + 0.04041306674480438, + 0.03381355106830597, + 0.0166280847042799, + 0.03641516715288162, + 0.02812592126429081, + 0.028140990063548088, + 0.017855999991297722, + 0.02479470893740654, + 0.016838183626532555, + 0.014017353765666485, + 0.014289133250713348, + 0.015763752162456512, + 0.009950936771929264, + 0.01624160446226597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0022050845436751842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980075217783451 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980075217783451, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980075217783451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31453078985214233, + 0.3531937599182129, + 0.04038131237030029, + 0.09073523432016373, + 0.058021433651447296, + 0.03832440450787544, + 0.032091833651065826, + 0.02918240614235401, + 0.01940847747027874, + 0.02413041703402996 + ], + "credit_map": [ + 0.04068262130022049, + 0.0371122770011425, + 0.05091141164302826, + 0.0410040058195591, + 0.040965620428323746, + 0.0329476073384285, + 0.04760877415537834, + 0.02329845167696476, + 0.03365630283951759, + 0.031538330018520355, + 0.045030031353235245, + 0.06555488705635071, + 0.02905602939426899, + 0.03163531795144081, + 0.07722283899784088, + 0.03949999064207077, + 0.0033769835717976093, + 0.002547966782003641, + 0.004826521035283804, + 0.003837947966530919, + 0.007599706295877695, + 0.00940422061830759, + 0.0038601262494921684, + 0.004927834961563349, + 0.006148486398160458, + 0.009519143030047417, + 0.008990623988211155, + 0.00864589773118496, + 0.01637939177453518, + 0.01985839195549488, + 0.009620011784136295, + 0.011573287658393383, + 0.005372441373765469, + 0.008410818874835968, + 0.0066517493687570095, + 0.0062019494362175465, + 0.008486328646540642, + 0.0075974417850375175, + 0.006951237563043833, + 0.008349462412297726, + 0.00452914834022522, + 0.004913621582090855, + 0.006784089840948582, + 0.002587769879028201, + 0.004303268156945705, + 0.005625513847917318, + 0.0031032352708280087, + 0.006477751769125462, + 0.003600579919293523, + 0.003345935605466366, + 0.004826522897928953, + 0.0031530235428363085, + 0.004267335869371891, + 0.004675221163779497, + 0.004054389428347349, + 0.00416882149875164, + 0.0034523895010352135, + 0.006692598108202219, + 0.003225735854357481, + 0.0043085673823952675, + 0.0025456268340349197, + 0.003614409826695919, + 0.0028029722161591053, + 0.00254010665230453, + 0.002033716067671776, + 0.002778769703581929, + 0.004057948011904955, + 0.0025665946304798126, + 0.0015015980461612344, + 0.002403301652520895, + 0.0020474386401474476, + 0.0020191085059195757, + 0.002589976182207465, + 0.002672657137736678, + 0.0042338441126048565, + 0.0029355972073972225, + 0.002720771124586463, + 0.0026015439070761204, + 0.0031722867861390114, + 0.003203736385330558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02980075217783451 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21449173986911774, + 0.13098061084747314, + 0.3552931845188141, + 0.08712784200906754, + 0.06364639848470688, + 0.044805221259593964, + 0.022552475333213806, + 0.019141538068652153, + 0.031156126409769058, + 0.030804792419075966 + ], + "credit_map": [ + 0.025761272758245468, + 0.038753870874643326, + 0.024569621309638023, + 0.019618919119238853, + 0.019361142069101334, + 0.04517974331974983, + 0.023350872099399567, + 0.017896318808197975, + 0.012182196602225304, + 0.015012145042419434, + 0.019731776788830757, + 0.021381473168730736, + 0.0124953743070364, + 0.017392626032233238, + 0.016179749742150307, + 0.016605278477072716, + 0.02302459441125393, + 0.023996368050575256, + 0.05242409184575081, + 0.032019756734371185, + 0.06454986333847046, + 0.07679588347673416, + 0.04119544103741646, + 0.04128721356391907, + 0.004891667980700731, + 0.0066878655925393105, + 0.007503296714276075, + 0.00940090324729681, + 0.017077522352337837, + 0.016783826053142548, + 0.013276850804686546, + 0.011505914852023125, + 0.005396557040512562, + 0.007412530016154051, + 0.010839215479791164, + 0.0070801302790641785, + 0.007350410800427198, + 0.008037865161895752, + 0.007936401292681694, + 0.009593290276825428, + 0.005849228706210852, + 0.005066449288278818, + 0.005701517220586538, + 0.004603069741278887, + 0.007326958701014519, + 0.006533960811793804, + 0.003570275381207466, + 0.006153766065835953, + 0.0024677636101841927, + 0.002226659096777439, + 0.0029960034880787134, + 0.0021266748663038015, + 0.0019154276233166456, + 0.0040352786891162395, + 0.002668027300387621, + 0.004116642288863659, + 0.0019182688556611538, + 0.004273851402103901, + 0.0021789262536913157, + 0.0030822097323834896, + 0.0013516271719709039, + 0.002574196085333824, + 0.0019078346667811275, + 0.0018546248320490122, + 0.0037863662000745535, + 0.004206233657896519, + 0.004723713733255863, + 0.003731646342203021, + 0.0026636722031980753, + 0.0037112971767783165, + 0.0036605666391551495, + 0.00467263488098979, + 0.0031854347325861454, + 0.004502971190959215, + 0.005093206185847521, + 0.003323198761790991, + 0.003974976018071175, + 0.003526996122673154, + 0.0037327075842767954, + 0.0034653018228709698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02328236773610115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot, the camera slowly tilting upward, begins at the glass and steel spire of a modern skyscraper. Against a gloomy, stormy sky, a massive, geometrically shaped alien mothership silently and majestically tears through the thick, rolling gray storm clouds. The vessel's sleek, dark metal hull gleams with an unearthly, iridescent sheen in the dim skylight. Its surface is covered in intricate, self-luminous blue patterns, like a giant circuit board, its structure seemingly undergoing minute, silent reconfigurations as it moves. With a slow, inexorable momentum, it majestically sweeps past the skyscraper's spire, its massive shadow instantly engulfing the cityscape below. As the vessel's energy field distorts the air as it passes, creating subtle ripples in the reflections on the spire's mirrored glass facade. The scene is imbued with an epic scale and a breathtaking, awe-inspiring stillness, creating the feeling of a grand and tense science fiction film.", + "seed": 28763, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28763_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": -0.0850875094998628, + "vlm_logprob": -0.0850875094998628, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.0850875094998628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2788069248199463 + ], + "opposite_logprob_tokens": [ + -1.5288069248199463 + ], + "logprob_reward": -0.2788069248199463, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2788069248199463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0606435127556324, + 0.06705548614263535, + 0.04510920122265816, + 0.26409366726875305, + 0.12242544442415237, + 0.1805865615606308, + 0.09943347424268723, + 0.045911114662885666, + 0.10280582308769226, + 0.011935671791434288 + ], + "credit_map": [ + 0.007082780823111534, + 0.010195172391831875, + 0.008296019397675991, + 0.011350185610353947, + 0.009009660221636295, + 0.005693124141544104, + 0.004102298058569431, + 0.004914278630167246, + 0.00773474620655179, + 0.009241494350135326, + 0.008395010605454445, + 0.00752859003841877, + 0.009202242828905582, + 0.014293443411588669, + 0.004657004494220018, + 0.006002954673022032, + 0.004188936203718185, + 0.005354080814868212, + 0.006473360117524862, + 0.0053825401701033115, + 0.007290844805538654, + 0.0062146494165062904, + 0.004413888789713383, + 0.005790898110717535, + 0.02198271080851555, + 0.026095833629369736, + 0.04078491032123566, + 0.042374610900878906, + 0.03540191054344177, + 0.02870825119316578, + 0.03233075886964798, + 0.03641470894217491, + 0.012677481397986412, + 0.013367493636906147, + 0.013917787931859493, + 0.01779669150710106, + 0.011930983513593674, + 0.01797480508685112, + 0.010985452681779861, + 0.023774750530719757, + 0.031807057559490204, + 0.015948941931128502, + 0.008312528021633625, + 0.013442943803966045, + 0.012609639205038548, + 0.05883356183767319, + 0.013628335669636726, + 0.026003563776612282, + 0.009415585547685623, + 0.011104141362011433, + 0.00858184415847063, + 0.015950074419379234, + 0.01401606760919094, + 0.01203507836908102, + 0.010144608095288277, + 0.01818608120083809, + 0.003992374055087566, + 0.0052238283678889275, + 0.005105846095830202, + 0.006238103378564119, + 0.0052380431443452835, + 0.008216721937060356, + 0.0061624906957149506, + 0.005733712110668421, + 0.007030101027339697, + 0.013109054416418076, + 0.01053654495626688, + 0.015927065163850784, + 0.014301138930022717, + 0.01384253241121769, + 0.01643737405538559, + 0.011622007936239243, + 0.0006461783195845783, + 0.001323989825323224, + 0.0012901965528726578, + 0.0018038798589259386, + 0.0014590584905818105, + 0.00154831912368536, + 0.002244155388325453, + 0.0016198963858187199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.2788069248199463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067553711123764515 + ], + "opposite_logprob_tokens": [ + -5.006755352020264 + ], + "logprob_reward": -0.0067553711123764515, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067553711123764515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10087509453296661, + 0.1088452935218811, + 0.19394609332084656, + 0.08329620957374573, + 0.11066921800374985, + 0.16139380633831024, + 0.10258439183235168, + 0.02894996851682663, + 0.08693555742502213, + 0.022504299879074097 + ], + "credit_map": [ + 0.011442996561527252, + 0.011935772374272346, + 0.012422684580087662, + 0.016414176672697067, + 0.014708569273352623, + 0.01502496749162674, + 0.008596042171120644, + 0.010329894721508026, + 0.011591043323278427, + 0.014137745834887028, + 0.013593392446637154, + 0.012973622418940067, + 0.017004355788230896, + 0.02376045286655426, + 0.0062928409315645695, + 0.009491849690675735, + 0.019520848989486694, + 0.01996488682925701, + 0.03469478338956833, + 0.024700190871953964, + 0.022672822698950768, + 0.0252370648086071, + 0.01844983734190464, + 0.028705671429634094, + 0.007823396474123001, + 0.008341754786670208, + 0.010688685812056065, + 0.013009587302803993, + 0.008736889809370041, + 0.009616633877158165, + 0.007192970719188452, + 0.017886299639940262, + 0.012208034284412861, + 0.012561360374093056, + 0.012686566449701786, + 0.011524513363838196, + 0.012618942186236382, + 0.018552282825112343, + 0.009557364508509636, + 0.020960157737135887, + 0.02776309661567211, + 0.015831364318728447, + 0.012765506282448769, + 0.01632489077746868, + 0.013887666165828705, + 0.040243420749902725, + 0.016957703977823257, + 0.017620157450437546, + 0.012998402118682861, + 0.014760423451662064, + 0.006018076557666063, + 0.013100027106702328, + 0.01660732552409172, + 0.011368499137461185, + 0.007472710683941841, + 0.020258938893675804, + 0.002010548021644354, + 0.0028718996327370405, + 0.003440198954194784, + 0.004867780953645706, + 0.0033636363223195076, + 0.004632229916751385, + 0.0045532644726336, + 0.0032104116398841143, + 0.0063062990084290504, + 0.010472891852259636, + 0.010929773561656475, + 0.01091750804334879, + 0.017362797632813454, + 0.012545480392873287, + 0.009929586201906204, + 0.008471229113638401, + 0.002303720684722066, + 0.0024706749245524406, + 0.002220937516540289, + 0.004066114779561758, + 0.0018113788682967424, + 0.003380063222721219, + 0.0032983592245727777, + 0.0029530527535825968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.0067553711123764515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201635157689452 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201635157689452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201635157689452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09132295846939087, + 0.07580330222845078, + 0.1379786729812622, + 0.14514413475990295, + 0.0731358453631401, + 0.1658107489347458, + 0.10934749990701675, + 0.025251056998968124, + 0.12335164099931717, + 0.052854206413030624 + ], + "credit_map": [ + 0.00956802163273096, + 0.01551020983606577, + 0.009778689593076706, + 0.015727225691080093, + 0.015305344946682453, + 0.008744425140321255, + 0.007958054542541504, + 0.008730985224246979, + 0.00783576350659132, + 0.010699952952563763, + 0.008664620108902454, + 0.010122133418917656, + 0.011835269629955292, + 0.01570197381079197, + 0.004771484527736902, + 0.0061721052043139935, + 0.012961902655661106, + 0.01418108120560646, + 0.020715823397040367, + 0.01852736994624138, + 0.01668699085712433, + 0.019263185560703278, + 0.013259625993669033, + 0.022382685914635658, + 0.013367601670324802, + 0.015084420330822468, + 0.020457953214645386, + 0.023006144911050797, + 0.01711275801062584, + 0.016329970210790634, + 0.018109789118170738, + 0.02167549543082714, + 0.006650074385106564, + 0.007363858632743359, + 0.010702768340706825, + 0.010352847166359425, + 0.008035426028072834, + 0.010076941922307014, + 0.006402844563126564, + 0.013551086187362671, + 0.02301662601530552, + 0.01343156024813652, + 0.015332619659602642, + 0.01608775183558464, + 0.014977764338254929, + 0.04735137149691582, + 0.015982147306203842, + 0.019630905240774155, + 0.01635320857167244, + 0.020329341292381287, + 0.00766694126650691, + 0.014129381626844406, + 0.0171351321041584, + 0.00936042983084917, + 0.008380352519452572, + 0.01599270850419998, + 0.0020534207578748465, + 0.002233936684206128, + 0.0033349727746099234, + 0.004352118354290724, + 0.0029040665831416845, + 0.0038893972523510456, + 0.0039007687009871006, + 0.0025823747273534536, + 0.011763122864067554, + 0.013164867646992207, + 0.014625993557274342, + 0.014953725971281528, + 0.024298453703522682, + 0.017785856500267982, + 0.016097698360681534, + 0.010661929845809937, + 0.004002211149781942, + 0.005936628207564354, + 0.00539430882781744, + 0.008464178070425987, + 0.00547950342297554, + 0.008022867143154144, + 0.009128732606768608, + 0.0064257788471877575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.002201635157689452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903291821479797 + ], + "opposite_logprob_tokens": [ + -2.4640328884124756 + ], + "logprob_reward": -0.08903291821479797, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903291821479797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18852460384368896, + 0.09971708804368973, + 0.1328979730606079, + 0.166017085313797, + 0.11953434348106384, + 0.0650281235575676, + 0.055112309753894806, + 0.10103283077478409, + 0.0375051312148571, + 0.034630466252565384 + ], + "credit_map": [ + 0.016559451818466187, + 0.024531595408916473, + 0.025251690298318863, + 0.03221118822693825, + 0.030696436762809753, + 0.019681299105286598, + 0.015415262430906296, + 0.02417769283056259, + 0.00962841510772705, + 0.01478339359164238, + 0.013630220666527748, + 0.017264746129512787, + 0.010769842192530632, + 0.015847869217395782, + 0.007891599088907242, + 0.009900997392833233, + 0.012414376251399517, + 0.012950768694281578, + 0.019976051524281502, + 0.017585694789886475, + 0.0168189387768507, + 0.018617914989590645, + 0.011536827310919762, + 0.022997405380010605, + 0.017430325970053673, + 0.016292355954647064, + 0.021218687295913696, + 0.02624783106148243, + 0.02224627695977688, + 0.017220554873347282, + 0.017744621261954308, + 0.027616430073976517, + 0.014155561104416847, + 0.013540363870561123, + 0.015490147285163403, + 0.012877699919044971, + 0.014387043192982674, + 0.01620524190366268, + 0.010613074526190758, + 0.02226521447300911, + 0.0132603719830513, + 0.00653380760923028, + 0.003901478834450245, + 0.006975156720727682, + 0.006008894648402929, + 0.014903700910508633, + 0.005530614871531725, + 0.007914094254374504, + 0.00781503040343523, + 0.008984980173408985, + 0.004062188323587179, + 0.0077301692217588425, + 0.00778083223849535, + 0.005618782714009285, + 0.004293323960155249, + 0.008827004581689835, + 0.007448931690305471, + 0.012801303528249264, + 0.013580406084656715, + 0.01933210901916027, + 0.01269179955124855, + 0.011831898242235184, + 0.01334878895431757, + 0.009997595101594925, + 0.0028990742284804583, + 0.0052132923156023026, + 0.0027869935147464275, + 0.0047986749559640884, + 0.007018744479864836, + 0.004720708355307579, + 0.006229345686733723, + 0.0038382948841899633, + 0.003055304056033492, + 0.003861194010823965, + 0.002933567389845848, + 0.004089668393135071, + 0.004810195881873369, + 0.006460142787545919, + 0.005469617899507284, + 0.003950775600969791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08903291821479797 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048640698194503784 + ], + "opposite_logprob_tokens": [ + -3.048640727996826 + ], + "logprob_reward": -0.048640698194503784, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048640698194503784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16848374903202057, + 0.07940758019685745, + 0.16770373284816742, + 0.3568119704723358, + 0.035633742809295654, + 0.059825632721185684, + 0.04437682405114174, + 0.02856452204287052, + 0.04150368645787239, + 0.01768859289586544 + ], + "credit_map": [ + 0.019596688449382782, + 0.022932622581720352, + 0.01814892329275608, + 0.02804216556251049, + 0.02557029388844967, + 0.028157677501440048, + 0.011401118710637093, + 0.0146342683583498, + 0.008940123952925205, + 0.009818830527365208, + 0.01155336108058691, + 0.009373032487928867, + 0.010765976272523403, + 0.01833510398864746, + 0.004574903286993504, + 0.006046249996870756, + 0.017063677310943604, + 0.018188849091529846, + 0.025742491707205772, + 0.019679665565490723, + 0.01631108671426773, + 0.020269623026251793, + 0.01525578461587429, + 0.03519256412982941, + 0.034701455384492874, + 0.03433503210544586, + 0.05299330875277519, + 0.06092271953821182, + 0.04196707531809807, + 0.03844359889626503, + 0.03473389893770218, + 0.05871491879224777, + 0.003899875795468688, + 0.004228387959301472, + 0.004358700942248106, + 0.0047916229814291, + 0.004013632424175739, + 0.005155657418072224, + 0.0030722860246896744, + 0.0061135790310800076, + 0.010628185234963894, + 0.005229257047176361, + 0.0029124936554580927, + 0.00578048313036561, + 0.0039233071729540825, + 0.019488820806145668, + 0.004687611013650894, + 0.007175471633672714, + 0.005518885795027018, + 0.004297053441405296, + 0.004091891925781965, + 0.0048659127205610275, + 0.006742722820490599, + 0.006367999128997326, + 0.00523344287648797, + 0.007258916273713112, + 0.002486190525814891, + 0.002681044163182378, + 0.004238893743604422, + 0.004737216513603926, + 0.003263360820710659, + 0.003476691199466586, + 0.0040652137249708176, + 0.003615912515670061, + 0.0029799561016261578, + 0.0057536582462489605, + 0.0034144390374422073, + 0.005466470029205084, + 0.009386923164129257, + 0.005661836825311184, + 0.005264532286673784, + 0.0035758691374212503, + 0.0015620543854311109, + 0.0023173168301582336, + 0.00172783515881747, + 0.0022498550824820995, + 0.002455358859151602, + 0.0031758430413901806, + 0.002178022637963295, + 0.0020223066676408052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.048640698194503784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot, the camera slowly tilting upward, begins at the glass and steel spire of a modern skyscraper. Against a gloomy, stormy sky, a massive, geometrically shaped alien mothership silently and majestically tears through the thick, rolling gray storm clouds. The vessel's sleek, dark metal hull gleams with an unearthly, iridescent sheen in the dim skylight. Its surface is covered in intricate, self-luminous blue patterns, like a giant circuit board, its structure seemingly undergoing minute, silent reconfigurations as it moves. With a slow, inexorable momentum, it majestically sweeps past the skyscraper's spire, its massive shadow instantly engulfing the cityscape below. As the vessel's energy field distorts the air as it passes, creating subtle ripples in the reflections on the spire's mirrored glass facade. The scene is imbued with an epic scale and a breathtaking, awe-inspiring stillness, creating the feeling of a grand and tense science fiction film.", + "seed": 28764, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28764_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": -0.051669635251164435, + "vlm_logprob": -0.051669635251164435, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.051669635251164435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0388554148375988 + ], + "opposite_logprob_tokens": [ + -4.41385555267334 + ], + "logprob_reward": -0.0388554148375988, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0388554148375988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1823994219303131, + 0.1262352466583252, + 0.046486519277095795, + 0.10324016213417053, + 0.04935425892472267, + 0.07557591795921326, + 0.1499636322259903, + 0.1374330222606659, + 0.10156594961881638, + 0.02774585224688053 + ], + "credit_map": [ + 0.018028657883405685, + 0.03353051468729973, + 0.017617376521229744, + 0.026232212781906128, + 0.038194455206394196, + 0.013214513659477234, + 0.026992399245500565, + 0.00858928170055151, + 0.0147283636033535, + 0.024070652201771736, + 0.015065140090882778, + 0.02146858721971512, + 0.017180897295475006, + 0.01017817948013544, + 0.014771085232496262, + 0.008772340603172779, + 0.004710857290774584, + 0.006554997060447931, + 0.00507937790825963, + 0.007901785895228386, + 0.008122518658638, + 0.005426125600934029, + 0.005286023952066898, + 0.0034048266243189573, + 0.009494704194366932, + 0.015396012924611568, + 0.0106834527105093, + 0.017903417348861694, + 0.010204887017607689, + 0.012070529162883759, + 0.019284989684820175, + 0.008202160708606243, + 0.00469164876267314, + 0.008384820073843002, + 0.007028480060398579, + 0.005984014365822077, + 0.007153058424592018, + 0.0049846963956952095, + 0.004544089548289776, + 0.006583445239812136, + 0.007713275495916605, + 0.015666253864765167, + 0.009555824100971222, + 0.004824550822377205, + 0.011386086232960224, + 0.010807079263031483, + 0.008658323436975479, + 0.006964521482586861, + 0.013458142057061195, + 0.023696765303611755, + 0.038188207894563675, + 0.013941573910415173, + 0.01879764534533024, + 0.016768651083111763, + 0.014770827256143093, + 0.010341809131205082, + 0.02897503972053528, + 0.018468674272298813, + 0.017678262665867805, + 0.016620097681879997, + 0.016333362087607384, + 0.01103459857404232, + 0.017194896936416626, + 0.011128084734082222, + 0.022616194561123848, + 0.011598816141486168, + 0.015930842608213425, + 0.01336800679564476, + 0.01067391037940979, + 0.009995375759899616, + 0.007688418962061405, + 0.009694378823041916, + 0.004172831773757935, + 0.0037619126960635185, + 0.0033608528319746256, + 0.0040456210263073444, + 0.003131616860628128, + 0.0030902819707989693, + 0.0025697157252579927, + 0.0036130165681242943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0388554148375988 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020604262128472328 + ], + "opposite_logprob_tokens": [ + -3.895604372024536 + ], + "logprob_reward": -0.020604262128472328, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020604262128472328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13791432976722717, + 0.07059512287378311, + 0.05768265202641487, + 0.19099263846874237, + 0.08911547809839249, + 0.094627745449543, + 0.12833592295646667, + 0.0901162400841713, + 0.05273252725601196, + 0.08788732439279556 + ], + "credit_map": [ + 0.014018837362527847, + 0.03162320330739021, + 0.014199674129486084, + 0.0188438780605793, + 0.0298844613134861, + 0.008632454089820385, + 0.014863085001707077, + 0.005848735570907593, + 0.00790392979979515, + 0.010551754385232925, + 0.006900755688548088, + 0.007953105494379997, + 0.007962467148900032, + 0.008081823587417603, + 0.012729698792099953, + 0.00851159356534481, + 0.006789993494749069, + 0.01087555568665266, + 0.0062100752256810665, + 0.008445197716355324, + 0.0060521992854774, + 0.008235238492488861, + 0.00596397090703249, + 0.005110421217978001, + 0.016453692689538002, + 0.04191022366285324, + 0.018612883985042572, + 0.02693426050245762, + 0.021927811205387115, + 0.030918728560209274, + 0.021857284009456635, + 0.012377751059830189, + 0.008337716571986675, + 0.017274493351578712, + 0.008397593162953854, + 0.01065345574170351, + 0.016494644805788994, + 0.011618129909038544, + 0.009708425030112267, + 0.006631016731262207, + 0.011029472574591637, + 0.01985449530184269, + 0.012000140734016895, + 0.006299126427620649, + 0.011829768307507038, + 0.015982268378138542, + 0.008742950856685638, + 0.008889525197446346, + 0.012500698678195477, + 0.01998593471944332, + 0.032170865684747696, + 0.012462240643799305, + 0.013756883330643177, + 0.01312884408980608, + 0.013879568316042423, + 0.010450887493789196, + 0.018451998010277748, + 0.0124161746352911, + 0.010366681031882763, + 0.008503510616719723, + 0.013418330810964108, + 0.006095499265938997, + 0.013036244548857212, + 0.007827804423868656, + 0.012897374108433723, + 0.006535331718623638, + 0.008939963765442371, + 0.0067490520887076855, + 0.004937148652970791, + 0.003854918759316206, + 0.0037016840651631355, + 0.005117054563015699, + 0.00870093610137701, + 0.00947926752269268, + 0.010263715870678425, + 0.011043542064726353, + 0.009550979360938072, + 0.013434350490570068, + 0.011815325357019901, + 0.013599203899502754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.020604262128472328 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790733456611633 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790733456611633, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790733456611633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17461828887462616, + 0.115456223487854, + 0.04101601988077164, + 0.14704017341136932, + 0.14540550112724304, + 0.04507739469408989, + 0.08143831044435501, + 0.08553802222013474, + 0.04580496996641159, + 0.1186051294207573 + ], + "credit_map": [ + 0.015701154246926308, + 0.048197150230407715, + 0.016472427174448967, + 0.019982369616627693, + 0.0456039234995842, + 0.009495910257101059, + 0.013898346573114395, + 0.005267009139060974, + 0.015376927331089973, + 0.020296957343816757, + 0.014833739027380943, + 0.013265273533761501, + 0.016560638323426247, + 0.010265950113534927, + 0.01674068719148636, + 0.008116047829389572, + 0.0038826237432658672, + 0.006298388820141554, + 0.0049250503070652485, + 0.007276548538357019, + 0.005857379641383886, + 0.005084638949483633, + 0.004616144113242626, + 0.0030752450693398714, + 0.012783128768205643, + 0.032707057893276215, + 0.012853182852268219, + 0.019729770720005035, + 0.01799297332763672, + 0.02170795202255249, + 0.01750081218779087, + 0.011765292845666409, + 0.008549287915229797, + 0.044210948050022125, + 0.010223561897873878, + 0.009256239980459213, + 0.042932383716106415, + 0.009899651631712914, + 0.008790455758571625, + 0.011542967520654202, + 0.005122731439769268, + 0.007745987270027399, + 0.005451400298625231, + 0.004362465813755989, + 0.005566116888076067, + 0.005514190532267094, + 0.005996215157210827, + 0.005318287760019302, + 0.007091097068041563, + 0.012518519535660744, + 0.019959868863224983, + 0.008251813240349293, + 0.009029993787407875, + 0.009682750329375267, + 0.008646579459309578, + 0.006257683504372835, + 0.013455189764499664, + 0.012553568929433823, + 0.009687014855444431, + 0.007265730760991573, + 0.015034617856144905, + 0.005846334621310234, + 0.014270354993641376, + 0.007425215560942888, + 0.010183881036937237, + 0.005592942703515291, + 0.007058209273964167, + 0.005546228960156441, + 0.004368813242763281, + 0.0032710011582821608, + 0.005283475387841463, + 0.004500419832766056, + 0.01799219474196434, + 0.01148209348320961, + 0.014606120064854622, + 0.01625090092420578, + 0.011266443878412247, + 0.01940334402024746, + 0.011297431774437428, + 0.01630660705268383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.009790733456611633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705351412296295 + ], + "opposite_logprob_tokens": [ + -2.1270534992218018 + ], + "logprob_reward": -0.12705351412296295, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705351412296295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23439620435237885, + 0.30652984976768494, + 0.047093383967876434, + 0.10592066496610641, + 0.0657028779387474, + 0.07256052643060684, + 0.04683803766965866, + 0.04743693768978119, + 0.024992847815155983, + 0.04852869361639023 + ], + "credit_map": [ + 0.027051880955696106, + 0.054341185837984085, + 0.025437047705054283, + 0.027560854330658913, + 0.04994437098503113, + 0.011685981415212154, + 0.026378139853477478, + 0.011996755376458168, + 0.03458067402243614, + 0.039291802793741226, + 0.03436952829360962, + 0.028691213577985764, + 0.0426923967897892, + 0.03307293355464935, + 0.06138208508491516, + 0.032449204474687576, + 0.005397173576056957, + 0.007599423639476299, + 0.005446416791528463, + 0.008428414352238178, + 0.006816502194851637, + 0.004564289469271898, + 0.005233870353549719, + 0.0036072914954274893, + 0.009087393060326576, + 0.012817480601370335, + 0.011277892626821995, + 0.01597015932202339, + 0.01358315721154213, + 0.014356475323438644, + 0.016184963285923004, + 0.012643138878047466, + 0.006569809280335903, + 0.013594729825854301, + 0.005091547966003418, + 0.006527626421302557, + 0.013047157786786556, + 0.010194486938416958, + 0.004292626865208149, + 0.006384894717484713, + 0.008605900220572948, + 0.010848339647054672, + 0.008902796544134617, + 0.006452517118304968, + 0.012798543088138103, + 0.008263641968369484, + 0.00882459245622158, + 0.007864193990826607, + 0.004811418242752552, + 0.008465757593512535, + 0.015086960047483444, + 0.0032884760294109583, + 0.004260862711817026, + 0.0033025750890374184, + 0.0042801774106919765, + 0.003341813338920474, + 0.008945251815021038, + 0.004358778707683086, + 0.005647133570164442, + 0.005449419375509024, + 0.005238794721662998, + 0.0032910401932895184, + 0.010059799067676067, + 0.004446722567081451, + 0.005200920160859823, + 0.0028921496123075485, + 0.003659346606582403, + 0.003132137004286051, + 0.0020797448232769966, + 0.002352227922528982, + 0.002615957288071513, + 0.0030603630002588034, + 0.008144071325659752, + 0.004909229464828968, + 0.006276429630815983, + 0.007352842949330807, + 0.0046667614951729774, + 0.006713174749165773, + 0.004372439347207546, + 0.0060937413945794106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.12705351412296295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204425171017647 + ], + "opposite_logprob_tokens": [ + -2.812044143676758 + ], + "logprob_reward": -0.06204425171017647, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204425171017647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22036747634410858, + 0.2796846330165863, + 0.029232511296868324, + 0.0717465803027153, + 0.1609288603067398, + 0.07897438853979111, + 0.03219946101307869, + 0.047900453209877014, + 0.029064549133181572, + 0.04990106821060181 + ], + "credit_map": [ + 0.02826157957315445, + 0.04475725442171097, + 0.02598590776324272, + 0.03235767036676407, + 0.040578294545412064, + 0.012411938048899174, + 0.027344169095158577, + 0.008670663461089134, + 0.04070300608873367, + 0.05625312030315399, + 0.03601425513625145, + 0.027533916756510735, + 0.043012116104364395, + 0.020885877311229706, + 0.0370120070874691, + 0.018270354717969894, + 0.002708060434088111, + 0.003335322719067335, + 0.003761563217267394, + 0.0057436744682490826, + 0.0033892407082021236, + 0.002874620957300067, + 0.004228449426591396, + 0.0031915802974253893, + 0.005855366587638855, + 0.008457032032310963, + 0.009367839433252811, + 0.011209864169359207, + 0.009971567429602146, + 0.008981427177786827, + 0.010794143192470074, + 0.0071093435399234295, + 0.008227699436247349, + 0.056725967675447464, + 0.007889930158853531, + 0.005716939456760883, + 0.057840313762426376, + 0.011250684037804604, + 0.004618764854967594, + 0.008658564649522305, + 0.007370572537183762, + 0.013820257969200611, + 0.00866345688700676, + 0.005368801765143871, + 0.016220660880208015, + 0.012621045112609863, + 0.00816150102764368, + 0.006748093757778406, + 0.0033293443266302347, + 0.004790641367435455, + 0.00707098376005888, + 0.004336860962212086, + 0.004008997231721878, + 0.002551051788032055, + 0.0031752721406519413, + 0.002936308505013585, + 0.00969226099550724, + 0.00676214974373579, + 0.005722472909837961, + 0.004832115489989519, + 0.006981219630688429, + 0.0035314876586198807, + 0.005891657434403896, + 0.004487087018787861, + 0.005332299042493105, + 0.002437779912725091, + 0.004224488977342844, + 0.00304462481290102, + 0.0039695557206869125, + 0.0034476248547434807, + 0.002898498671129346, + 0.0037096773739904165, + 0.00805820245295763, + 0.0046167452819645405, + 0.006806339602917433, + 0.007462049834430218, + 0.004408482927829027, + 0.006790496874600649, + 0.004732661880552769, + 0.0070260874927043915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.06204425171017647 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic overhead shot, the camera slowly tilting upward, begins at the glass and steel spire of a modern skyscraper. Against a gloomy, stormy sky, a massive, geometrically shaped alien mothership silently and majestically tears through the thick, rolling gray storm clouds. The vessel's sleek, dark metal hull gleams with an unearthly, iridescent sheen in the dim skylight. Its surface is covered in intricate, self-luminous blue patterns, like a giant circuit board, its structure seemingly undergoing minute, silent reconfigurations as it moves. With a slow, inexorable momentum, it majestically sweeps past the skyscraper's spire, its massive shadow instantly engulfing the cityscape below. As the vessel's energy field distorts the air as it passes, creating subtle ripples in the reflections on the spire's mirrored glass facade. The scene is imbued with an epic scale and a breathtaking, awe-inspiring stillness, creating the feeling of a grand and tense science fiction film.", + "seed": 28765, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_1/index_1588_seed_28765_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": -0.3416666079312563, + "vlm_logprob": -0.3416666079312563, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.3416666079312563, + "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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5013396739959717 + ], + "opposite_logprob_tokens": [ + -1.0013396739959717 + ], + "logprob_reward": -0.5013396739959717, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5013396739959717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26796525716781616, + 0.09917692095041275, + 0.05681787058711052, + 0.11406456679105759, + 0.12237553298473358, + 0.08098652213811874, + 0.12823283672332764, + 0.07600735872983932, + 0.031219718977808952, + 0.023153390735387802 + ], + "credit_map": [ + 0.016908073797822, + 0.022222787141799927, + 0.08195068687200546, + 0.02728079818189144, + 0.016382483765482903, + 0.05694469064474106, + 0.020510762929916382, + 0.025764964520931244, + 0.008739985525608063, + 0.012580637820065022, + 0.013826022855937481, + 0.020872849971055984, + 0.01604503206908703, + 0.01047426462173462, + 0.009843554347753525, + 0.006794569082558155, + 0.004971964750438929, + 0.00776734109967947, + 0.008033121936023235, + 0.009636282920837402, + 0.009857827797532082, + 0.006548920646309853, + 0.005575973074883223, + 0.00442644115537405, + 0.00695739034563303, + 0.024335812777280807, + 0.011935382150113583, + 0.01630779542028904, + 0.024416832253336906, + 0.007562160026282072, + 0.014777052216231823, + 0.007772135082632303, + 0.008398754522204399, + 0.024466874077916145, + 0.01162116602063179, + 0.01592056080698967, + 0.021912971511483192, + 0.010788236744701862, + 0.018341923132538795, + 0.010925044305622578, + 0.007060730364173651, + 0.007489729207009077, + 0.011664717458188534, + 0.01419559121131897, + 0.01150067150592804, + 0.008693587966263294, + 0.01226020697504282, + 0.00812128372490406, + 0.00515045365318656, + 0.0045307353138923645, + 0.009612166322767735, + 0.003118036314845085, + 0.05017749220132828, + 0.045744188129901886, + 0.0042001474648714066, + 0.005699607077986002, + 0.007949696853756905, + 0.00942035112529993, + 0.014728575013577938, + 0.010252966545522213, + 0.007860035635530949, + 0.008670477196574211, + 0.007748070638626814, + 0.009377183392643929, + 0.003122973721474409, + 0.0021694274619221687, + 0.004622467327862978, + 0.0041154311038553715, + 0.004605742637068033, + 0.004866490606218576, + 0.0028125536628067493, + 0.004904633853584528, + 0.0023896254133433104, + 0.003038157010450959, + 0.0038340298924595118, + 0.0020283409394323826, + 0.002220909344032407, + 0.0032748307567089796, + 0.001960617955774069, + 0.004406880121678114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.5013396739959717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094762831926346 + ], + "opposite_logprob_tokens": [ + -4.14109468460083 + ], + "logprob_reward": -0.016094762831926346, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094762831926346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2492789775133133, + 0.03681711480021477, + 0.04379279911518097, + 0.07472722977399826, + 0.055069420486688614, + 0.03859088942408562, + 0.11579401046037674, + 0.20198866724967957, + 0.04831209406256676, + 0.1356288343667984 + ], + "credit_map": [ + 0.017383214086294174, + 0.033157721161842346, + 0.036209724843502045, + 0.029752960428595543, + 0.026343973353505135, + 0.0524507500231266, + 0.028189536184072495, + 0.025791067630052567, + 0.004244365729391575, + 0.005448095500469208, + 0.005071437917649746, + 0.005553834140300751, + 0.0050468179397284985, + 0.00481762457638979, + 0.0036203048657625914, + 0.0030146301724016666, + 0.004264228045940399, + 0.0048372624441981316, + 0.006989110726863146, + 0.007844372652471066, + 0.004794202744960785, + 0.004740123637020588, + 0.005815669894218445, + 0.004507830832153559, + 0.0043062735348939896, + 0.011837435886263847, + 0.008468867279589176, + 0.010920850560069084, + 0.0127010652795434, + 0.005646054167300463, + 0.014572381973266602, + 0.0062742955051362514, + 0.0026287464424967766, + 0.01605081744492054, + 0.003838407341390848, + 0.005427342839539051, + 0.015404061414301395, + 0.004785442259162664, + 0.004155260510742664, + 0.0027793392073363066, + 0.003809238551184535, + 0.003025553422048688, + 0.0051960120908916, + 0.004244140814989805, + 0.006504213437438011, + 0.006041723303496838, + 0.0053906189277768135, + 0.004379385616630316, + 0.009090348146855831, + 0.0062373243272304535, + 0.017681369557976723, + 0.0052935718558728695, + 0.03119257092475891, + 0.030007699504494667, + 0.008182474412024021, + 0.008108644746243954, + 0.01910216733813286, + 0.02605924941599369, + 0.04022039845585823, + 0.02608627825975418, + 0.017891930416226387, + 0.026920584961771965, + 0.020537961274385452, + 0.025170084089040756, + 0.004442145116627216, + 0.004610768985003233, + 0.013452274724841118, + 0.0048911673948168755, + 0.00452495226636529, + 0.00800689123570919, + 0.003550954395905137, + 0.004832936450839043, + 0.01527310162782669, + 0.01726585440337658, + 0.03361796587705612, + 0.012404924258589745, + 0.011233611963689327, + 0.02198096364736557, + 0.011612138710916042, + 0.012240263633430004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016094762831926346 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13294829428195953, + 0.07870125770568848, + 0.02979177236557007, + 0.2223948985338211, + 0.07656868547201157, + 0.04865006357431412, + 0.04444700479507446, + 0.05997344106435776, + 0.037910740822553635, + 0.26861393451690674 + ], + "credit_map": [ + 0.008546902798116207, + 0.014710078947246075, + 0.027336066588759422, + 0.01453588530421257, + 0.012852903455495834, + 0.03224034234881401, + 0.01206691563129425, + 0.010659191757440567, + 0.007953490130603313, + 0.0076783220283687115, + 0.016488565132021904, + 0.008291666395962238, + 0.00938477274030447, + 0.014212330803275108, + 0.005565588362514973, + 0.00912652537226677, + 0.0019000242464244366, + 0.0034536097664386034, + 0.0037584106903523207, + 0.0046257879585027695, + 0.0049610380083322525, + 0.003268598113209009, + 0.00500595523044467, + 0.0028183478862047195, + 0.012905986979603767, + 0.04894523322582245, + 0.02370942011475563, + 0.026448186486959457, + 0.04146980494260788, + 0.017752405256032944, + 0.03644361346960068, + 0.01472022570669651, + 0.0022541938815265894, + 0.028955446556210518, + 0.0023670452646911144, + 0.0029565421864390373, + 0.03052056021988392, + 0.004328043665736914, + 0.003178369952365756, + 0.0020084886346012354, + 0.004906353075057268, + 0.004580780863761902, + 0.0066405707038939, + 0.007877788506448269, + 0.0052336701191961765, + 0.0062852222472429276, + 0.0075795310549438, + 0.005546144209802151, + 0.00278011173941195, + 0.002521950053051114, + 0.004016165621578693, + 0.00339441723190248, + 0.011744413524866104, + 0.013760599307715893, + 0.0033665045630186796, + 0.002862843219190836, + 0.006380775012075901, + 0.006239307578653097, + 0.012252256274223328, + 0.00690945889800787, + 0.005511926021426916, + 0.007743057329207659, + 0.005991403944790363, + 0.008945253677666187, + 0.0037838774733245373, + 0.0034213135950267315, + 0.005347409751266241, + 0.0057783410884439945, + 0.0049570053815841675, + 0.004264255985617638, + 0.004618935752660036, + 0.005739602260291576, + 0.03389684855937958, + 0.023851333186030388, + 0.06534507125616074, + 0.031006891280412674, + 0.02136782556772232, + 0.04352893307805061, + 0.025960581377148628, + 0.023656459525227547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot an overhead view that slowly tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0540099143981934 + ], + "opposite_logprob_tokens": [ + -0.42900994420051575 + ], + "logprob_reward": -1.0540099143981934, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0540099143981934, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29735225439071655, + 0.08186060935258865, + 0.07105087488889694, + 0.1858353465795517, + 0.06386169791221619, + 0.06124144792556763, + 0.050312601029872894, + 0.07151538878679276, + 0.028842054307460785, + 0.0881277397274971 + ], + "credit_map": [ + 0.019101567566394806, + 0.026664210483431816, + 0.07918954640626907, + 0.037730906158685684, + 0.02290181629359722, + 0.046717219054698944, + 0.03635219857096672, + 0.02869478426873684, + 0.00913168303668499, + 0.008384795859456062, + 0.01379544660449028, + 0.010888788849115372, + 0.011274049989879131, + 0.013650395907461643, + 0.006562330760061741, + 0.008173121139407158, + 0.005690047051757574, + 0.006999834906309843, + 0.009491033852100372, + 0.012688009068369865, + 0.011979884468019009, + 0.008522940799593925, + 0.009225231595337391, + 0.006453897804021835, + 0.016955293715000153, + 0.024966983124613762, + 0.025578878819942474, + 0.023751338943839073, + 0.026303470134735107, + 0.020905980840325356, + 0.031209692358970642, + 0.016163703054189682, + 0.005803139880299568, + 0.01056136004626751, + 0.0070725963450968266, + 0.0051896292716264725, + 0.013080866076052189, + 0.008630246855318546, + 0.007904237136244774, + 0.005619617644697428, + 0.005440184846520424, + 0.006320782471448183, + 0.0090695945546031, + 0.010349858552217484, + 0.0063989413902163506, + 0.005759104620665312, + 0.011587281711399555, + 0.006315696984529495, + 0.0043029808439314365, + 0.0024591046385467052, + 0.0063790432177484035, + 0.003740625688806176, + 0.013541601598262787, + 0.013663292862474918, + 0.0031459422316402197, + 0.0030800069216638803, + 0.005660728085786104, + 0.006744077894836664, + 0.011423102580010891, + 0.007226092740893364, + 0.014561772346496582, + 0.013788418844342232, + 0.005506595596671104, + 0.006604600697755814, + 0.0027939961291849613, + 0.002009290037676692, + 0.007124411873519421, + 0.0025391501840204, + 0.003981349058449268, + 0.004771590698510408, + 0.0024506940972059965, + 0.0031715724617242813, + 0.011263654567301273, + 0.007721369154751301, + 0.02587057463824749, + 0.006953991949558258, + 0.007648907136172056, + 0.014737811870872974, + 0.005771269556134939, + 0.008160156197845936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -1.0540099143981934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, geometrically shaped alien mothership visible tearing through thick, rolling gray storm clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the mothership have a sleek, dark metal hull with self-luminous blue patterns resembling a circuit board?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the mothership sweep past a glass and steel spire of a modern skyscraper, casting a massive shadow over the city below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the passing vessel create an energy field that distorts the air and causes ripples in the reflections on the spire's mirrored glass facade?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709099054336548 + ], + "opposite_logprob_tokens": [ + -2.1270909309387207 + ], + "logprob_reward": -0.12709099054336548, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709099054336548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23044002056121826, + 0.11360510438680649, + 0.04543137922883034, + 0.26822149753570557, + 0.07897019386291504, + 0.0369446724653244, + 0.02937816083431244, + 0.052909642457962036, + 0.0446431040763855, + 0.0994563028216362 + ], + "credit_map": [ + 0.01817375421524048, + 0.033402979373931885, + 0.06375477463006973, + 0.02037312276661396, + 0.01964162290096283, + 0.03798481076955795, + 0.016538266092538834, + 0.020570652559399605, + 0.009263675659894943, + 0.015275854617357254, + 0.014943281188607216, + 0.014365192502737045, + 0.016999419778585434, + 0.02543221041560173, + 0.007469806354492903, + 0.009855661541223526, + 0.003297171788290143, + 0.007019811775535345, + 0.00806734524667263, + 0.007032296154648066, + 0.004770110361278057, + 0.0037019364535808563, + 0.006240679416805506, + 0.005302025005221367, + 0.017491266131401062, + 0.07659391313791275, + 0.022718189284205437, + 0.019889503717422485, + 0.06478305906057358, + 0.020572787150740623, + 0.02993742562830448, + 0.01623532921075821, + 0.0030977842397987843, + 0.02869775891304016, + 0.003248422872275114, + 0.004701533820480108, + 0.026237675920128822, + 0.0035017277114093304, + 0.006404160056263208, + 0.0030811179894953966, + 0.003607657738029957, + 0.0037683388218283653, + 0.0048146615736186504, + 0.007794630713760853, + 0.004128580912947655, + 0.002673040609806776, + 0.0065909200347959995, + 0.0035668383352458477, + 0.0022250707261264324, + 0.002371877897530794, + 0.0030213212594389915, + 0.003598883980885148, + 0.006826505996286869, + 0.0069043319672346115, + 0.002760721603408456, + 0.0016694430960342288, + 0.005443534813821316, + 0.00705377571284771, + 0.010358323343098164, + 0.007155792787671089, + 0.004991176538169384, + 0.006201855838298798, + 0.005594091489911079, + 0.006111090071499348, + 0.005105777643620968, + 0.003572068177163601, + 0.00845762062817812, + 0.0036962430458515882, + 0.006886155344545841, + 0.006997961550951004, + 0.003643728094175458, + 0.006283543538302183, + 0.012418688274919987, + 0.00717067439109087, + 0.030975213274359703, + 0.007704008370637894, + 0.00871400535106659, + 0.016648493707180023, + 0.006114631425589323, + 0.009710581041872501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.12709099054336548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_tensor([696, 696])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_tensor([696, 696])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..957a40df1e79565213091b1d32997d943075c185 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_tensor([696, 696])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, the camera seems fixed to the round headlight of a Porsche 911 rally car, shaking and vibrating violently as the car lurches across the track. The round glass lampshade is splattered with pixelated brown mud, while the car's vibrant red paint appears incredibly saturated under the bright, even, and nearly shadowless arcade lighting. Each violent jolt causes the entire image to momentarily lose focus before snapping back into focus, the camera following the dirt track composed of large, clearly visible square pixels, a blend of dark brown, ochre, and earthy yellow, forming a bumpy path. The entire world is rendered in a vibrant arcade art style, with color saturation pushed to extremes. The scenery on both sides of the track is stretched into horizontal, vibrant stripes of color at high speed, and all objects have a subtle, hard, pixelated outline around their edges, creating an atmosphere of high-speed, intense, and retro gaming excitement.", + "seed": 3367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3367_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": -0.11292338501662016, + "vlm_logprob": -0.11292338501662016, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.11292338501662016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13426946103572845 + ], + "opposite_logprob_tokens": [ + -2.1342694759368896 + ], + "logprob_reward": -0.13426946103572845, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13426946103572845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3512059152126312, + 0.15842093527317047, + 0.06568291038274765, + 0.034646954387426376, + 0.023893114179372787, + 0.03581356629729271, + 0.05778686702251434, + 0.05670364946126938, + 0.18192167580127716, + 0.033924344927072525 + ], + "credit_map": [ + 0.04043566435575485, + 0.027687707915902138, + 0.04330427944660187, + 0.049020569771528244, + 0.036340292543172836, + 0.04706000164151192, + 0.055097293108701706, + 0.0522601380944252, + 0.020420720800757408, + 0.008627371862530708, + 0.020194578915834427, + 0.02767019346356392, + 0.019768089056015015, + 0.01523720845580101, + 0.02272905968129635, + 0.02377372607588768, + 0.005797934718430042, + 0.0034519359469413757, + 0.009726885706186295, + 0.010226255282759666, + 0.009606418199837208, + 0.0078505277633667, + 0.00893761869519949, + 0.01008534524589777, + 0.00292788352817297, + 0.0026057299692183733, + 0.004388234578073025, + 0.005145746748894453, + 0.005399394314736128, + 0.003544558770954609, + 0.00536606227979064, + 0.005269348621368408, + 0.0030824055429548025, + 0.0015471660299226642, + 0.0025176084600389004, + 0.002879919484257698, + 0.003628215054050088, + 0.0024060357827693224, + 0.0038043621461838484, + 0.004027404356747866, + 0.005478159990161657, + 0.0021300234366208315, + 0.003464219393208623, + 0.004934431053698063, + 0.005692943464964628, + 0.003999808337539434, + 0.004916236735880375, + 0.005197749473154545, + 0.00919509306550026, + 0.0035298301372677088, + 0.007552148308604956, + 0.006675132550299168, + 0.007000178564339876, + 0.004721286240965128, + 0.0100244777277112, + 0.009088726714253426, + 0.006954994052648544, + 0.003950509242713451, + 0.007784391287714243, + 0.005975362844765186, + 0.010806453414261341, + 0.003625384531915188, + 0.00889330543577671, + 0.008713256567716599, + 0.017704328522086143, + 0.00852460227906704, + 0.01785716414451599, + 0.026492195203900337, + 0.026926521211862564, + 0.0146141042932868, + 0.04390554502606392, + 0.025897257030010223, + 0.0035634736996144056, + 0.0018408439354971051, + 0.002845345064997673, + 0.005363255273550749, + 0.006111884955316782, + 0.0035489050205796957, + 0.005824408493936062, + 0.0048262327909469604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.13426946103572845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054984647780656815 + ], + "opposite_logprob_tokens": [ + -2.9299845695495605 + ], + "logprob_reward": -0.054984647780656815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054984647780656815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3121839761734009, + 0.1296655535697937, + 0.0594974085688591, + 0.06961977481842041, + 0.05500095337629318, + 0.019854558631777763, + 0.03384172543883324, + 0.04644080623984337, + 0.2346845269203186, + 0.03921070694923401 + ], + "credit_map": [ + 0.032694753259420395, + 0.024094322696328163, + 0.03388051688671112, + 0.05443310737609863, + 0.0321471132338047, + 0.04207279533147812, + 0.04674827679991722, + 0.046113088726997375, + 0.011180676519870758, + 0.007477536331862211, + 0.015465890988707542, + 0.025558818131685257, + 0.017551764845848083, + 0.011122122406959534, + 0.02346407063305378, + 0.017844675108790398, + 0.00476563535630703, + 0.0029196639079600573, + 0.007832144387066364, + 0.008409064263105392, + 0.009228304959833622, + 0.006505436263978481, + 0.011113898828625679, + 0.008723261766135693, + 0.006510090082883835, + 0.004712310619652271, + 0.006948898080736399, + 0.009937510825693607, + 0.010178543627262115, + 0.007225851994007826, + 0.012947688810527325, + 0.011158882640302181, + 0.004294476471841335, + 0.002748520113527775, + 0.0040947506204247475, + 0.010308069176971912, + 0.00980914756655693, + 0.002885393798351288, + 0.012292800471186638, + 0.00856779981404543, + 0.0036717180628329515, + 0.0010883067734539509, + 0.0019143334357067943, + 0.0024239688646048307, + 0.0022749470081180334, + 0.002555340062826872, + 0.0028169627767056227, + 0.0031089833937585354, + 0.004369903821498156, + 0.002459451323375106, + 0.004689350258558989, + 0.004203402902930975, + 0.004449968226253986, + 0.004606378264725208, + 0.004517662338912487, + 0.0045456127263605595, + 0.004925163462758064, + 0.002610672963783145, + 0.006144634447991848, + 0.005221796687692404, + 0.008107640780508518, + 0.004018431529402733, + 0.007827726192772388, + 0.0075847432017326355, + 0.023184577003121376, + 0.012682759203016758, + 0.019838612526655197, + 0.02733953855931759, + 0.038032304495573044, + 0.017046721652150154, + 0.06116525083780289, + 0.03539476916193962, + 0.006161337718367577, + 0.002640525111928582, + 0.003727296832948923, + 0.005451614037156105, + 0.0065545267425477505, + 0.0034371190704405308, + 0.006172006484121084, + 0.0050662849098443985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.054984647780656815 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545477598905563 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545477598905563, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545477598905563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44768422842025757, + 0.12596239149570465, + 0.07238752394914627, + 0.051351148635149, + 0.0369088314473629, + 0.0518156960606575, + 0.03449331596493721, + 0.054943159222602844, + 0.07001098245382309, + 0.05444267392158508 + ], + "credit_map": [ + 0.05500918626785278, + 0.04253397136926651, + 0.06254922598600388, + 0.06116362288594246, + 0.03934725001454353, + 0.06654997915029526, + 0.05109657347202301, + 0.06943443417549133, + 0.010590989142656326, + 0.008312533609569073, + 0.019467981532216072, + 0.020812025293707848, + 0.014639466069638729, + 0.011755848303437233, + 0.019873734563589096, + 0.020509827882051468, + 0.005591684952378273, + 0.0034555611200630665, + 0.010604414157569408, + 0.011643740348517895, + 0.009799560531973839, + 0.006501308176666498, + 0.012667706236243248, + 0.012123558670282364, + 0.006061358377337456, + 0.0030677164904773235, + 0.005138164386153221, + 0.00677919015288353, + 0.007376622874289751, + 0.004746514372527599, + 0.010011778213083744, + 0.008169812150299549, + 0.005562315694987774, + 0.002103824634104967, + 0.004075579810887575, + 0.003492871532216668, + 0.006093280389904976, + 0.004426495172083378, + 0.005102134309709072, + 0.006052336189895868, + 0.006851361133158207, + 0.004058097489178181, + 0.007325081154704094, + 0.006130093242973089, + 0.006552977487444878, + 0.007982124574482441, + 0.0054162233136594296, + 0.0074997409246861935, + 0.00685833441093564, + 0.002420570468530059, + 0.004735354799777269, + 0.003945192787796259, + 0.0032316534779965878, + 0.0030291506554931402, + 0.004899989347904921, + 0.005373073276132345, + 0.005367042496800423, + 0.003603591350838542, + 0.009102452546358109, + 0.005929047241806984, + 0.007762139663100243, + 0.004924164619296789, + 0.009158173575997353, + 0.009096554480493069, + 0.006900168489664793, + 0.004263097885996103, + 0.006965488661080599, + 0.008960730396211147, + 0.0077232192270457745, + 0.007121343165636063, + 0.017383070662617683, + 0.010693863965570927, + 0.010000721551477909, + 0.0035553898196667433, + 0.005623804870992899, + 0.006987116765230894, + 0.007988063618540764, + 0.005261915735900402, + 0.00767464330419898, + 0.007351023610681295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012545477598905563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486209809780121 + ], + "opposite_logprob_tokens": [ + -1.2236210107803345 + ], + "logprob_reward": -0.3486209809780121, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486209809780121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.183935284614563, + 0.0926373302936554, + 0.04670381173491478, + 0.055182456970214844, + 0.08641588687896729, + 0.0375705286860466, + 0.045269954949617386, + 0.042163338512182236, + 0.3834397792816162, + 0.02668163552880287 + ], + "credit_map": [ + 0.02785707823932171, + 0.011594540439546108, + 0.032955847680568695, + 0.02617851085960865, + 0.019877688959240913, + 0.01461948174983263, + 0.021693432703614235, + 0.029158707708120346, + 0.008033119142055511, + 0.007134337909519672, + 0.010403887368738651, + 0.011735779233276844, + 0.011186831630766392, + 0.008315525949001312, + 0.020378056913614273, + 0.015449783764779568, + 0.0068216207437217236, + 0.0031240261159837246, + 0.00562394829466939, + 0.007328391540795565, + 0.004330350086092949, + 0.007025399710983038, + 0.006400718353688717, + 0.006049353629350662, + 0.006135511212050915, + 0.003824142273515463, + 0.008784265257418156, + 0.007503763772547245, + 0.005657824221998453, + 0.00550121022388339, + 0.00919915921986103, + 0.008576573804020882, + 0.007099303882569075, + 0.0042280093766748905, + 0.0074674333445727825, + 0.015236462466418743, + 0.014305206947028637, + 0.0047907354310154915, + 0.01995117776095867, + 0.013337556272745132, + 0.005184157285839319, + 0.001773769618012011, + 0.003632726613432169, + 0.005188348703086376, + 0.005412831902503967, + 0.004498002585023642, + 0.00680425763130188, + 0.005076427944004536, + 0.0040380386635661125, + 0.0031746961176395416, + 0.005009631160646677, + 0.0061339810490608215, + 0.00747228367254138, + 0.002645381959155202, + 0.010065984912216663, + 0.0067299543879926205, + 0.0034172083251178265, + 0.003293287707492709, + 0.004224652424454689, + 0.005759603343904018, + 0.008418393321335316, + 0.004432419314980507, + 0.006868537049740553, + 0.0057492367923259735, + 0.03249787166714668, + 0.0195616465061903, + 0.03364088758826256, + 0.051047779619693756, + 0.06042800843715668, + 0.022642413154244423, + 0.10259971767663956, + 0.06102144345641136, + 0.005596154369413853, + 0.0025940313935279846, + 0.0028819816652685404, + 0.0027065465692430735, + 0.0029965743888169527, + 0.002386529929935932, + 0.003715568920597434, + 0.003804245498031378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.3486209809780121 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196357689797878 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196357689797878, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196357689797878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3414038419723511, + 0.12106240540742874, + 0.1022573858499527, + 0.06631461530923843, + 0.040167152881622314, + 0.038607317954301834, + 0.058259546756744385, + 0.08032922446727753, + 0.10767412930727005, + 0.04392432048916817 + ], + "credit_map": [ + 0.04207371175289154, + 0.023262955248355865, + 0.03828766569495201, + 0.05660490691661835, + 0.03259296715259552, + 0.0461924783885479, + 0.05035219341516495, + 0.052036967128515244, + 0.015951184555888176, + 0.006810541730374098, + 0.015101092867553234, + 0.023647822439670563, + 0.012613930739462376, + 0.009933147579431534, + 0.01874573901295662, + 0.01825895346701145, + 0.010610273107886314, + 0.006044115871191025, + 0.013764873147010803, + 0.018075857311487198, + 0.012542092241346836, + 0.012061064131557941, + 0.013922843150794506, + 0.015236277133226395, + 0.006317076738923788, + 0.0042494479566812515, + 0.008254245854914188, + 0.009225443005561829, + 0.008424564264714718, + 0.006458651274442673, + 0.012697925791144371, + 0.010687267407774925, + 0.004617614671587944, + 0.002481220057234168, + 0.0037922589108347893, + 0.005097477696835995, + 0.006813634652644396, + 0.004101719241589308, + 0.007057319860905409, + 0.006205910351127386, + 0.009944378398358822, + 0.0022885564249008894, + 0.004780692979693413, + 0.0038146949373185635, + 0.003813656745478511, + 0.004582024645060301, + 0.00356360268779099, + 0.005819713696837425, + 0.008044982329010963, + 0.00415637344121933, + 0.008179064840078354, + 0.007710753008723259, + 0.006513406056910753, + 0.007505620829761028, + 0.008060173131525517, + 0.008089177310466766, + 0.007563937921077013, + 0.004979576915502548, + 0.007118588779121637, + 0.00904998742043972, + 0.01690865121781826, + 0.009115428663790226, + 0.01380196027457714, + 0.011791097931563854, + 0.013378730043768883, + 0.008350308053195477, + 0.01084089558571577, + 0.012799040414392948, + 0.014827090315520763, + 0.009447205811738968, + 0.02338475175201893, + 0.014646107330918312, + 0.006862589158117771, + 0.0025639997329562902, + 0.003938158508390188, + 0.005890903528779745, + 0.007495461963117123, + 0.003889908781275153, + 0.007367144338786602, + 0.005916159134358168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014196357689797878 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera seems fixed to the round headlight of a Porsche 911 rally car, shaking and vibrating violently as the car lurches across the track. The round glass lampshade is splattered with pixelated brown mud, while the car's vibrant red paint appears incredibly saturated under the bright, even, and nearly shadowless arcade lighting. Each violent jolt causes the entire image to momentarily lose focus before snapping back into focus, the camera following the dirt track composed of large, clearly visible square pixels, a blend of dark brown, ochre, and earthy yellow, forming a bumpy path. The entire world is rendered in a vibrant arcade art style, with color saturation pushed to extremes. The scenery on both sides of the track is stretched into horizontal, vibrant stripes of color at high speed, and all objects have a subtle, hard, pixelated outline around their edges, creating an atmosphere of high-speed, intense, and retro gaming excitement.", + "seed": 3368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3368_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": -0.11792111545801162, + "vlm_logprob": -0.11792111545801162, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.11792111545801162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0597248449921608 + ], + "opposite_logprob_tokens": [ + -3.434724807739258 + ], + "logprob_reward": -0.0597248449921608, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0597248449921608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037142008543014526, + 0.02420463226735592, + 0.03256228193640709, + 0.029831551015377045, + 0.015319897793233395, + 0.714569628238678, + 0.00864678155630827, + 0.006386726628988981, + 0.12075550109148026, + 0.010581000708043575 + ], + "credit_map": [ + 0.0034060273319482803, + 0.005520652513951063, + 0.004852361045777798, + 0.00403175363317132, + 0.004750297404825687, + 0.005818549543619156, + 0.004293679725378752, + 0.004468685481697321, + 0.00185029034037143, + 0.002250479767099023, + 0.005234047304838896, + 0.0039860219694674015, + 0.001603009644895792, + 0.0021117564756423235, + 0.002398578217253089, + 0.004770449362695217, + 0.005081895273178816, + 0.0023182055447250605, + 0.005112999584525824, + 0.0032361482735723257, + 0.003418160369619727, + 0.00548224663361907, + 0.003423927817493677, + 0.004488695878535509, + 0.0036896923556923866, + 0.0056915185414254665, + 0.00410036277025938, + 0.0037979239132255316, + 0.003670481964945793, + 0.0031221567187458277, + 0.002235756255686283, + 0.0035236577969044447, + 0.001512137707322836, + 0.0021679357159882784, + 0.0028257984668016434, + 0.0013952229637652636, + 0.0022491607815027237, + 0.0014603675808757544, + 0.0011991035426035523, + 0.002510170917958021, + 0.02789665386080742, + 0.01154233142733574, + 0.27492305636405945, + 0.021129092201590538, + 0.014453493058681488, + 0.042488571256399155, + 0.011841226369142532, + 0.3102951943874359, + 0.00130450667347759, + 0.0007115962216630578, + 0.0011165421456098557, + 0.0006682667881250381, + 0.0011259869206696749, + 0.0012590663973242044, + 0.0007914122543297708, + 0.0016694042133167386, + 0.0007745940238237381, + 0.0005259529571048915, + 0.0009327292791567743, + 0.0007317457348108292, + 0.0009062820463441312, + 0.0009428193443454802, + 0.0006354591459967196, + 0.0009371439809910953, + 0.0050801923498511314, + 0.004893414210528135, + 0.04075205326080322, + 0.008296371437609196, + 0.005559847690165043, + 0.00770193338394165, + 0.0050271619111299515, + 0.04344453290104866, + 0.0007746681221760809, + 0.000958252523560077, + 0.0024956234265118837, + 0.0011616663541644812, + 0.001153425546362996, + 0.001443884102627635, + 0.0007450978155247867, + 0.0018483821768313646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0597248449921608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20160658657550812 + ], + "opposite_logprob_tokens": [ + -1.7016066312789917 + ], + "logprob_reward": -0.20160658657550812, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20160658657550812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10055451840162277, + 0.01608632691204548, + 0.06596788763999939, + 0.0423409603536129, + 0.03466949611902237, + 0.6057283282279968, + 0.02262219786643982, + 0.02198546752333641, + 0.08095578104257584, + 0.009089023806154728 + ], + "credit_map": [ + 0.00993043277412653, + 0.014039726927876472, + 0.011981219984591007, + 0.009734328836202621, + 0.014536209404468536, + 0.015597785823047161, + 0.013175144791603088, + 0.011559665203094482, + 0.001582750235684216, + 0.002317905891686678, + 0.0022003643680363894, + 0.0014918859815225005, + 0.0019583296962082386, + 0.0020562349818646908, + 0.0017039959784597158, + 0.002774859545752406, + 0.009705735370516777, + 0.004594751633703709, + 0.009515475481748581, + 0.006548543460667133, + 0.007193420082330704, + 0.011836481280624866, + 0.007944156415760517, + 0.008629323914647102, + 0.004551697988063097, + 0.00787291582673788, + 0.006222536321729422, + 0.005240521393716335, + 0.005739935208112001, + 0.004409251268953085, + 0.0037532267160713673, + 0.00455087423324585, + 0.003150784410536289, + 0.006004516035318375, + 0.005335887894034386, + 0.0037504585925489664, + 0.0045027839951217175, + 0.004134648479521275, + 0.0028403173200786114, + 0.004950099159032106, + 0.023644207045435905, + 0.009585783816874027, + 0.2338130623102188, + 0.017906272783875465, + 0.01216179970651865, + 0.03474197909235954, + 0.0098425867035985, + 0.2640326917171478, + 0.0020640376023948193, + 0.002018204191699624, + 0.0028472752310335636, + 0.001981248613446951, + 0.004390760324895382, + 0.002901684492826462, + 0.0021352304611355066, + 0.0042837574146687984, + 0.003109424840658903, + 0.0016741144936531782, + 0.002963075879961252, + 0.0022219070233404636, + 0.002619042294099927, + 0.0027800549287348986, + 0.0024735727347433567, + 0.004144276957958937, + 0.005442297551780939, + 0.006798038724809885, + 0.0181739442050457, + 0.010023625567555428, + 0.0072464714758098125, + 0.007675741799175739, + 0.006590297445654869, + 0.019005361944437027, + 0.0006798660615459085, + 0.0011049869935959578, + 0.0016653559869155288, + 0.0013721627183258533, + 0.0008602086454629898, + 0.0010901920031756163, + 0.0006568722892552614, + 0.001659379224292934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.20160658657550812 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020591650158166885 + ], + "opposite_logprob_tokens": [ + -3.8955917358398438 + ], + "logprob_reward": -0.020591650158166885, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020591650158166885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04821882024407387, + 0.06362409144639969, + 0.14004148542881012, + 0.04468798264861107, + 0.041546911001205444, + 0.5420594215393066, + 0.026799701154232025, + 0.029070312157273293, + 0.05530388653278351, + 0.008647388778626919 + ], + "credit_map": [ + 0.005980520509183407, + 0.007936440408229828, + 0.006751824636012316, + 0.005924235563725233, + 0.005356040317565203, + 0.006484169978648424, + 0.004232416395097971, + 0.005553169641643763, + 0.004683401435613632, + 0.007424936629831791, + 0.017422733828425407, + 0.00816336553543806, + 0.003489731578156352, + 0.0028868902008980513, + 0.003112073289230466, + 0.016440965235233307, + 0.021288609132170677, + 0.009586554951965809, + 0.02382718026638031, + 0.014104803092777729, + 0.01227644830942154, + 0.02227165549993515, + 0.014831612817943096, + 0.021854620426893234, + 0.004740085452795029, + 0.008311630226671696, + 0.006903427187353373, + 0.005776877049356699, + 0.005912385880947113, + 0.004323808941990137, + 0.0038318384904414415, + 0.00488792872056365, + 0.0037646384444087744, + 0.005785135552287102, + 0.006416900549083948, + 0.003633196232840419, + 0.007539140991866589, + 0.004540643189102411, + 0.004202044568955898, + 0.005665214266628027, + 0.020848285406827927, + 0.008688932284712791, + 0.20914731919765472, + 0.016281086951494217, + 0.009931239299476147, + 0.03156227245926857, + 0.009463829919695854, + 0.23613645136356354, + 0.002512421226128936, + 0.002793719060719013, + 0.0033366421703249216, + 0.002352411625906825, + 0.003937275614589453, + 0.003889617742970586, + 0.0023280528839677572, + 0.005649559665471315, + 0.004489011596888304, + 0.0022638163063675165, + 0.0037530148401856422, + 0.0033146704081445932, + 0.0036707622930407524, + 0.0038186898455023766, + 0.003523459890857339, + 0.004236887209117413, + 0.005272202659398317, + 0.00601548096165061, + 0.007805311121046543, + 0.009213416837155819, + 0.006531420163810253, + 0.007049765437841415, + 0.006052965763956308, + 0.0073633198626339436, + 0.0007374078268185258, + 0.0010086024412885308, + 0.0018025052268058062, + 0.0014054913772270083, + 0.0008305067894980311, + 0.0010519000934436917, + 0.0005742856883443892, + 0.0012366895098239183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.020591650158166885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133270144462585 + ], + "opposite_logprob_tokens": [ + -1.4063327312469482 + ], + "logprob_reward": -0.28133270144462585, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133270144462585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018682757392525673, + 0.025704532861709595, + 0.02189534157514572, + 0.004584233742207289, + 0.008991297334432602, + 0.8817000985145569, + 0.005514601711183786, + 0.012064586393535137, + 0.018914297223091125, + 0.0019482222851365805 + ], + "credit_map": [ + 0.001346826204098761, + 0.0017839506035670638, + 0.003845375729724765, + 0.002358637284487486, + 0.0018344641430303454, + 0.0030560872983187437, + 0.0017540279077365994, + 0.0027033889200538397, + 0.002006963361054659, + 0.0032608327455818653, + 0.00440274877473712, + 0.0037539361510425806, + 0.002527763834223151, + 0.0027441729325801134, + 0.0022249920293688774, + 0.0047831228002905846, + 0.0037656333297491074, + 0.0015861536376178265, + 0.0037569839041680098, + 0.0021734191104769707, + 0.002173015382140875, + 0.0031455557327717543, + 0.0019260379485785961, + 0.003368541831150651, + 0.0004057681653648615, + 0.0005631005624309182, + 0.0010215325746685266, + 0.0006958502344787121, + 0.0005853719776496291, + 0.00040390260983258486, + 0.00040516917943023145, + 0.0005035383510403335, + 0.0008759896154515445, + 0.0015219158958643675, + 0.001164373243227601, + 0.0007029533735476434, + 0.0015966524370014668, + 0.0009741924586705863, + 0.0011990484781563282, + 0.0009561713086441159, + 0.03443086892366409, + 0.01409151591360569, + 0.3396146595478058, + 0.026032011955976486, + 0.01801401562988758, + 0.05172724649310112, + 0.014430237002670765, + 0.38335952162742615, + 0.000945913780014962, + 0.0005667724763043225, + 0.000774362706579268, + 0.0003603838267736137, + 0.000865795707795769, + 0.0005777747137472034, + 0.0004564804839901626, + 0.0009671177249401808, + 0.0012926602503284812, + 0.0012603150680661201, + 0.0017649090150371194, + 0.001265598344616592, + 0.001544426311738789, + 0.0017285184003412724, + 0.0010347526986151934, + 0.002173406770452857, + 0.000865599955432117, + 0.0005923811113461852, + 0.0062029254622757435, + 0.0008164204191416502, + 0.001054410357028246, + 0.001969132572412491, + 0.0007878139149397612, + 0.006625614129006863, + 0.0002271783450851217, + 0.00024205910449381918, + 0.00027937456616200507, + 0.0002784414100460708, + 0.00028042125632055104, + 0.0002272413403261453, + 0.00020437692000996321, + 0.00020912931358907372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.28133270144462585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04420800134539604, + 0.024206457659602165, + 0.078451968729496, + 0.03618709370493889, + 0.014472408220171928, + 0.6673873066902161, + 0.019171971827745438, + 0.0274930689483881, + 0.07214146107435226, + 0.016280170530080795 + ], + "credit_map": [ + 0.004969542380422354, + 0.005763949826359749, + 0.005287433974444866, + 0.0043189506977796555, + 0.00674087181687355, + 0.006688114255666733, + 0.004790617618709803, + 0.005648522637784481, + 0.002026011236011982, + 0.002986083272844553, + 0.00415390357375145, + 0.002218435052782297, + 0.002837294014170766, + 0.0028977494221180677, + 0.0025269831530749798, + 0.004559998866170645, + 0.011331568472087383, + 0.0055519649758934975, + 0.012332097627222538, + 0.007971320301294327, + 0.007616572082042694, + 0.013044084422290325, + 0.009311135858297348, + 0.01129322312772274, + 0.003974877763539553, + 0.006502160336822271, + 0.005427264608442783, + 0.004424307961016893, + 0.005182536318898201, + 0.003633304266259074, + 0.003146635601297021, + 0.003896007314324379, + 0.0020880368538200855, + 0.00196291645988822, + 0.001997979823499918, + 0.0015645119128748775, + 0.002077162731438875, + 0.00150671589653939, + 0.0014927516458556056, + 0.0017823332455009222, + 0.02573644183576107, + 0.010436169803142548, + 0.2584846317768097, + 0.020035473629832268, + 0.011738546192646027, + 0.03765406832098961, + 0.011365748941898346, + 0.29193630814552307, + 0.0022499782498925924, + 0.0012602281058207154, + 0.002843537600710988, + 0.0012764199636876583, + 0.002593200420960784, + 0.0032019848003983498, + 0.001441090484149754, + 0.004305531736463308, + 0.0045003206469118595, + 0.0019653597846627235, + 0.003888822393491864, + 0.002502131275832653, + 0.0028488743118941784, + 0.003291171044111252, + 0.002910022158175707, + 0.005586369428783655, + 0.004963243845850229, + 0.006259246729314327, + 0.015708686783909798, + 0.009153526276350021, + 0.006859608925879002, + 0.006878007669001818, + 0.005976928863674402, + 0.01634221337735653, + 0.0010379357263445854, + 0.0009965505450963974, + 0.004765698686242104, + 0.002712759654968977, + 0.0008084053988568485, + 0.0010811416432261467, + 0.0009620024356991053, + 0.003915676847100258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.02634979411959648 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera seems fixed to the round headlight of a Porsche 911 rally car, shaking and vibrating violently as the car lurches across the track. The round glass lampshade is splattered with pixelated brown mud, while the car's vibrant red paint appears incredibly saturated under the bright, even, and nearly shadowless arcade lighting. Each violent jolt causes the entire image to momentarily lose focus before snapping back into focus, the camera following the dirt track composed of large, clearly visible square pixels, a blend of dark brown, ochre, and earthy yellow, forming a bumpy path. The entire world is rendered in a vibrant arcade art style, with color saturation pushed to extremes. The scenery on both sides of the track is stretched into horizontal, vibrant stripes of color at high speed, and all objects have a subtle, hard, pixelated outline around their edges, creating an atmosphere of high-speed, intense, and retro gaming excitement.", + "seed": 3369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3369_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": -0.04937595445662737, + "vlm_logprob": -0.04937595445662737, + "vlm_yes_no_margin": 3.600000023841858, + "vlm_token_credit": -0.04937595445662737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05255616083741188 + ], + "opposite_logprob_tokens": [ + -3.677556276321411 + ], + "logprob_reward": -0.05255616083741188, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05255616083741188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05625825375318527, + 0.014077607542276382, + 0.02920326590538025, + 0.01627654768526554, + 0.1419859528541565, + 0.019974878057837486, + 0.012815646827220917, + 0.04509030655026436, + 0.644150972366333, + 0.02016659826040268 + ], + "credit_map": [ + 0.004617630969733, + 0.005726738832890987, + 0.0075695570558309555, + 0.005348704755306244, + 0.006778135895729065, + 0.010016890242695808, + 0.00792067963629961, + 0.008279912173748016, + 0.0015338357770815492, + 0.0015339655801653862, + 0.002562592038884759, + 0.0013232037890702486, + 0.0015597203746438026, + 0.0018340464448556304, + 0.0021173141431063414, + 0.0016129299765452743, + 0.002429206157103181, + 0.0013513071462512016, + 0.0021726798731833696, + 0.00415226211771369, + 0.0027580189052969217, + 0.0020598946139216423, + 0.008092457428574562, + 0.006187438033521175, + 0.0020885297562927008, + 0.0016865028301253915, + 0.0016156822675839067, + 0.0025453413836658, + 0.0021123853512108326, + 0.0013827190268784761, + 0.0025503961369395256, + 0.0022949916310608387, + 0.012786963023245335, + 0.025740696117281914, + 0.01094099972397089, + 0.014663204550743103, + 0.00797593966126442, + 0.00391964428126812, + 0.025502536445856094, + 0.04045598581433296, + 0.0019278355175629258, + 0.002097174758091569, + 0.003596050664782524, + 0.002547197276726365, + 0.0013873933348804712, + 0.0023791049607098103, + 0.0029796173330396414, + 0.003060505725443363, + 0.0024737617932260036, + 0.0015206042444333434, + 0.0015222918009385467, + 0.001090004574507475, + 0.0012713925680145621, + 0.0023756909649819136, + 0.0008500022813677788, + 0.0017118987161666155, + 0.0026482611428946257, + 0.0014264693018049002, + 0.0015112805413082242, + 0.0014940063701942563, + 0.015977418050169945, + 0.0030623581260442734, + 0.0161477942019701, + 0.0028227141592651606, + 0.012098397128283978, + 0.005882862489670515, + 0.014588011428713799, + 0.06383243948221207, + 0.017318639904260635, + 0.01043187826871872, + 0.28426745533943176, + 0.23573124408721924, + 0.002692691283300519, + 0.004006416071206331, + 0.0021984074264764786, + 0.001654094085097313, + 0.0022331043146550655, + 0.0033344817347824574, + 0.001628040918149054, + 0.002419363707304001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.05255616083741188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282505214214325 + ], + "opposite_logprob_tokens": [ + -2.017825126647949 + ], + "logprob_reward": -0.14282505214214325, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282505214214325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027947446331381798, + 0.004381587728857994, + 0.008988780900835991, + 0.007690508849918842, + 0.07709435373544693, + 0.008850856684148312, + 0.010400442406535149, + 0.00754210352897644, + 0.8413702249526978, + 0.00573366042226553 + ], + "credit_map": [ + 0.0029117255471646786, + 0.002997131086885929, + 0.00407354487106204, + 0.0029161199927330017, + 0.0033724410459399223, + 0.003461546963080764, + 0.004264059942215681, + 0.00395087618380785, + 0.0006262696115300059, + 0.0003725752467289567, + 0.0008074135985225439, + 0.0006492417305707932, + 0.00043596155592240393, + 0.0003909156657755375, + 0.0005619996227324009, + 0.0005372106097638607, + 0.0007287324988283217, + 0.0005006759311072528, + 0.0007819269667379558, + 0.0011037165531888604, + 0.000858670799061656, + 0.0005745951784774661, + 0.002518514171242714, + 0.0019219490932300687, + 0.0009191547869704664, + 0.000716155394911766, + 0.000847559014800936, + 0.001367484568618238, + 0.0006950684473849833, + 0.0007193544879555702, + 0.0010076462058350444, + 0.0014180863508954644, + 0.004688996355980635, + 0.009947222657501698, + 0.0028108975384384394, + 0.007390828803181648, + 0.0030619509052485228, + 0.00195232720579952, + 0.020033281296491623, + 0.027208851650357246, + 0.0008752014255151153, + 0.0010639801621437073, + 0.0018533567199483514, + 0.0007970004808157682, + 0.0007223061984404922, + 0.0008160651195794344, + 0.0014649360673502088, + 0.0012580102775245905, + 0.0023358326870948076, + 0.0011236123973503709, + 0.001083612092770636, + 0.0008779846248216927, + 0.0009056082344613969, + 0.0020708779338747263, + 0.0006311619654297829, + 0.0013717524707317352, + 0.001077507622539997, + 0.0004301748122088611, + 0.0009346617734991014, + 0.00040715717477723956, + 0.0008879544329829514, + 0.0007204785943031311, + 0.001025988720357418, + 0.002058180281892419, + 0.01537211611866951, + 0.007314973510801792, + 0.018931183964014053, + 0.08170446008443832, + 0.022831356152892113, + 0.013795861043035984, + 0.37104105949401855, + 0.3103792369365692, + 0.00037469191011041403, + 0.0003515729040373117, + 0.0009146327502094209, + 0.0005384914693422616, + 0.0007295041577890515, + 0.0014954829821363091, + 0.0005794080207124352, + 0.000749876257032156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.14282505214214325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632498629391193 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632498629391193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632498629391193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10097921639680862, + 0.06432220339775085, + 0.0573846735060215, + 0.08050698786973953, + 0.22116585075855255, + 0.09589632600545883, + 0.057853858917951584, + 0.07089252769947052, + 0.2037298083305359, + 0.047268521040678024 + ], + "credit_map": [ + 0.012275095097720623, + 0.009240335784852505, + 0.01277987752109766, + 0.010945284739136696, + 0.010120454244315624, + 0.017600972205400467, + 0.014911198988556862, + 0.013105994090437889, + 0.008132587186992168, + 0.007395269349217415, + 0.010334856808185577, + 0.006855091545730829, + 0.006813652813434601, + 0.007159560453146696, + 0.009801420383155346, + 0.007829765789210796, + 0.007263581268489361, + 0.005209694616496563, + 0.006865126080811024, + 0.0068509806878864765, + 0.004382556304335594, + 0.004086795262992382, + 0.011407899670302868, + 0.011318041011691093, + 0.007491119205951691, + 0.007618291303515434, + 0.008835350163280964, + 0.015147280879318714, + 0.008924884721636772, + 0.00660887872800231, + 0.011873763054609299, + 0.014007428660988808, + 0.019124945625662804, + 0.03716401010751724, + 0.05329868197441101, + 0.03829494118690491, + 0.01328338123857975, + 0.007983794435858727, + 0.024101614952087402, + 0.02791448123753071, + 0.009655303321778774, + 0.01076381653547287, + 0.0201624296605587, + 0.008471070788800716, + 0.008607328869402409, + 0.010278178378939629, + 0.014384690672159195, + 0.013573508709669113, + 0.009073133580386639, + 0.007025968749076128, + 0.004051804542541504, + 0.005748058669269085, + 0.0075701456516981125, + 0.011675558052957058, + 0.005131871905177832, + 0.00757732056081295, + 0.010279612615704536, + 0.003320450661703944, + 0.006967204622924328, + 0.00417688162997365, + 0.00964317936450243, + 0.008952844887971878, + 0.008882769383490086, + 0.01866958476603031, + 0.003505943575873971, + 0.0033353327307850122, + 0.006035500671714544, + 0.013349181972444057, + 0.006413034629076719, + 0.004323014989495277, + 0.08565202355384827, + 0.08111578226089478, + 0.0038729270454496145, + 0.00491903442889452, + 0.006587935145944357, + 0.00405103387311101, + 0.005036333110183477, + 0.011928626336157322, + 0.004232580307871103, + 0.006640047766268253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.007632498629391193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328772470355034 + ], + "opposite_logprob_tokens": [ + -3.773287773132324 + ], + "logprob_reward": -0.02328772470355034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328772470355034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04053575173020363, + 0.015195483341813087, + 0.02070735953748226, + 0.039627496153116226, + 0.43854618072509766, + 0.014596965163946152, + 0.02743152715265751, + 0.02959337644279003, + 0.3589596152305603, + 0.014806166291236877 + ], + "credit_map": [ + 0.006369824055582285, + 0.0037486820947378874, + 0.0055152117274701595, + 0.004579475615173578, + 0.0036105120088905096, + 0.005869278684258461, + 0.0048959217965602875, + 0.0059468504041433334, + 0.0014211148954927921, + 0.0014261399628594518, + 0.0024383242707699537, + 0.0016793729737401009, + 0.001833303482271731, + 0.003030891064554453, + 0.0017967293970286846, + 0.0015696086920797825, + 0.00229281815700233, + 0.0010508442064747214, + 0.0021584867499768734, + 0.0023410951253026724, + 0.0018474952084943652, + 0.0014696336584165692, + 0.0054811895824968815, + 0.004065799526870251, + 0.0038795899599790573, + 0.0029328642413020134, + 0.003315647365525365, + 0.00917118415236473, + 0.0027087440248578787, + 0.0032454803586006165, + 0.006692902185022831, + 0.007681087125092745, + 0.048448264598846436, + 0.09574023634195328, + 0.06317996233701706, + 0.05851472169160843, + 0.02701435796916485, + 0.013994422741234303, + 0.04494577646255493, + 0.08670849353075027, + 0.001439366489648819, + 0.0015628351829946041, + 0.0021231877617537975, + 0.0013522328808903694, + 0.001645736861974001, + 0.0018257491756230593, + 0.002025889465585351, + 0.0026219685096293688, + 0.004784478340297937, + 0.003076496534049511, + 0.002004696289077401, + 0.0029080724343657494, + 0.0034224523697048426, + 0.005010575987398624, + 0.0024377242662012577, + 0.0037870334926992655, + 0.003864115569740534, + 0.0025557030458003283, + 0.0031904522329568863, + 0.002066184999421239, + 0.004801016766577959, + 0.005193781573325396, + 0.004637554287910461, + 0.0032845709938555956, + 0.006661206018179655, + 0.003250762354582548, + 0.00809134729206562, + 0.03351133316755295, + 0.00962730310857296, + 0.0059179035015404224, + 0.15804803371429443, + 0.13385175168514252, + 0.0013125865953043103, + 0.0008972866926342249, + 0.0019583115354180336, + 0.0019489794503897429, + 0.0013619586825370789, + 0.0026202609296888113, + 0.0022107723634690046, + 0.002496011322364211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.02328772470355034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578335970640182 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578335970640182, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578335970640182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03874187543988228, + 0.024175407364964485, + 0.046083007007837296, + 0.015915874391794205, + 0.43615075945854187, + 0.020280729979276657, + 0.016249587759375572, + 0.027434948831796646, + 0.3569496273994446, + 0.018018221482634544 + ], + "credit_map": [ + 0.007030493579804897, + 0.0044943708926439285, + 0.005143244285136461, + 0.004293711856007576, + 0.004114293027669191, + 0.005377046298235655, + 0.004849294200539589, + 0.003439421532675624, + 0.0029955438803881407, + 0.003297435352578759, + 0.004709060303866863, + 0.002073577605187893, + 0.003055595327168703, + 0.0025230261962860823, + 0.002889323979616165, + 0.0026318449527025223, + 0.0032329210080206394, + 0.002538125729188323, + 0.003269697306677699, + 0.006395617965608835, + 0.004938025958836079, + 0.003466295078396797, + 0.012535279616713524, + 0.009707044810056686, + 0.002956039272248745, + 0.0013501535868272185, + 0.0020322606433182955, + 0.002272033831104636, + 0.001294471905566752, + 0.0016078978078439832, + 0.002221082104369998, + 0.0021819360554218292, + 0.03479181230068207, + 0.07281756401062012, + 0.040551137179136276, + 0.04677731916308403, + 0.02171330153942108, + 0.01298938225954771, + 0.08189409226179123, + 0.12461614608764648, + 0.001710229553282261, + 0.0025258378591388464, + 0.003937509376555681, + 0.0020731324329972267, + 0.0015993473352864385, + 0.002176600508391857, + 0.0029950453899800777, + 0.003263027872890234, + 0.0035491585731506348, + 0.001544772181659937, + 0.0014415920013561845, + 0.0014555968809872866, + 0.001851756009273231, + 0.002833455102518201, + 0.0014026644639670849, + 0.0021705911494791508, + 0.0037104396615177393, + 0.0011127432808279991, + 0.0026389132253825665, + 0.0008713257848285139, + 0.00599774019792676, + 0.0019291837234050035, + 0.004548765253275633, + 0.006625837180763483, + 0.006357340142130852, + 0.0029994328506290913, + 0.008083399385213852, + 0.03200387954711914, + 0.009645038284361362, + 0.0059877727180719376, + 0.15693630278110504, + 0.13493643701076508, + 0.0016073405276983976, + 0.0010022831847891212, + 0.0023715945426374674, + 0.0019307530019432306, + 0.0019183330005034804, + 0.00434698024764657, + 0.0021909940987825394, + 0.0026499414816498756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.020578335970640182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera seems fixed to the round headlight of a Porsche 911 rally car, shaking and vibrating violently as the car lurches across the track. The round glass lampshade is splattered with pixelated brown mud, while the car's vibrant red paint appears incredibly saturated under the bright, even, and nearly shadowless arcade lighting. Each violent jolt causes the entire image to momentarily lose focus before snapping back into focus, the camera following the dirt track composed of large, clearly visible square pixels, a blend of dark brown, ochre, and earthy yellow, forming a bumpy path. The entire world is rendered in a vibrant arcade art style, with color saturation pushed to extremes. The scenery on both sides of the track is stretched into horizontal, vibrant stripes of color at high speed, and all objects have a subtle, hard, pixelated outline around their edges, creating an atmosphere of high-speed, intense, and retro gaming excitement.", + "seed": 3370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_10/index_696_seed_3370_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": -0.060720712319016455, + "vlm_logprob": -0.060720712319016455, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.060720712319016455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03419117256999016 + ], + "opposite_logprob_tokens": [ + -4.159191131591797 + ], + "logprob_reward": -0.03419117256999016, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03419117256999016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24094587564468384, + 0.10202240198850632, + 0.10578222572803497, + 0.03365718200802803, + 0.14861612021923065, + 0.05925515294075012, + 0.024868084117770195, + 0.08325858414173126, + 0.06314965337514877, + 0.13844473659992218 + ], + "credit_map": [ + 0.03018675558269024, + 0.025577176362276077, + 0.03240175172686577, + 0.01992957293987274, + 0.0419427715241909, + 0.03162315860390663, + 0.032174184918403625, + 0.027110500261187553, + 0.010577659122645855, + 0.008518174290657043, + 0.01803659275174141, + 0.010966485366225243, + 0.01531253196299076, + 0.010053981095552444, + 0.017929906025528908, + 0.010627065785229206, + 0.017552955076098442, + 0.011608184315264225, + 0.021785641089081764, + 0.013565251603722572, + 0.006131469737738371, + 0.009711751714348793, + 0.01202055998146534, + 0.013406408950686455, + 0.0047324784100055695, + 0.004835538566112518, + 0.007034237962216139, + 0.0032642639707773924, + 0.002608424751088023, + 0.004306858405470848, + 0.003201023442670703, + 0.003674356499686837, + 0.017688827589154243, + 0.01240594033151865, + 0.02951200120151043, + 0.012499532662332058, + 0.01503918319940567, + 0.011272480711340904, + 0.02918357029557228, + 0.02101457677781582, + 0.010512998327612877, + 0.0054077645763754845, + 0.007954902946949005, + 0.006039738655090332, + 0.004142367281019688, + 0.005076205357909203, + 0.007926223799586296, + 0.012194951996207237, + 0.003184225643053651, + 0.002211535582318902, + 0.004470171872526407, + 0.0037404352333396673, + 0.0026558421086519957, + 0.0021048542112112045, + 0.0036110710352659225, + 0.0028899486642330885, + 0.010427427478134632, + 0.018699003383517265, + 0.00901397131383419, + 0.010533198714256287, + 0.008335779421031475, + 0.009251968003809452, + 0.009894408285617828, + 0.007102827541530132, + 0.006445751525461674, + 0.006587297655642033, + 0.009016386233270168, + 0.00854996033012867, + 0.008614052087068558, + 0.006265809293836355, + 0.00952084269374609, + 0.008149553090333939, + 0.013744045980274677, + 0.011217864230275154, + 0.014252502471208572, + 0.016588324680924416, + 0.019716626033186913, + 0.012902415357530117, + 0.030439110472798347, + 0.019583838060498238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03419117256999016 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032401978969574 + ], + "opposite_logprob_tokens": [ + -2.3503239154815674 + ], + "logprob_reward": -0.10032401978969574, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032401978969574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10894491523504257, + 0.0742376521229744, + 0.044943928718566895, + 0.10843463987112045, + 0.1953529268503189, + 0.01392386108636856, + 0.039074573665857315, + 0.29434168338775635, + 0.02323766052722931, + 0.09750815480947495 + ], + "credit_map": [ + 0.010386094450950623, + 0.010615584440529346, + 0.012127204798161983, + 0.010196817107498646, + 0.022248949855566025, + 0.010230287909507751, + 0.021126288920640945, + 0.012013684026896954, + 0.005893317051231861, + 0.00656670518219471, + 0.008316379971802235, + 0.008345866575837135, + 0.012758053839206696, + 0.009188375435769558, + 0.0151179488748312, + 0.008051009848713875, + 0.00605617044493556, + 0.004278060980141163, + 0.009130357764661312, + 0.00521802855655551, + 0.003395805135369301, + 0.004528984427452087, + 0.005891794338822365, + 0.006444727070629597, + 0.012186020612716675, + 0.011258364655077457, + 0.015686633065342903, + 0.014498899690806866, + 0.013802945613861084, + 0.01168801635503769, + 0.01640976406633854, + 0.012904004193842411, + 0.01705433987081051, + 0.014299186877906322, + 0.022827187553048134, + 0.021479789167642593, + 0.03574446961283684, + 0.016403397545218468, + 0.04507016763091087, + 0.022474393248558044, + 0.0021471783984452486, + 0.001165467663668096, + 0.001425927970558405, + 0.0012678985949605703, + 0.0010656174272298813, + 0.0014480575919151306, + 0.0018697649938985705, + 0.003533949377015233, + 0.0038923544343560934, + 0.004840000532567501, + 0.005205848254263401, + 0.003878170857205987, + 0.007456423249095678, + 0.004897215869277716, + 0.004959444515407085, + 0.003945119213312864, + 0.035646721720695496, + 0.07799355685710907, + 0.0281188003718853, + 0.04016566649079323, + 0.027233216911554337, + 0.03367847576737404, + 0.025842634961009026, + 0.02566261775791645, + 0.001926626544445753, + 0.0023154057562351227, + 0.003620858071371913, + 0.0035800584591925144, + 0.0026020773220807314, + 0.0025989331770688295, + 0.0038832759018987417, + 0.002710426691919565, + 0.008235805667936802, + 0.008192251436412334, + 0.00978719163686037, + 0.009684238582849503, + 0.017313536256551743, + 0.008092544041574001, + 0.024485237896442413, + 0.011717353016138077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.10032401978969574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104761406779289 + ], + "opposite_logprob_tokens": [ + -5.5041046142578125 + ], + "logprob_reward": -0.004104761406779289, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104761406779289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10822425782680511, + 0.04304050654172897, + 0.05225469544529915, + 0.05545779690146446, + 0.24250756204128265, + 0.02999808080494404, + 0.03197045996785164, + 0.33673903346061707, + 0.024388089776039124, + 0.07541950047016144 + ], + "credit_map": [ + 0.013622104190289974, + 0.010434480383992195, + 0.019770219922065735, + 0.011849812231957912, + 0.017583085224032402, + 0.010011069476604462, + 0.014508347027003765, + 0.01044513564556837, + 0.004557995591312647, + 0.004331208299845457, + 0.005897941999137402, + 0.004198012873530388, + 0.004240072797983885, + 0.00469161057844758, + 0.010844620876014233, + 0.004279040731489658, + 0.005862452555447817, + 0.007635672111064196, + 0.005630671977996826, + 0.006833203136920929, + 0.005163085181266069, + 0.005918153561651707, + 0.007122782990336418, + 0.008088676258921623, + 0.00718841515481472, + 0.005067468620836735, + 0.009876907803118229, + 0.007456944789737463, + 0.007021790370345116, + 0.004835520405322313, + 0.007188816554844379, + 0.006821935530751944, + 0.03626653924584389, + 0.03442322462797165, + 0.04881737753748894, + 0.030256200581789017, + 0.017392607405781746, + 0.021033208817243576, + 0.02852989174425602, + 0.025788508355617523, + 0.004320777487009764, + 0.0032655224204063416, + 0.0032464968971908092, + 0.0032163094729185104, + 0.002448351588100195, + 0.0030346496496349573, + 0.004421055316925049, + 0.006044916808605194, + 0.0037360023707151413, + 0.003548327600583434, + 0.0037373972591012716, + 0.0031122697982937098, + 0.005941009148955345, + 0.0036474161315709352, + 0.004447105806320906, + 0.0038009292911738157, + 0.04006332904100418, + 0.0886472836136818, + 0.03338999301195145, + 0.04731032997369766, + 0.03002835065126419, + 0.03940029442310333, + 0.02954401634633541, + 0.028355438262224197, + 0.0027423715218901634, + 0.002606760710477829, + 0.0030872924253344536, + 0.0040067341178655624, + 0.0032342311460524797, + 0.0018364894203841686, + 0.003934791777282953, + 0.0029394193552434444, + 0.008465951308608055, + 0.005373517517000437, + 0.0073778750374913216, + 0.008362041786313057, + 0.01292203739285469, + 0.00760162016376853, + 0.015503041446208954, + 0.009813415817916393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.004104761406779289 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603468805551529 + ], + "opposite_logprob_tokens": [ + -1.9103468656539917 + ], + "logprob_reward": -0.1603468805551529, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603468805551529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16682219505310059, + 0.12216747552156448, + 0.12538465857505798, + 0.11498554050922394, + 0.05694582313299179, + 0.01600581593811512, + 0.04596492275595665, + 0.2710530757904053, + 0.026173334568738937, + 0.05449705570936203 + ], + "credit_map": [ + 0.019807223230600357, + 0.016541821882128716, + 0.024692567065358162, + 0.015556911006569862, + 0.02648228220641613, + 0.023371579125523567, + 0.019753267988562584, + 0.020616555586457253, + 0.01785232685506344, + 0.009393290616571903, + 0.04035799205303192, + 0.014393175020813942, + 0.008260497823357582, + 0.01004038006067276, + 0.010337813757359982, + 0.011532007716596127, + 0.0161733515560627, + 0.011892817914485931, + 0.02776562049984932, + 0.014387552626430988, + 0.010799918323755264, + 0.013467015698552132, + 0.014059556648135185, + 0.016838831827044487, + 0.010742422193288803, + 0.011746379546821117, + 0.02392641268670559, + 0.015426560305058956, + 0.009558222256600857, + 0.013299115002155304, + 0.013802265748381615, + 0.016484182327985764, + 0.008623789064586163, + 0.006192936096340418, + 0.0074453409761190414, + 0.007753685582429171, + 0.007741743233054876, + 0.004194923676550388, + 0.007872051559388638, + 0.007121359929442406, + 0.0018651994178071618, + 0.002062353305518627, + 0.001677967607975006, + 0.0016461730701848865, + 0.002306215465068817, + 0.0016617279034107924, + 0.0033157027792185545, + 0.0014704769710078835, + 0.004575695376843214, + 0.005203349981456995, + 0.005807836074382067, + 0.007567779626697302, + 0.006713282782584429, + 0.005067109130322933, + 0.006435598712414503, + 0.004594273399561644, + 0.0340360663831234, + 0.0666152611374855, + 0.025927841663360596, + 0.0348099060356617, + 0.026325134560465813, + 0.02922852709889412, + 0.029284676536917686, + 0.024825716391205788, + 0.002920333994552493, + 0.002396794967353344, + 0.0038921304512768984, + 0.0035080474335700274, + 0.0035292450338602066, + 0.0021193118300288916, + 0.004671220667660236, + 0.0031362513545900583, + 0.006343346554785967, + 0.004239839501678944, + 0.007486862596124411, + 0.005950260441750288, + 0.008103884756565094, + 0.005080730188637972, + 0.009499840438365936, + 0.007792300544679165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.1603468805551529 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 focused on the round headlight of a Porsche 911 rally car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the round glass lampshade splattered with pixelated brown mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's paint vibrant red and saturated under bright, nearly shadowless lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the image exhibit violent shaking, focus loss, and horizontal striped motion blur consistent with high-speed movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in a vibrant arcade art style with hard pixelated outlines and a pixelated dirt track?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1091262474656105, + 0.07167385518550873, + 0.08130185306072235, + 0.07666738331317902, + 0.18734899163246155, + 0.011057280004024506, + 0.015389987267553806, + 0.39655736088752747, + 0.0102858766913414, + 0.04059118777513504 + ], + "credit_map": [ + 0.015738490968942642, + 0.009859174489974976, + 0.01619284600019455, + 0.011105663143098354, + 0.01517779752612114, + 0.01226248499006033, + 0.01321176253259182, + 0.015578017570078373, + 0.008347772993147373, + 0.006365105509757996, + 0.018659943714737892, + 0.010871818289160728, + 0.004601378459483385, + 0.007367250043898821, + 0.00820981152355671, + 0.007250769063830376, + 0.00954838003963232, + 0.008820177987217903, + 0.011724271811544895, + 0.011144725605845451, + 0.006881244480609894, + 0.008997529745101929, + 0.012435036711394787, + 0.01175048016011715, + 0.010171826928853989, + 0.0071459305472671986, + 0.014210475608706474, + 0.011287006549537182, + 0.008584578521549702, + 0.0062681520357728004, + 0.009288059547543526, + 0.009711341001093388, + 0.026649175211787224, + 0.02737349085509777, + 0.038275256752967834, + 0.023273969069123268, + 0.014475355856120586, + 0.016003094613552094, + 0.02160712704062462, + 0.019691532477736473, + 0.0020430355798453093, + 0.0010228623868897557, + 0.0012282247189432383, + 0.0010023392969742417, + 0.0007527020061388612, + 0.001020830008201301, + 0.0022044244688004255, + 0.001782862120307982, + 0.0023025544360280037, + 0.001865029102191329, + 0.0030835685320198536, + 0.0014128006296232343, + 0.001407610485330224, + 0.0014995909295976162, + 0.002270631492137909, + 0.0015481997979804873, + 0.047180596739053726, + 0.10384753346443176, + 0.03906179592013359, + 0.055086374282836914, + 0.03589558228850365, + 0.0456012524664402, + 0.035905707627534866, + 0.033978499472141266, + 0.0012738894438371062, + 0.0009902598103508353, + 0.0012602428905665874, + 0.0011476530926302075, + 0.0011778469197452068, + 0.0009527817601338029, + 0.002167726866900921, + 0.0013154748594388366, + 0.004857703577727079, + 0.0031885639764368534, + 0.00558227114379406, + 0.004853058140724897, + 0.005205598194152117, + 0.0038905302062630653, + 0.007561979349702597, + 0.005451480392366648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.004636727273464203 + } + ], + "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/0000022/group_11/index_tensor([9375, 9375])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_tensor([9375, 9375])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f8e3aefdea8d8ae6b5f26acef9db4c630bf8daf9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_tensor([9375, 9375])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, lingering close-up captures the portrait of strategist Zhou Yu, rendered entirely in the traditional and iconic ink painting style. Thick, decisive calligraphic brushstrokes saturated with deep black ink define the sharp edges and overlapping plates of his ornate battle armor, while lighter gray tones suggest its metallic sheen. His sharp, angular features are twisted into a grimace of deep, anguished frustration; his forehead is furrowed, his lips pressed into a thin, bloodless line, and the muscles in his jaw are visibly clenched in suppressed anger. His background, a rough, textured backdrop resembling rice paper, is painted with faint, abstract ink splashes, evoking a sense of inner turmoil. The lighting is high-contrast and dramatic, with the pure white of the unpainted paper creating harsh highlights on one side of his face and the sharp lines of his nose, while the other side is plunged into deep ink shadows. The camera remains completely still, holding this long, lingering shot, forcing the viewer to confront the raw, silent pain etched on the face, creating an intense and powerful psychological atmosphere filled with a deep sense of frustration.", + "seed": 29285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_9375_seed_29285_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": -0.15950329583138229, + "vlm_logprob": -0.15950329583138229, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.15950329583138229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.6413748264312744 + ], + "opposite_logprob_tokens": [ + -0.7663748264312744 + ], + "logprob_reward": -0.6413748264312744, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6413748264312744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08949502557516098, + 0.07605097442865372, + 0.04879692569375038, + 0.32361772656440735, + 0.03169610723853111, + 0.048681315034627914, + 0.04093964770436287, + 0.10240261256694794, + 0.1371770203113556, + 0.10114265233278275 + ], + "credit_map": [ + 0.011523697525262833, + 0.005450131371617317, + 0.02267635613679886, + 0.006298780906945467, + 0.00791983213275671, + 0.005346226505935192, + 0.024914775043725967, + 0.005365218501538038, + 0.011597604490816593, + 0.004182623233646154, + 0.01665177009999752, + 0.0032743592746555805, + 0.004345386289060116, + 0.0038897099439054728, + 0.02862345054745674, + 0.003486072877421975, + 0.005767833907157183, + 0.002841681009158492, + 0.010586186312139034, + 0.0026259494479745626, + 0.00398894352838397, + 0.0031339358538389206, + 0.017117634415626526, + 0.0027347595896571875, + 0.10802990943193436, + 0.02553042583167553, + 0.08287500590085983, + 0.02234542742371559, + 0.030950075015425682, + 0.01904163882136345, + 0.022118311375379562, + 0.012726940214633942, + 0.006627657450735569, + 0.004168351646512747, + 0.003983048722147942, + 0.003276955569162965, + 0.004436319228261709, + 0.0031989736016839743, + 0.003924807067960501, + 0.00207999418489635, + 0.006947536952793598, + 0.0038705840706825256, + 0.00789868738502264, + 0.0029825144447386265, + 0.0052208127453923225, + 0.003423477988690138, + 0.015404774807393551, + 0.0029329275712370872, + 0.00488358736038208, + 0.003075741697102785, + 0.011024300940334797, + 0.003277062438428402, + 0.003568678628653288, + 0.0028369356878101826, + 0.010041682980954647, + 0.0022316589020192623, + 0.014437312260270119, + 0.005010106600821018, + 0.031155521050095558, + 0.005684417672455311, + 0.006307943258434534, + 0.004305637441575527, + 0.030166255310177803, + 0.0053354171104729176, + 0.015041835606098175, + 0.008350593969225883, + 0.036648161709308624, + 0.008014842867851257, + 0.009076139889657497, + 0.005122005008161068, + 0.048391010612249374, + 0.006532431114464998, + 0.015252258628606796, + 0.006011425517499447, + 0.028447283431887627, + 0.007104429416358471, + 0.007090432103723288, + 0.006145046558231115, + 0.02489750273525715, + 0.006194268818944693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.6413748264312744 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.04869270324707031 + ], + "opposite_logprob_tokens": [ + -3.0486927032470703 + ], + "logprob_reward": -0.04869270324707031, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869270324707031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03379959985613823, + 0.03003634698688984, + 0.22686374187469482, + 0.05577990412712097, + 0.022620826959609985, + 0.01631172001361847, + 0.01742594875395298, + 0.10034314543008804, + 0.1020788922905922, + 0.39473989605903625 + ], + "credit_map": [ + 0.00835184846073389, + 0.0020479527302086353, + 0.0025524692609906197, + 0.0031402925960719585, + 0.0022546052932739258, + 0.0028425855562090874, + 0.008919723331928253, + 0.0036901221610605717, + 0.008935592137277126, + 0.003054526401683688, + 0.002556642284616828, + 0.0018884586170315742, + 0.0029341906774789095, + 0.002675087656825781, + 0.005720783025026321, + 0.002271064091473818, + 0.08446168154478073, + 0.01018257811665535, + 0.040952183306217194, + 0.013399099931120872, + 0.013780409470200539, + 0.014253735542297363, + 0.04125167801976204, + 0.008582374081015587, + 0.019866943359375, + 0.0033188832458108664, + 0.015427587553858757, + 0.0034698264207690954, + 0.0040113613940775394, + 0.003126186551526189, + 0.004221667535603046, + 0.0023374445736408234, + 0.0032808869145810604, + 0.0016138439532369375, + 0.003991236444562674, + 0.0013027683598920703, + 0.001995807746425271, + 0.0010199210373684764, + 0.008024963550269604, + 0.0013913981383666396, + 0.002223911928012967, + 0.0014817764749750495, + 0.0035199758131057024, + 0.0017101061530411243, + 0.0012386585585772991, + 0.000915036303922534, + 0.0041901362128555775, + 0.0010321185691282153, + 0.0030989754013717175, + 0.0023215655237436295, + 0.003614702494814992, + 0.001859518582932651, + 0.0015289910370483994, + 0.0013404272031039, + 0.002517678774893284, + 0.0011440899688750505, + 0.013152544386684895, + 0.004381429869681597, + 0.026641985401511192, + 0.006283849477767944, + 0.006310905795544386, + 0.004485915414988995, + 0.03400478512048721, + 0.005081736948341131, + 0.010404195636510849, + 0.0034461466129869223, + 0.03168226778507233, + 0.005031248088926077, + 0.0063714999705553055, + 0.003682216629385948, + 0.036919914186000824, + 0.004541401751339436, + 0.038907796144485474, + 0.012201694771647453, + 0.1264210194349289, + 0.018735144287347794, + 0.023599065840244293, + 0.01104618702083826, + 0.14776229858398438, + 0.01606668345630169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.04869270324707031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.08902855962514877 + ], + "opposite_logprob_tokens": [ + -2.4640285968780518 + ], + "logprob_reward": -0.08902855962514877, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902855962514877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12356843799352646, + 0.04955441132187843, + 0.1954483985900879, + 0.14313790202140808, + 0.06343092024326324, + 0.03968841955065727, + 0.03467243164777756, + 0.05207842215895653, + 0.05743992328643799, + 0.24098066985607147 + ], + "credit_map": [ + 0.014983093366026878, + 0.01270345225930214, + 0.017887812107801437, + 0.010899153538048267, + 0.015792055055499077, + 0.00998346135020256, + 0.03213237598538399, + 0.00918705016374588, + 0.012081706896424294, + 0.003761337138712406, + 0.013095421716570854, + 0.004479206167161465, + 0.004951253533363342, + 0.004155298229306936, + 0.004298406187444925, + 0.002731787972152233, + 0.07224229723215103, + 0.008767737075686455, + 0.03718235716223717, + 0.011752583086490631, + 0.011753016151487827, + 0.013035791926085949, + 0.03304389864206314, + 0.007670748047530651, + 0.01791287027299404, + 0.013387642800807953, + 0.022227704524993896, + 0.014609089121222496, + 0.022391313686966896, + 0.010533544234931469, + 0.033489394932985306, + 0.008586361072957516, + 0.01712377369403839, + 0.006450468208640814, + 0.01186416670680046, + 0.006568677723407745, + 0.006501412484794855, + 0.004931089002639055, + 0.006746949627995491, + 0.003244393737986684, + 0.009042607620358467, + 0.0031083119101822376, + 0.006223937030881643, + 0.003500143066048622, + 0.004510167986154556, + 0.0028745238669216633, + 0.006845039781183004, + 0.0035836934112012386, + 0.004733125679194927, + 0.004146951716393232, + 0.006220923271030188, + 0.002796330489218235, + 0.0035930287558585405, + 0.0020705314818769693, + 0.008524561300873756, + 0.002586984308436513, + 0.0049322848208248615, + 0.00308045232668519, + 0.013535820879042149, + 0.003406840842217207, + 0.004473178647458553, + 0.0031330841593444347, + 0.016478857025504112, + 0.0030379104427993298, + 0.007887744344770908, + 0.0031469978857785463, + 0.00989676546305418, + 0.004785981494933367, + 0.0044762687757611275, + 0.003687012242153287, + 0.020024746656417847, + 0.0035344150383025408, + 0.02895284630358219, + 0.008692456409335136, + 0.07376006990671158, + 0.01129373162984848, + 0.014518006704747677, + 0.007498135790228844, + 0.08612685650587082, + 0.010138601996004581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.08902855962514877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.008635676465928555 + ], + "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": [ + 0.08560557663440704, + 0.02184203267097473, + 0.08735513687133789, + 0.024062754586338997, + 0.02040201425552368, + 0.010871061123907566, + 0.009226936846971512, + 0.01622714102268219, + 0.03101409412920475, + 0.6933932304382324 + ], + "credit_map": [ + 0.013707984238862991, + 0.004455242771655321, + 0.013144699856638908, + 0.004520820453763008, + 0.006817635148763657, + 0.005360662471503019, + 0.03236803039908409, + 0.005230494309216738, + 0.0023711230605840683, + 0.001250491593964398, + 0.0053922743536531925, + 0.0013093143934383988, + 0.0019804968032985926, + 0.0011168476194143295, + 0.007198907434940338, + 0.0012225755490362644, + 0.031285278499126434, + 0.004003879614174366, + 0.013732129707932472, + 0.005377144552767277, + 0.005077891517430544, + 0.006043097469955683, + 0.018360113725066185, + 0.0034755985252559185, + 0.007222093641757965, + 0.0019945884123444557, + 0.005526875611394644, + 0.001986610936000943, + 0.002214635256677866, + 0.001285335049033165, + 0.0022700373083353043, + 0.0015625774394720793, + 0.0047270311042666435, + 0.0020317118614912033, + 0.002382694510743022, + 0.0024057740811258554, + 0.002093326533213258, + 0.0012529515661299229, + 0.003989045508205891, + 0.0015194781590253115, + 0.0014522928977385163, + 0.0011214326368644834, + 0.0026724652852863073, + 0.0009657178889028728, + 0.0008470692555420101, + 0.0006555577856488526, + 0.002272498793900013, + 0.0008840259397402406, + 0.001330021070316434, + 0.0016636750660836697, + 0.0011759380577132106, + 0.0012608792167156935, + 0.0012096803402528167, + 0.0009185121743939817, + 0.0010284458985552192, + 0.000639784149825573, + 0.002728231018409133, + 0.0011075169313699007, + 0.005103677976876497, + 0.0011240711901336908, + 0.0010183628182858229, + 0.0007444547372870147, + 0.0037083523347973824, + 0.0006924716290086508, + 0.00472259521484375, + 0.0020815564785152674, + 0.00842244178056717, + 0.0019638764206320047, + 0.0024465040769428015, + 0.001750432187691331, + 0.007947118021547794, + 0.0016795675037428737, + 0.06993155181407928, + 0.020840421319007874, + 0.21681994199752808, + 0.03280427306890488, + 0.04159598797559738, + 0.020893454551696777, + 0.2598522901535034, + 0.030655285343527794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 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 subject a portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.009784713387489319 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.009784713387489319, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784713387489319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06630579382181168, + 0.07609010487794876, + 0.059341929852962494, + 0.032674938440322876, + 0.01584913581609726, + 0.04187435656785965, + 0.030224250629544258, + 0.13588158786296844, + 0.28373005986213684, + 0.25802773237228394 + ], + "credit_map": [ + 0.01460185181349516, + 0.005262450315058231, + 0.00520630506798625, + 0.004439541604369879, + 0.005105944350361824, + 0.005942299496382475, + 0.017235124483704567, + 0.008512279950082302, + 0.026493461802601814, + 0.004517815075814724, + 0.016010845080018044, + 0.004588447976857424, + 0.005440708249807358, + 0.004688955843448639, + 0.01094660721719265, + 0.0034032755065709352, + 0.019034745171666145, + 0.003798407269641757, + 0.004879120271652937, + 0.0038530805613845587, + 0.004804715979844332, + 0.003968273289501667, + 0.016061242669820786, + 0.002942353254184127, + 0.009712784551084042, + 0.0032349522225558758, + 0.005558960605412722, + 0.0025791802909225225, + 0.0034863301552832127, + 0.002719477517530322, + 0.003163594752550125, + 0.002219661371782422, + 0.00232389522716403, + 0.0017017341451719403, + 0.002527795499190688, + 0.001797524164430797, + 0.002063729567453265, + 0.0014556354144588113, + 0.002628907561302185, + 0.001349915866740048, + 0.006566420663148165, + 0.0033986042253673077, + 0.006668156012892723, + 0.003696967149153352, + 0.0032781714107841253, + 0.003678339533507824, + 0.011845177039504051, + 0.0027425221633166075, + 0.005485274363309145, + 0.0026581576094031334, + 0.00539960153400898, + 0.0025320625863969326, + 0.002660585567355156, + 0.00193599087651819, + 0.0076532503589987755, + 0.0018993283156305552, + 0.014490853063762188, + 0.006494052708148956, + 0.036087766289711, + 0.0067780460231006145, + 0.008953003212809563, + 0.006082476116716862, + 0.050268322229385376, + 0.006727077532559633, + 0.02954312600195408, + 0.009829793125391006, + 0.0848207175731659, + 0.01341189630329609, + 0.017706742510199547, + 0.010185166262090206, + 0.10523868352174759, + 0.012993957847356796, + 0.033568914979696274, + 0.015168456360697746, + 0.07006832212209702, + 0.015554841607809067, + 0.017858780920505524, + 0.013371669687330723, + 0.08056624233722687, + 0.011870544403791428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.009784713387489319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, lingering close-up captures the portrait of strategist Zhou Yu, rendered entirely in the traditional and iconic ink painting style. Thick, decisive calligraphic brushstrokes saturated with deep black ink define the sharp edges and overlapping plates of his ornate battle armor, while lighter gray tones suggest its metallic sheen. His sharp, angular features are twisted into a grimace of deep, anguished frustration; his forehead is furrowed, his lips pressed into a thin, bloodless line, and the muscles in his jaw are visibly clenched in suppressed anger. His background, a rough, textured backdrop resembling rice paper, is painted with faint, abstract ink splashes, evoking a sense of inner turmoil. The lighting is high-contrast and dramatic, with the pure white of the unpainted paper creating harsh highlights on one side of his face and the sharp lines of his nose, while the other side is plunged into deep ink shadows. The camera remains completely still, holding this long, lingering shot, forcing the viewer to confront the raw, silent pain etched on the face, creating an intense and powerful psychological atmosphere filled with a deep sense of frustration.", + "seed": 29286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_9375_seed_29286_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": -0.2776198510080576, + "vlm_logprob": -0.2776198510080576, + "vlm_yes_no_margin": 2.175000011920929, + "vlm_token_credit": -0.2776198510080576, + "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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0594258308410645 + ], + "opposite_logprob_tokens": [ + -0.43442589044570923 + ], + "logprob_reward": -1.0594258308410645, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0594258308410645, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2889508008956909, + 0.11243576556444168, + 0.18852248787879944, + 0.048546191304922104, + 0.04706878960132599, + 0.08321958780288696, + 0.055828411132097244, + 0.048631370067596436, + 0.07036109268665314, + 0.056435439735651016 + ], + "credit_map": [ + 0.04646453261375427, + 0.04106398671865463, + 0.09551346302032471, + 0.029030457139015198, + 0.027737772092223167, + 0.014620324596762657, + 0.019310787320137024, + 0.015209497883915901, + 0.021332217380404472, + 0.018328676000237465, + 0.0179592352360487, + 0.010034698061645031, + 0.014065468683838844, + 0.007843729108572006, + 0.007408428005874157, + 0.015463323332369328, + 0.0657341405749321, + 0.014400853775441647, + 0.019046587869524956, + 0.018919022753834724, + 0.006644205655902624, + 0.010004987940192223, + 0.010418740101158619, + 0.04335397109389305, + 0.007038269657641649, + 0.006386390887200832, + 0.009572191163897514, + 0.007195168174803257, + 0.006708216853439808, + 0.004132645670324564, + 0.004607053007930517, + 0.002906261943280697, + 0.006383578758686781, + 0.007414110936224461, + 0.01054088119417429, + 0.007398225367069244, + 0.0055438983254134655, + 0.0033987786155194044, + 0.003575267968699336, + 0.002814055886119604, + 0.032422054558992386, + 0.02477698028087616, + 0.005973910912871361, + 0.004485752433538437, + 0.005252467468380928, + 0.001995037542656064, + 0.0030598086304962635, + 0.005253585521131754, + 0.006018565967679024, + 0.006888326723128557, + 0.00910148024559021, + 0.0071798050776124, + 0.0128738172352314, + 0.00532662495970726, + 0.005416748113930225, + 0.003023045137524605, + 0.007756934501230717, + 0.005490258801728487, + 0.014528986066579819, + 0.005845782347023487, + 0.006473185494542122, + 0.0026900204829871655, + 0.0032503639813512564, + 0.002595841186121106, + 0.012232664041221142, + 0.008455866016447544, + 0.021326927468180656, + 0.009742080233991146, + 0.008049254305660725, + 0.0033492932561784983, + 0.004309941083192825, + 0.0028950755950063467, + 0.007979141548275948, + 0.009422294795513153, + 0.011076829396188259, + 0.008358067832887173, + 0.008543715812265873, + 0.0034933125134557486, + 0.0046362788416445255, + 0.0029258050490170717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -1.0594258308410645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.2255755066871643 + ], + "opposite_logprob_tokens": [ + -1.6005754470825195 + ], + "logprob_reward": -0.2255755066871643, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255755066871643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10100407153367996, + 0.07616481930017471, + 0.0708022266626358, + 0.042177993804216385, + 0.29330408573150635, + 0.11830345541238785, + 0.012883160263299942, + 0.04081128165125847, + 0.18241675198078156, + 0.062132202088832855 + ], + "credit_map": [ + 0.01806396059691906, + 0.01150154322385788, + 0.028694676235318184, + 0.013579877093434334, + 0.011185484938323498, + 0.005695379804819822, + 0.004606328438967466, + 0.007676821667701006, + 0.022747768089175224, + 0.024041296914219856, + 0.011989590711891651, + 0.005745536647737026, + 0.00486033083871007, + 0.0021347200963646173, + 0.0026465828996151686, + 0.001998982857912779, + 0.009907105937600136, + 0.023549091070890427, + 0.01023754384368658, + 0.006578714586794376, + 0.005903916899114847, + 0.0027203240897506475, + 0.003688857425004244, + 0.008216668851673603, + 0.006201745476573706, + 0.00541540514677763, + 0.01044427789747715, + 0.007288418710231781, + 0.005314989481121302, + 0.0021826960146427155, + 0.003467222908511758, + 0.0018632367718964815, + 0.06904973089694977, + 0.02998024970293045, + 0.10007379949092865, + 0.04583991318941116, + 0.016449173912405968, + 0.00912175327539444, + 0.014962077140808105, + 0.00782736949622631, + 0.027645722031593323, + 0.016238391399383545, + 0.017938900738954544, + 0.012409090995788574, + 0.005783505737781525, + 0.005901598371565342, + 0.006642347667366266, + 0.0257438812404871, + 0.0016319951973855495, + 0.002877174410969019, + 0.001905182609334588, + 0.0020460239611566067, + 0.0016508458647876978, + 0.000916816876269877, + 0.0010782371973618865, + 0.0007768827490508556, + 0.006890381686389446, + 0.0043092817068099976, + 0.009466578252613544, + 0.0068419682793319225, + 0.0031472959090024233, + 0.0024036241229623556, + 0.00317892013117671, + 0.004573224578052759, + 0.045063335448503494, + 0.02144325152039528, + 0.054703276604413986, + 0.02263590879738331, + 0.015462171286344528, + 0.007639703340828419, + 0.00899454578757286, + 0.006474555004388094, + 0.01283639669418335, + 0.010366636328399181, + 0.011924995109438896, + 0.010940407402813435, + 0.005321696400642395, + 0.003098459914326668, + 0.004130586516112089, + 0.0035130211617797613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.2255755066871643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.033697985112667084 + ], + "opposite_logprob_tokens": [ + -3.408698081970215 + ], + "logprob_reward": -0.033697985112667084, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697985112667084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21073821187019348, + 0.027383290231227875, + 0.2996189594268799, + 0.027332114055752754, + 0.12765222787857056, + 0.07045686990022659, + 0.13511940836906433, + 0.029039449989795685, + 0.024628128856420517, + 0.04803137853741646 + ], + "credit_map": [ + 0.026068558916449547, + 0.02929401583969593, + 0.05109764635562897, + 0.024567216634750366, + 0.02871119976043701, + 0.014633630402386189, + 0.01796489953994751, + 0.018401028588414192, + 0.005303544923663139, + 0.006235840730369091, + 0.003777481848374009, + 0.0027721580117940903, + 0.0038584109861403704, + 0.0015547111397609115, + 0.0020685680210590363, + 0.0018125720089301467, + 0.07911419123411179, + 0.04493606463074684, + 0.02711433358490467, + 0.029990563169121742, + 0.014071502722799778, + 0.015457667410373688, + 0.016622072085738182, + 0.07231254875659943, + 0.003265922889113426, + 0.0036934115923941135, + 0.007216520141810179, + 0.0030951451044529676, + 0.0034617322962731123, + 0.002493061823770404, + 0.0022741188295185566, + 0.0018322003306820989, + 0.03355760872364044, + 0.009576546028256416, + 0.0470869280397892, + 0.012955390848219395, + 0.010137324221432209, + 0.004227162338793278, + 0.00579792307689786, + 0.0043133278377354145, + 0.008255324326455593, + 0.025202324613928795, + 0.005855503026396036, + 0.0061430553905665874, + 0.004401408135890961, + 0.003145960858091712, + 0.003251057118177414, + 0.014202226884663105, + 0.019746417179703712, + 0.016475440934300423, + 0.02551238238811493, + 0.013893705792725086, + 0.027290593832731247, + 0.012400172650814056, + 0.010613945312798023, + 0.009186736308038235, + 0.0034751291386783123, + 0.0028955417219549417, + 0.004774364642798901, + 0.0026343981735408306, + 0.006912827026098967, + 0.0031177240889519453, + 0.002832293976098299, + 0.0023971691261976957, + 0.00418315501883626, + 0.003372280392795801, + 0.004678705241531134, + 0.0022921343334019184, + 0.004399924073368311, + 0.0016784786712378263, + 0.001826631254516542, + 0.0021968171931803226, + 0.0076178149320185184, + 0.006391739472746849, + 0.007918860763311386, + 0.006172879133373499, + 0.009580397047102451, + 0.004027124028652906, + 0.0032279600854963064, + 0.0030945991165935993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.033697985112667084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.02634189836680889 + ], + "opposite_logprob_tokens": [ + -3.6513419151306152 + ], + "logprob_reward": -0.02634189836680889, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634189836680889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12451499700546265, + 0.05852721631526947, + 0.07277495414018631, + 0.038359347730875015, + 0.28859567642211914, + 0.13583484292030334, + 0.024960152804851532, + 0.08354423195123672, + 0.06939768046140671, + 0.1034909039735794 + ], + "credit_map": [ + 0.017592323943972588, + 0.015268117189407349, + 0.03580934554338455, + 0.01342840027064085, + 0.016936667263507843, + 0.007445712108165026, + 0.00928702112287283, + 0.008747398853302002, + 0.01810196228325367, + 0.006981828715652227, + 0.01209221687167883, + 0.005916151683777571, + 0.0035221856087446213, + 0.002590290969237685, + 0.002560299588367343, + 0.006762281991541386, + 0.012925447896122932, + 0.010702019557356834, + 0.0165345910936594, + 0.007896525785326958, + 0.003551312955096364, + 0.0033180012833327055, + 0.00418749637901783, + 0.013659561052918434, + 0.006056904327124357, + 0.004882668145000935, + 0.010762262158095837, + 0.005442642141133547, + 0.004159783944487572, + 0.0025499367620795965, + 0.0028001065365970135, + 0.0017050427850335836, + 0.07189144939184189, + 0.025301555171608925, + 0.11929456144571304, + 0.03661995753645897, + 0.012463407590985298, + 0.006710745394229889, + 0.009794637560844421, + 0.006519364193081856, + 0.02993847243487835, + 0.05047408118844032, + 0.013114272616803646, + 0.010995065793395042, + 0.01002692524343729, + 0.0033866302110254765, + 0.004798269830644131, + 0.013101131655275822, + 0.0035935640335083008, + 0.0036716428585350513, + 0.005763487424701452, + 0.00430489843711257, + 0.003026606049388647, + 0.0014419369399547577, + 0.0019098359625786543, + 0.0012481799349188805, + 0.015463129617273808, + 0.00969523936510086, + 0.02126433700323105, + 0.015424842946231365, + 0.0056240553967654705, + 0.0035578764509409666, + 0.0054183229804039, + 0.0070964256301522255, + 0.015191580168902874, + 0.010545420460402966, + 0.015021510422229767, + 0.007280163001269102, + 0.008929358795285225, + 0.003777276026085019, + 0.005633055232465267, + 0.003019314259290695, + 0.01827024482190609, + 0.016360171139240265, + 0.020576855167746544, + 0.020646117627620697, + 0.010043417103588581, + 0.005053053144365549, + 0.007516362238675356, + 0.005024679005146027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02634189836680889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.04305803403258324 + ], + "opposite_logprob_tokens": [ + -3.168057918548584 + ], + "logprob_reward": -0.04305803403258324, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305803403258324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17385484278202057, + 0.2573484778404236, + 0.20176680386066437, + 0.04840509220957756, + 0.031107740476727486, + 0.17121760547161102, + 0.025628285482525826, + 0.028791086748242378, + 0.02702394127845764, + 0.0348561629652977 + ], + "credit_map": [ + 0.02949507348239422, + 0.017916370183229446, + 0.06338761001825333, + 0.019593339413404465, + 0.016137562692165375, + 0.007657299283891916, + 0.011154486797749996, + 0.00851309671998024, + 0.03937733173370361, + 0.06506730616092682, + 0.05560887232422829, + 0.03138846904039383, + 0.016166452318429947, + 0.008480445481836796, + 0.01103798858821392, + 0.0302216038107872, + 0.06717289239168167, + 0.016941439360380173, + 0.021313995122909546, + 0.020007528364658356, + 0.007254165131598711, + 0.010509087704122066, + 0.011329973116517067, + 0.04723770171403885, + 0.006941627245396376, + 0.00666482700034976, + 0.00918357353657484, + 0.008019517175853252, + 0.006908631883561611, + 0.003956803120672703, + 0.003770391922444105, + 0.0029597205575555563, + 0.006161292549222708, + 0.003586504142731428, + 0.008186809718608856, + 0.0031946348026394844, + 0.00451540295034647, + 0.0015667823608964682, + 0.0022579296492040157, + 0.001638383255340159, + 0.06580537557601929, + 0.05650775879621506, + 0.012309851124882698, + 0.008512676693499088, + 0.010738678276538849, + 0.00333040370605886, + 0.005847450811415911, + 0.008165414445102215, + 0.003494767937809229, + 0.005560002289712429, + 0.00541054317727685, + 0.0038053097669035196, + 0.002373429946601391, + 0.0013869330286979675, + 0.002270822413265705, + 0.00132647599093616, + 0.0046058581210672855, + 0.002881312742829323, + 0.008797171525657177, + 0.0030500732827931643, + 0.003683974966406822, + 0.0015231372090056539, + 0.0023147696629166603, + 0.0019347872585058212, + 0.004124497063457966, + 0.0033130133524537086, + 0.006127640604972839, + 0.0032344742212444544, + 0.004606830421835184, + 0.001739951316267252, + 0.0021908567287027836, + 0.0016866765217855573, + 0.005842450074851513, + 0.005067233927547932, + 0.007617582567036152, + 0.005544805899262428, + 0.004038357175886631, + 0.0020203241147100925, + 0.0029849521815776825, + 0.00174045586027205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04305803403258324 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, lingering close-up captures the portrait of strategist Zhou Yu, rendered entirely in the traditional and iconic ink painting style. Thick, decisive calligraphic brushstrokes saturated with deep black ink define the sharp edges and overlapping plates of his ornate battle armor, while lighter gray tones suggest its metallic sheen. His sharp, angular features are twisted into a grimace of deep, anguished frustration; his forehead is furrowed, his lips pressed into a thin, bloodless line, and the muscles in his jaw are visibly clenched in suppressed anger. His background, a rough, textured backdrop resembling rice paper, is painted with faint, abstract ink splashes, evoking a sense of inner turmoil. The lighting is high-contrast and dramatic, with the pure white of the unpainted paper creating harsh highlights on one side of his face and the sharp lines of his nose, while the other side is plunged into deep ink shadows. The camera remains completely still, holding this long, lingering shot, forcing the viewer to confront the raw, silent pain etched on the face, creating an intense and powerful psychological atmosphere filled with a deep sense of frustration.", + "seed": 29287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_9375_seed_29287_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": -0.17819717405363916, + "vlm_logprob": -0.17819717405363916, + "vlm_yes_no_margin": 2.8499999761581423, + "vlm_token_credit": -0.17819717405363916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.5908410549163818 + ], + "opposite_logprob_tokens": [ + -0.8408410549163818 + ], + "logprob_reward": -0.5908410549163818, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5908410549163818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2679271996021271, + 0.08155298233032227, + 0.05597149580717087, + 0.06749893724918365, + 0.07724983990192413, + 0.03578192740678787, + 0.26250770688056946, + 0.04673314094543457, + 0.06151239573955536, + 0.043264396488666534 + ], + "credit_map": [ + 0.04677809402346611, + 0.02822057530283928, + 0.07130059599876404, + 0.0391254648566246, + 0.01706436462700367, + 0.017811277881264687, + 0.025137636810541153, + 0.022489206865429878, + 0.01571381278336048, + 0.007639298215508461, + 0.02060692198574543, + 0.011187693104147911, + 0.004075917880982161, + 0.0061653982847929, + 0.008929264731705189, + 0.007234668824821711, + 0.008972199633717537, + 0.004732878878712654, + 0.0070802150294184685, + 0.013328847475349903, + 0.003703173715621233, + 0.004061524290591478, + 0.00920192152261734, + 0.00489073758944869, + 0.01241039577871561, + 0.00767791923135519, + 0.0131644606590271, + 0.011405043303966522, + 0.0038053439930081367, + 0.005350572057068348, + 0.008170894347131252, + 0.005514311604201794, + 0.01126114185899496, + 0.008574874140322208, + 0.01577666774392128, + 0.01236370112746954, + 0.004845567978918552, + 0.007060796953737736, + 0.00991376768797636, + 0.007453312631696463, + 0.005726780276745558, + 0.004311019089072943, + 0.005854932591319084, + 0.005410095676779747, + 0.002801052061840892, + 0.0034174860920757055, + 0.005045171361416578, + 0.0032153890933841467, + 0.03587821498513222, + 0.03238510712981224, + 0.03575928136706352, + 0.055930621922016144, + 0.016604013741016388, + 0.02224792167544365, + 0.042731333523988724, + 0.020971199497580528, + 0.008312019519507885, + 0.005382648669183254, + 0.01132235024124384, + 0.006979946047067642, + 0.002723133424296975, + 0.002997656585648656, + 0.005548801738768816, + 0.003466581227257848, + 0.009980350732803345, + 0.006562578957527876, + 0.007110293488949537, + 0.012236223556101322, + 0.005528916139155626, + 0.005946698598563671, + 0.00936182588338852, + 0.004785510245710611, + 0.006781910080462694, + 0.003968213684856892, + 0.006391141563653946, + 0.010525698773562908, + 0.002413427457213402, + 0.0025769914500415325, + 0.007247450295835733, + 0.003359562950208783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.5908410549163818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.1603747010231018 + ], + "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": [ + 0.1909244805574417, + 0.0850411057472229, + 0.034847963601350784, + 0.1350821554660797, + 0.04163030534982681, + 0.29206007719039917, + 0.1336626261472702, + 0.031728073954582214, + 0.03531983122229576, + 0.019703395664691925 + ], + "credit_map": [ + 0.03143525496125221, + 0.013725774362683296, + 0.04255827143788338, + 0.034813378006219864, + 0.01324870903044939, + 0.015877412632107735, + 0.020297188311815262, + 0.018968483433127403, + 0.017642851918935776, + 0.0060017709620296955, + 0.023111209273338318, + 0.011543315835297108, + 0.004966833163052797, + 0.005915781483054161, + 0.008283778093755245, + 0.007575566414743662, + 0.005782141815871, + 0.004101411439478397, + 0.008003045804798603, + 0.005676025990396738, + 0.0028110367711633444, + 0.002215732354670763, + 0.003958456683903933, + 0.0023001127410680056, + 0.022802215069532394, + 0.015026920475065708, + 0.02305612340569496, + 0.025381941348314285, + 0.008477064780890942, + 0.009482878260314465, + 0.019207868725061417, + 0.011647138744592667, + 0.006971708964556456, + 0.004604138433933258, + 0.0056936671026051044, + 0.007485286332666874, + 0.003827923210337758, + 0.003467259230092168, + 0.006111626047641039, + 0.003468696027994156, + 0.054322388023138046, + 0.033670149743556976, + 0.0635916143655777, + 0.046572793275117874, + 0.018053701147437096, + 0.020529408007860184, + 0.030484335497021675, + 0.024835683405399323, + 0.019739726558327675, + 0.015533487312495708, + 0.013938413932919502, + 0.03501748666167259, + 0.007098447065800428, + 0.00959351472556591, + 0.02300122380256653, + 0.009740320034325123, + 0.005007438361644745, + 0.003065927419811487, + 0.003398105502128601, + 0.00868623424321413, + 0.0018536279676482081, + 0.0019492883002385497, + 0.005205625668168068, + 0.002561824396252632, + 0.006529655307531357, + 0.003512128023430705, + 0.005020358599722385, + 0.007287575397640467, + 0.0025736393872648478, + 0.002596758771687746, + 0.004797879606485367, + 0.0030018335673958063, + 0.0027265120297670364, + 0.0016937789041548967, + 0.0022402657195925713, + 0.005212195683270693, + 0.0014801295474171638, + 0.001291971537284553, + 0.0034018384758383036, + 0.0016567031852900982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "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 subject a portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.12703430652618408 + ], + "opposite_logprob_tokens": [ + -2.1270341873168945 + ], + "logprob_reward": -0.12703430652618408, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703430652618408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11691568046808243, + 0.13525548577308655, + 0.1950414925813675, + 0.07439103722572327, + 0.08240511268377304, + 0.07585536688566208, + 0.05054912716150284, + 0.021517295390367508, + 0.2147258222103119, + 0.0333436019718647 + ], + "credit_map": [ + 0.01840142533183098, + 0.010678243823349476, + 0.017630649730563164, + 0.024142425507307053, + 0.01013166829943657, + 0.009569120593369007, + 0.0188080333173275, + 0.00755411759018898, + 0.025076013058423996, + 0.012221496552228928, + 0.029879840090870857, + 0.016194850206375122, + 0.010253554210066795, + 0.015414887107908726, + 0.013613175600767136, + 0.012601670809090137, + 0.03746449947357178, + 0.019896656274795532, + 0.05399513989686966, + 0.029705019667744637, + 0.011260507628321648, + 0.011588148772716522, + 0.018918011337518692, + 0.012213517911732197, + 0.013477024622261524, + 0.008504752069711685, + 0.015250827185809612, + 0.011250512674450874, + 0.004742810502648354, + 0.005585242994129658, + 0.00894027017056942, + 0.006639604456722736, + 0.015501406975090504, + 0.007217313628643751, + 0.01908734068274498, + 0.017476994544267654, + 0.0045540775172412395, + 0.0042777108028531075, + 0.009249439463019371, + 0.0050408258102834225, + 0.014020268805325031, + 0.00936056300997734, + 0.014847165904939175, + 0.010963666252791882, + 0.005530520807951689, + 0.0060274009592831135, + 0.008008042350411415, + 0.007097743917256594, + 0.00879924837499857, + 0.005130869336426258, + 0.012279987335205078, + 0.007554970681667328, + 0.0026917632203549147, + 0.0036567954812198877, + 0.005545560736209154, + 0.004889932461082935, + 0.0031815427355468273, + 0.002142922952771187, + 0.002491955878213048, + 0.005524426233023405, + 0.0013349491637200117, + 0.001635134918615222, + 0.003141791094094515, + 0.002064572647213936, + 0.03642052412033081, + 0.017940502613782883, + 0.03666984289884567, + 0.059892524033784866, + 0.008684386499226093, + 0.007276749704033136, + 0.03676771745085716, + 0.011073581874370575, + 0.00559267308562994, + 0.002746770391240716, + 0.00789688341319561, + 0.006621523294597864, + 0.001853928784839809, + 0.0017677531577646732, + 0.005047715734690428, + 0.0018163564382120967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.12703430652618408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.008635676465928555 + ], + "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": [ + 0.2617184817790985, + 0.058605894446372986, + 0.11050810664892197, + 0.24385754764080048, + 0.03389354050159454, + 0.16792947053909302, + 0.033087048679590225, + 0.03447379171848297, + 0.042008284479379654, + 0.01391790620982647 + ], + "credit_map": [ + 0.042038436979055405, + 0.02023560181260109, + 0.06846845895051956, + 0.032411545515060425, + 0.014961480163037777, + 0.02620105631649494, + 0.02274811640381813, + 0.03465375676751137, + 0.010473774746060371, + 0.004047941416501999, + 0.013987448997795582, + 0.005510879214853048, + 0.0043157972395420074, + 0.006200277712196112, + 0.006585658062249422, + 0.007484117057174444, + 0.020338496193289757, + 0.011736501939594746, + 0.030884094536304474, + 0.01791251078248024, + 0.005883499048650265, + 0.006099270191043615, + 0.010987291112542152, + 0.006666433997452259, + 0.04339131340384483, + 0.02626846358180046, + 0.043446119874715805, + 0.04632191359996796, + 0.013555019162595272, + 0.017837801948189735, + 0.0335511676967144, + 0.019485726952552795, + 0.004309629090130329, + 0.0036357881035655737, + 0.003129312302917242, + 0.007820526137948036, + 0.0025651524774730206, + 0.0034042003098875284, + 0.005160709843039513, + 0.0038682203739881516, + 0.030290020629763603, + 0.01543460413813591, + 0.031614191830158234, + 0.034582436084747314, + 0.009468933567404747, + 0.011251498945057392, + 0.021103620529174805, + 0.014184148982167244, + 0.004905202891677618, + 0.003784659318625927, + 0.0037966936361044645, + 0.007996846921741962, + 0.0019159631337970495, + 0.0028090861160308123, + 0.005454444792121649, + 0.002424150239676237, + 0.005420479457825422, + 0.0033038740511983633, + 0.003921298775821924, + 0.009669994935393333, + 0.001948544173501432, + 0.0018731579184532166, + 0.005682099144905806, + 0.002654341049492359, + 0.0066197654232382774, + 0.00406612129881978, + 0.005928561091423035, + 0.011467894539237022, + 0.0020577681716531515, + 0.0019638047087937593, + 0.007283100858330727, + 0.0026212672237306833, + 0.002198135480284691, + 0.0012482807505875826, + 0.0015029815258458257, + 0.002723392564803362, + 0.0012806883314624429, + 0.0011584359453991055, + 0.002442630473524332, + 0.0013633607886731625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "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 subject a portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.004100131336599588 + ], + "opposite_logprob_tokens": [ + -5.504100322723389 + ], + "logprob_reward": -0.004100131336599588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100131336599588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1568523496389389, + 0.07465402781963348, + 0.057142097502946854, + 0.11956385523080826, + 0.08442999422550201, + 0.09526531398296356, + 0.15055660903453827, + 0.1794658601284027, + 0.03421643376350403, + 0.047853466123342514 + ], + "credit_map": [ + 0.026329748332500458, + 0.014092660509049892, + 0.041761741042137146, + 0.023920483887195587, + 0.00967456866055727, + 0.01112779788672924, + 0.015420773066580296, + 0.014524582773447037, + 0.013220511376857758, + 0.007983342744410038, + 0.015025049448013306, + 0.010413493029773235, + 0.0053051309660077095, + 0.0065454598516225815, + 0.008900073356926441, + 0.007260962389409542, + 0.010163775645196438, + 0.004978131968528032, + 0.011930372565984726, + 0.011654021218419075, + 0.004181662108749151, + 0.0041045197285711765, + 0.006145155522972345, + 0.003984457813203335, + 0.022644123062491417, + 0.010199316777288914, + 0.02755475789308548, + 0.01995382271707058, + 0.007430701050907373, + 0.007372463122010231, + 0.01378480065613985, + 0.010623881593346596, + 0.014572990126907825, + 0.009021805599331856, + 0.019844481721520424, + 0.014902923256158829, + 0.004723391495645046, + 0.006682642735540867, + 0.008363029919564724, + 0.006318729370832443, + 0.012788753025233746, + 0.010645946487784386, + 0.017425935715436935, + 0.013861688785254955, + 0.006514515727758408, + 0.011223040521144867, + 0.010575439780950546, + 0.012229997664690018, + 0.02654961682856083, + 0.015974244102835655, + 0.035704609006643295, + 0.023958595469594002, + 0.008855407126247883, + 0.011376767419278622, + 0.016147762537002563, + 0.011989599093794823, + 0.027986910194158554, + 0.016062837094068527, + 0.02985834702849388, + 0.04414062947034836, + 0.010021696798503399, + 0.01101380679756403, + 0.028743218630552292, + 0.01163841038942337, + 0.006059895269572735, + 0.0028822203166782856, + 0.008278514258563519, + 0.005662903655320406, + 0.002300016349181533, + 0.0022403192706406116, + 0.004049079492688179, + 0.0027434853836894035, + 0.008207350969314575, + 0.004656271543353796, + 0.010503464378416538, + 0.010043718852102757, + 0.0026860195212066174, + 0.002525473013520241, + 0.005940462928265333, + 0.0032907044515013695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.004100131336599588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, lingering close-up captures the portrait of strategist Zhou Yu, rendered entirely in the traditional and iconic ink painting style. Thick, decisive calligraphic brushstrokes saturated with deep black ink define the sharp edges and overlapping plates of his ornate battle armor, while lighter gray tones suggest its metallic sheen. His sharp, angular features are twisted into a grimace of deep, anguished frustration; his forehead is furrowed, his lips pressed into a thin, bloodless line, and the muscles in his jaw are visibly clenched in suppressed anger. His background, a rough, textured backdrop resembling rice paper, is painted with faint, abstract ink splashes, evoking a sense of inner turmoil. The lighting is high-contrast and dramatic, with the pure white of the unpainted paper creating harsh highlights on one side of his face and the sharp lines of his nose, while the other side is plunged into deep ink shadows. The camera remains completely still, holding this long, lingering shot, forcing the viewer to confront the raw, silent pain etched on the face, creating an intense and powerful psychological atmosphere filled with a deep sense of frustration.", + "seed": 29288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_11/index_9375_seed_29288_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": -0.21822387175634503, + "vlm_logprob": -0.21822387175634503, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.21822387175634503, + "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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any movement?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9126030206680298 + ], + "opposite_logprob_tokens": [ + -0.5376030206680298 + ], + "logprob_reward": -0.9126030206680298, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9126030206680298, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17571954429149628, + 0.08755321800708771, + 0.11543478816747665, + 0.08277789503335953, + 0.09456920623779297, + 0.06631487607955933, + 0.16343529522418976, + 0.038075536489486694, + 0.11955031752586365, + 0.05656932294368744 + ], + "credit_map": [ + 0.04321247339248657, + 0.018673710525035858, + 0.022406283766031265, + 0.021214675158262253, + 0.012367445975542068, + 0.015034092590212822, + 0.021175192669034004, + 0.021635662764310837, + 0.019246524199843407, + 0.007894718088209629, + 0.011593974195420742, + 0.017013199627399445, + 0.006044280715286732, + 0.007449976168572903, + 0.009495500475168228, + 0.008815047331154346, + 0.03169412165880203, + 0.011399065144360065, + 0.014484599232673645, + 0.011046817526221275, + 0.006609193980693817, + 0.012038237415254116, + 0.010129442438483238, + 0.018033307045698166, + 0.019343538209795952, + 0.007556644268333912, + 0.00903293862938881, + 0.00842658244073391, + 0.0063162692822515965, + 0.009358272887766361, + 0.008257023058831692, + 0.014486627653241158, + 0.03039836883544922, + 0.010346910916268826, + 0.013445031829178333, + 0.01146178413182497, + 0.005195487290620804, + 0.006647604517638683, + 0.007995824329555035, + 0.009078191593289375, + 0.01664489321410656, + 0.006761588621884584, + 0.00968131422996521, + 0.008295738138258457, + 0.0052069369703531265, + 0.005756513215601444, + 0.0048804315738379955, + 0.009087459184229374, + 0.030055727809667587, + 0.014712020754814148, + 0.017765307798981667, + 0.026903973892331123, + 0.011996332556009293, + 0.020245229825377464, + 0.014552721753716469, + 0.027203988283872604, + 0.008043128065764904, + 0.004834974650293589, + 0.004412480629980564, + 0.004339416511356831, + 0.003929351456463337, + 0.003434379119426012, + 0.004346018191426992, + 0.00473578879609704, + 0.008819559589028358, + 0.004306306131184101, + 0.005259898491203785, + 0.004494957160204649, + 0.005827657878398895, + 0.04715622588992119, + 0.005900135729461908, + 0.03778557479381561, + 0.007444783113896847, + 0.004858286585658789, + 0.004392136353999376, + 0.008524694480001926, + 0.005615726578980684, + 0.01126436423510313, + 0.005435313563793898, + 0.009034018032252789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.9126030206680298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.1129675880074501 + ], + "opposite_logprob_tokens": [ + -2.2379674911499023 + ], + "logprob_reward": -0.1129675880074501, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129675880074501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07421153038740158, + 0.2565748691558838, + 0.1436154544353485, + 0.06104807183146477, + 0.11323266476392746, + 0.055670205503702164, + 0.06094007194042206, + 0.05862002819776535, + 0.0744146853685379, + 0.1016724556684494 + ], + "credit_map": [ + 0.01000257022678852, + 0.009917255491018295, + 0.008626080118119717, + 0.011970327235758305, + 0.007188367657363415, + 0.0070223077200353146, + 0.010369046591222286, + 0.009115573018789291, + 0.05373712629079819, + 0.022052044048905373, + 0.03029455617070198, + 0.05427660793066025, + 0.01513112522661686, + 0.026217859238386154, + 0.021087095141410828, + 0.033778440207242966, + 0.03601900115609169, + 0.015407618135213852, + 0.02039247378706932, + 0.015778958797454834, + 0.007092960178852081, + 0.015020770020782948, + 0.0110743073746562, + 0.02282935567200184, + 0.01436921488493681, + 0.006439323537051678, + 0.0067182546481490135, + 0.007950745522975922, + 0.004374445881694555, + 0.006377586163580418, + 0.005464611109346151, + 0.009353887289762497, + 0.03196680545806885, + 0.010678577236831188, + 0.014547097496688366, + 0.016565660014748573, + 0.006880364380776882, + 0.0101839080452919, + 0.0088871531188488, + 0.013523082248866558, + 0.015240272507071495, + 0.005870138294994831, + 0.007358972914516926, + 0.007086163852363825, + 0.004642547108232975, + 0.004732178058475256, + 0.005407496355473995, + 0.005332434084266424, + 0.012266743928194046, + 0.006603891495615244, + 0.0067511871457099915, + 0.009184212423861027, + 0.0069126104936003685, + 0.006030069198459387, + 0.006208323873579502, + 0.006983030121773481, + 0.012585686519742012, + 0.0055841137655079365, + 0.006243590265512466, + 0.005245986860245466, + 0.004199751187115908, + 0.008957971818745136, + 0.004997919779270887, + 0.010805007070302963, + 0.005073156673461199, + 0.002962101949378848, + 0.0030555189587175846, + 0.003357117995619774, + 0.0030999653972685337, + 0.029612809419631958, + 0.0037470990791916847, + 0.02350691333413124, + 0.013971037231385708, + 0.011021442711353302, + 0.008751630783081055, + 0.02288767695426941, + 0.009136677719652653, + 0.012684248387813568, + 0.01018256414681673, + 0.013037178665399551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.1129675880074501 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.005956872832030058 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956872832030058, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956872832030058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2551252841949463, + 0.07412263005971909, + 0.19486956298351288, + 0.04717843607068062, + 0.04825164005160332, + 0.0947413444519043, + 0.03461635112762451, + 0.10373841971158981, + 0.055283065885305405, + 0.09207324683666229 + ], + "credit_map": [ + 0.061222560703754425, + 0.026477791368961334, + 0.03316677361726761, + 0.029620695859193802, + 0.023032115772366524, + 0.018749399110674858, + 0.03259620815515518, + 0.030259760096669197, + 0.008110364899039268, + 0.007225313223898411, + 0.006949362810701132, + 0.014775886200368404, + 0.010760940611362457, + 0.005721492227166891, + 0.0138841038569808, + 0.006695171818137169, + 0.04763203486800194, + 0.01921859383583069, + 0.024030204862356186, + 0.020432177931070328, + 0.015519998036324978, + 0.019249895587563515, + 0.020491691306233406, + 0.028294960036873817, + 0.006492651533335447, + 0.004945935215801001, + 0.004121946170926094, + 0.006771325133740902, + 0.007727579213678837, + 0.004365840926766396, + 0.006865035742521286, + 0.005888121668249369, + 0.010717334225773811, + 0.00668254354968667, + 0.004531920421868563, + 0.005476351827383041, + 0.005980193614959717, + 0.0041895401664078236, + 0.005395702086389065, + 0.005278049502521753, + 0.024268541485071182, + 0.010144529864192009, + 0.011786566115915775, + 0.014462246559560299, + 0.00768248550593853, + 0.007196098566055298, + 0.009515290148556232, + 0.009685582481324673, + 0.009404141455888748, + 0.003505992004647851, + 0.004479102790355682, + 0.0036725772079080343, + 0.0024245386011898518, + 0.0033218993339687586, + 0.003118041902780533, + 0.004690059460699558, + 0.02503392845392227, + 0.012522739358246326, + 0.013837536796927452, + 0.012091190554201603, + 0.010580959729850292, + 0.007431994657963514, + 0.012308113276958466, + 0.009931953623890877, + 0.007801768369972706, + 0.0036427604500204325, + 0.003273374866694212, + 0.002995927119627595, + 0.003938155248761177, + 0.016128765419125557, + 0.003647098783403635, + 0.013855214230716228, + 0.009621585719287395, + 0.005015517584979534, + 0.006666928995400667, + 0.00727872783318162, + 0.007103589829057455, + 0.028783641755580902, + 0.0064514558762311935, + 0.021151795983314514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.005956872832030058 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.054946962743997574 + ], + "opposite_logprob_tokens": [ + -2.9299468994140625 + ], + "logprob_reward": -0.054946962743997574, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054946962743997574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11627758294343948, + 0.20928354561328888, + 0.21744054555892944, + 0.09726355969905853, + 0.17122158408164978, + 0.017339274287223816, + 0.04189566150307655, + 0.05928434059023857, + 0.030101215466856956, + 0.03989267349243164 + ], + "credit_map": [ + 0.03218856081366539, + 0.012486516498029232, + 0.01465697679668665, + 0.01964707113802433, + 0.0068205855786800385, + 0.008711965754628181, + 0.009855851531028748, + 0.011910060420632362, + 0.052277643233537674, + 0.02215844765305519, + 0.026399558410048485, + 0.031235603615641594, + 0.010249987244606018, + 0.019613949581980705, + 0.01620117574930191, + 0.031147176399827003, + 0.05780293419957161, + 0.021287553012371063, + 0.02644105814397335, + 0.02181953936815262, + 0.012900315225124359, + 0.022977305576205254, + 0.0198531374335289, + 0.034358710050582886, + 0.026427265256643295, + 0.00869414396584034, + 0.012669292278587818, + 0.013589792884886265, + 0.005521134473383427, + 0.009276775643229485, + 0.0070516509003937244, + 0.014033501967787743, + 0.036155395209789276, + 0.013928302563726902, + 0.019605111330747604, + 0.039537250995635986, + 0.010594156570732594, + 0.016576912254095078, + 0.012390743009746075, + 0.02243371680378914, + 0.0026162483263760805, + 0.0023784409277141094, + 0.0019051812123507261, + 0.0026383469812572002, + 0.0022851023823022842, + 0.0016235408838838339, + 0.0018726219423115253, + 0.0020197939593344927, + 0.010679321363568306, + 0.003429022151976824, + 0.0059023224748671055, + 0.00474896002560854, + 0.00331109412945807, + 0.004313088487833738, + 0.003643117379397154, + 0.0058687347918748856, + 0.011233882047235966, + 0.006825149059295654, + 0.0062162140384316444, + 0.009020869620144367, + 0.005947531200945377, + 0.006721792742609978, + 0.006267627701163292, + 0.007051273714751005, + 0.004615663550794125, + 0.0028010800015181303, + 0.002888474380597472, + 0.0033309112768620253, + 0.0032696572598069906, + 0.005625465884804726, + 0.0028650888707488775, + 0.004704875405877829, + 0.0038363798521459103, + 0.0028234466444700956, + 0.003319170093163848, + 0.005728020332753658, + 0.0023032943718135357, + 0.011185414157807827, + 0.002966902684420347, + 0.007730043958872557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.054946962743997574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 portrait of a man identified as strategist Zhou Yu?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style rendered entirely in a traditional ink painting style with thick calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character display a grimace of deep frustration with furrowed brows, pressed lips, and clenched jaw muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a rough textured surface resembling rice paper with faint abstract ink splashes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera completely still holding a static close-up shot without any 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": [ + -0.0046449145302176476 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.0046449145302176476, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046449145302176476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09732876718044281, + 0.09806738048791885, + 0.07843859493732452, + 0.09201577305793762, + 0.08788701891899109, + 0.032929036766290665, + 0.034184686839580536, + 0.051278091967105865, + 0.05061790347099304, + 0.3772526979446411 + ], + "credit_map": [ + 0.02280273102223873, + 0.013643135316669941, + 0.01169187854975462, + 0.013629897497594357, + 0.007911375723779202, + 0.007554321084171534, + 0.009158995002508163, + 0.010936444625258446, + 0.01715659536421299, + 0.009042962454259396, + 0.011279894039034843, + 0.021677499637007713, + 0.005650507286190987, + 0.009606516920030117, + 0.011047585867345333, + 0.01260583195835352, + 0.018580032512545586, + 0.008004640229046345, + 0.009130843915045261, + 0.01475716382265091, + 0.004167858045548201, + 0.006880955770611763, + 0.006416983436793089, + 0.010500127449631691, + 0.027588700875639915, + 0.010364587418735027, + 0.012589149177074432, + 0.01003539003431797, + 0.0053130448795855045, + 0.007336597424000502, + 0.0077866907231509686, + 0.011001626029610634, + 0.02101767249405384, + 0.007665273733437061, + 0.011739113368093967, + 0.015389487147331238, + 0.00438171299174428, + 0.008144786581397057, + 0.0063295564614236355, + 0.013219427317380905, + 0.007155192084610462, + 0.004079409874975681, + 0.003316916059702635, + 0.006344247609376907, + 0.0027384173590689898, + 0.002712182467803359, + 0.0034798497799783945, + 0.0031028236262500286, + 0.006437778938561678, + 0.0031246067956089973, + 0.002993799978867173, + 0.005810695234686136, + 0.0030720457434654236, + 0.004094668198376894, + 0.0034566817339509726, + 0.005194410681724548, + 0.012451673857867718, + 0.005664308089762926, + 0.00652908394113183, + 0.006591674406081438, + 0.004323064349591732, + 0.00475195050239563, + 0.004785694647580385, + 0.00618064496666193, + 0.006370916962623596, + 0.005231705028563738, + 0.004614123608916998, + 0.005190908443182707, + 0.007864916697144508, + 0.006853358820080757, + 0.006517123430967331, + 0.007974857464432716, + 0.03525875508785248, + 0.023178474977612495, + 0.027089854702353477, + 0.025000229477882385, + 0.03684514760971069, + 0.11415157467126846, + 0.029747722670435905, + 0.08598095923662186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0046449145302176476 + } + ], + "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/0000022/group_12/index_tensor([3638, 3638])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_tensor([3638, 3638])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3726984636968c538a18a3815d221a1253760347 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_tensor([3638, 3638])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a static, low-angle wide shot, a viscous, unstoppable torrent of incandescent magma slowly but surely intrudes from the right side of the frame, engulfing the dry, cracked, black earth in the foreground. The core of this magma glows a fiery white, while its surface is a dazzling orange-red fluid, covered by a thin, dark red crust that constantly tears and heals as it flows. As the magma front encounters the cold air, its shiny surface rapidly wrinkles and darkens. With a muffled crackling sound, it rapidly cools and solidifies into a vast expanse of rugged, porous black volcanic rock. Despite solidification, a dark red residual glow still pulses within the deep cracks of the newly formed rock, creating a dramatic contrast with the intense glow of the liquid magma advancing behind it. The scene is illuminated solely by the magma itself, casting dancing shadows amidst the steaming heat and thin sulfurous fumes, creating an apocalyptic atmosphere imbued with primal power, both majestic and awe-inspiring.", + "seed": 21357, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_3638_seed_21357_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": -0.05261740293353796, + "vlm_logprob": -0.05261740293353796, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.05261740293353796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19119183719158173 + ], + "opposite_logprob_tokens": [ + -1.9411917924880981 + ], + "logprob_reward": -0.19119183719158173, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.19119183719158173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15190351009368896, + 0.1562272310256958, + 0.06592928618192673, + 0.1414562314748764, + 0.05514337122440338, + 0.15275292098522186, + 0.11751531809568405, + 0.03017420880496502, + 0.07124818116426468, + 0.05764970928430557 + ], + "credit_map": [ + 0.011362701654434204, + 0.01215145643800497, + 0.036923810839653015, + 0.023050518706440926, + 0.01513191033154726, + 0.01160438172519207, + 0.017388291656970978, + 0.024290449917316437, + 0.012448349967598915, + 0.023867139592766762, + 0.025704622268676758, + 0.03436862304806709, + 0.011386026628315449, + 0.012647399678826332, + 0.02323024533689022, + 0.01257482822984457, + 0.009462271817028522, + 0.00841642264276743, + 0.00912797823548317, + 0.008921177126467228, + 0.008230136707425117, + 0.0044655767269432545, + 0.00869005173444748, + 0.008615676313638687, + 0.01341423112899065, + 0.014313940890133381, + 0.023006804287433624, + 0.015455720014870167, + 0.012761149555444717, + 0.014053497463464737, + 0.030496543273329735, + 0.017954347655177116, + 0.006761264055967331, + 0.006929159630089998, + 0.008181880228221416, + 0.00591264758259058, + 0.007429958321154118, + 0.007886809296905994, + 0.006008094176650047, + 0.006033558864146471, + 0.020785747095942497, + 0.016296060755848885, + 0.02644309401512146, + 0.018275324255228043, + 0.013875599950551987, + 0.0142906429246068, + 0.023259487003087997, + 0.019526977092027664, + 0.014217437244951725, + 0.011852757073938847, + 0.017516445368528366, + 0.02413291670382023, + 0.010801561176776886, + 0.010288690216839314, + 0.014886599965393543, + 0.013818913139402866, + 0.003923907410353422, + 0.003918548114597797, + 0.002905055182054639, + 0.0040639606304466724, + 0.005514027550816536, + 0.004465149715542793, + 0.0029959685634821653, + 0.0023875923361629248, + 0.005662228912115097, + 0.008770436979830265, + 0.010807265527546406, + 0.016779646277427673, + 0.006155872717499733, + 0.008232519030570984, + 0.0075605520978569984, + 0.007279657293111086, + 0.006473444402217865, + 0.0055476827546954155, + 0.006436820141971111, + 0.0075585064478218555, + 0.00596512109041214, + 0.010083320550620556, + 0.00999316107481718, + 0.005591655150055885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.19119183719158173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598072558641434 + ], + "opposite_logprob_tokens": [ + -3.8955981731414795 + ], + "logprob_reward": -0.020598072558641434, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598072558641434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15666286647319794, + 0.12060278654098511, + 0.06952271610498428, + 0.2035660594701767, + 0.061722010374069214, + 0.16104020178318024, + 0.03353198990225792, + 0.04994312301278114, + 0.051723625510931015, + 0.09168457984924316 + ], + "credit_map": [ + 0.01892266236245632, + 0.015501420944929123, + 0.03398522362112999, + 0.020526748150587082, + 0.021473834291100502, + 0.010834789834916592, + 0.012134925462305546, + 0.02328326925635338, + 0.014434757642447948, + 0.0173659548163414, + 0.014678409323096275, + 0.019018474966287613, + 0.01184711791574955, + 0.007460532244294882, + 0.02248620241880417, + 0.013311340473592281, + 0.008375277742743492, + 0.009847685694694519, + 0.010258005931973457, + 0.008002598769962788, + 0.009910210967063904, + 0.0077471318654716015, + 0.008938184939324856, + 0.006443622522056103, + 0.02364633046090603, + 0.01847855933010578, + 0.038914475589990616, + 0.020780431106686592, + 0.01691102422773838, + 0.026898253709077835, + 0.025681406259536743, + 0.032255593687295914, + 0.005625831428915262, + 0.005888157989829779, + 0.0062197172082960606, + 0.0099156079813838, + 0.007096063811331987, + 0.005947585217654705, + 0.012195038609206676, + 0.00883401557803154, + 0.01561292726546526, + 0.015652503818273544, + 0.027903379872441292, + 0.021144699305295944, + 0.0159746203571558, + 0.014637402258813381, + 0.028811294585466385, + 0.021303389221429825, + 0.004246638622134924, + 0.003620593110099435, + 0.003672190010547638, + 0.005072343163192272, + 0.0035850098356604576, + 0.0038184963632375, + 0.006010985467582941, + 0.003505736356601119, + 0.006385226268321276, + 0.0056574707850813866, + 0.004124610219150782, + 0.004863150883466005, + 0.00893037486821413, + 0.006372259464114904, + 0.008030027151107788, + 0.005580007564276457, + 0.004595451522618532, + 0.008227568119764328, + 0.006230785045772791, + 0.011780339293181896, + 0.005389186087995768, + 0.005542389117181301, + 0.004896878264844418, + 0.005061028525233269, + 0.008344241417944431, + 0.019772561267018318, + 0.010346587747335434, + 0.013154134154319763, + 0.00867283996194601, + 0.00611437251791358, + 0.011681370437145233, + 0.013598473742604256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.020598072558641434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23559288680553436, + 0.16452570259571075, + 0.05200392007827759, + 0.10822920501232147, + 0.07913956791162491, + 0.10077966749668121, + 0.06800840049982071, + 0.0441117025911808, + 0.05207325890660286, + 0.0955357551574707 + ], + "credit_map": [ + 0.018123608082532883, + 0.020428331568837166, + 0.039159417152404785, + 0.04374876618385315, + 0.025013066828250885, + 0.013419999741017818, + 0.038928791880607605, + 0.036770883947610855, + 0.017812103033065796, + 0.020646482706069946, + 0.021288570016622543, + 0.020430605858564377, + 0.02110634371638298, + 0.01980862021446228, + 0.018111927434802055, + 0.025321034714579582, + 0.004904131405055523, + 0.007290643639862537, + 0.008885170333087444, + 0.006195590365678072, + 0.005476272664964199, + 0.005670812446624041, + 0.005222805310040712, + 0.00835848692804575, + 0.012358059175312519, + 0.011555359698832035, + 0.019631771370768547, + 0.0137596745043993, + 0.010087920352816582, + 0.011685996316373348, + 0.015159936621785164, + 0.013990482315421104, + 0.011392898857593536, + 0.010136410593986511, + 0.012313133105635643, + 0.011237775906920433, + 0.008521262556314468, + 0.00953700765967369, + 0.007013421040028334, + 0.00898765493184328, + 0.007650922983884811, + 0.00850889552384615, + 0.014198699966073036, + 0.014410283416509628, + 0.009527635760605335, + 0.009078644216060638, + 0.024150049313902855, + 0.013254532590508461, + 0.008379782550036907, + 0.007650061976164579, + 0.008232701569795609, + 0.009069661609828472, + 0.007157622836530209, + 0.008024184964597225, + 0.009962446056306362, + 0.009531940333545208, + 0.003756935242563486, + 0.005174006801098585, + 0.004612377379089594, + 0.009655137546360493, + 0.005753750912845135, + 0.005496446508914232, + 0.006100579630583525, + 0.003562465775758028, + 0.004792103078216314, + 0.007832793518900871, + 0.0054305922240018845, + 0.011406389065086842, + 0.005697937682271004, + 0.005620397627353668, + 0.00601440342143178, + 0.0052786413580179214, + 0.009080098941922188, + 0.018494484946131706, + 0.014144358225166798, + 0.012013298459351063, + 0.009402441792190075, + 0.010028552263975143, + 0.009894954040646553, + 0.012477567419409752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04985805228352547, + 0.05292940512299538, + 0.0524764247238636, + 0.12785868346691132, + 0.15632450580596924, + 0.223504900932312, + 0.06857903301715851, + 0.08985505253076553, + 0.06707783043384552, + 0.11153615266084671 + ], + "credit_map": [ + 0.008695676922798157, + 0.004703408572822809, + 0.007275944575667381, + 0.006223843898624182, + 0.006090983748435974, + 0.004499085713177919, + 0.005439921282231808, + 0.006929187104105949, + 0.004379585850983858, + 0.006967877969145775, + 0.008762428537011147, + 0.008532339707016945, + 0.004244547802954912, + 0.003899595933035016, + 0.010040477849543095, + 0.006102553103119135, + 0.006077558733522892, + 0.007944517768919468, + 0.009931412525475025, + 0.005681334063410759, + 0.006229122634977102, + 0.0038087107241153717, + 0.006137822289019823, + 0.006665944587439299, + 0.009901970624923706, + 0.011575022712349892, + 0.024982983246445656, + 0.016330519691109657, + 0.010014107450842857, + 0.011217005550861359, + 0.026341747492551804, + 0.01749531924724579, + 0.014263325370848179, + 0.014915269799530506, + 0.018297307193279266, + 0.033769164234399796, + 0.010888857766985893, + 0.01495625264942646, + 0.02927151322364807, + 0.019962800666689873, + 0.02180221676826477, + 0.01869203895330429, + 0.037544652819633484, + 0.027276724576950073, + 0.018353024497628212, + 0.020682429894804955, + 0.048441264778375626, + 0.030712544918060303, + 0.008277474902570248, + 0.0076740337535738945, + 0.00870698131620884, + 0.0075964354909956455, + 0.007456162478774786, + 0.009187208488583565, + 0.008729390799999237, + 0.010951343923807144, + 0.006324330810457468, + 0.008508702740073204, + 0.009901721961796284, + 0.015491204336285591, + 0.01155648473650217, + 0.009646028280258179, + 0.018023166805505753, + 0.010403409600257874, + 0.0060006435960531235, + 0.008539675734937191, + 0.007322126533836126, + 0.014377984218299389, + 0.007827782072126865, + 0.009374219924211502, + 0.007148423232138157, + 0.0064869713969528675, + 0.009505298919975758, + 0.020341770723462105, + 0.014471099711954594, + 0.015029270201921463, + 0.011076805181801319, + 0.013134554959833622, + 0.015641124919056892, + 0.012336233630776405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.018187187612056732 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07933484762907028, + 0.07808175683021545, + 0.06851454079151154, + 0.09302658587694168, + 0.1494765728712082, + 0.06676428765058517, + 0.04090574011206627, + 0.08031139522790909, + 0.0843757912516594, + 0.2592085301876068 + ], + "credit_map": [ + 0.008449617773294449, + 0.006846671458333731, + 0.019072411581873894, + 0.012115011923015118, + 0.008465131744742393, + 0.005048495717346668, + 0.006793285720050335, + 0.012544221244752407, + 0.005963809322565794, + 0.012847529724240303, + 0.009715475142002106, + 0.016938749700784683, + 0.0067992485128343105, + 0.004740196745842695, + 0.012850398197770119, + 0.008226347155869007, + 0.0075412942096591, + 0.008600138127803802, + 0.010456699877977371, + 0.015290486626327038, + 0.006421670783311129, + 0.0057601300068199635, + 0.008263888768851757, + 0.00618022633716464, + 0.007073124870657921, + 0.008016694337129593, + 0.020865418016910553, + 0.011037172749638557, + 0.006284071132540703, + 0.010761883109807968, + 0.015863848850131035, + 0.013124367222189903, + 0.015052558854222298, + 0.014956554397940636, + 0.021491872146725655, + 0.02813723310828209, + 0.013428355567157269, + 0.01405332051217556, + 0.026095915585756302, + 0.016260752454400063, + 0.006003553979098797, + 0.005942192859947681, + 0.011079197749495506, + 0.008305865339934826, + 0.006583771202713251, + 0.00585632398724556, + 0.014317753724753857, + 0.008675627410411835, + 0.00464651919901371, + 0.004344226326793432, + 0.005010681692510843, + 0.00729218078777194, + 0.0036855298094451427, + 0.00490097189322114, + 0.006243078038096428, + 0.004782549571245909, + 0.007606137078255415, + 0.01019973587244749, + 0.008737688884139061, + 0.014034519903361797, + 0.00983535498380661, + 0.009706406854093075, + 0.012368948198854923, + 0.007822594605386257, + 0.008010969497263432, + 0.011121608316898346, + 0.007283215411007404, + 0.01631823740899563, + 0.010237419977784157, + 0.01492177415639162, + 0.008084913715720177, + 0.00839765090495348, + 0.026174679398536682, + 0.07180873304605484, + 0.030299881473183632, + 0.0320548340678215, + 0.024929117411375046, + 0.015134554356336594, + 0.01713685691356659, + 0.04166985675692558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.026366978883743286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, low-angle wide shot, a viscous, unstoppable torrent of incandescent magma slowly but surely intrudes from the right side of the frame, engulfing the dry, cracked, black earth in the foreground. The core of this magma glows a fiery white, while its surface is a dazzling orange-red fluid, covered by a thin, dark red crust that constantly tears and heals as it flows. As the magma front encounters the cold air, its shiny surface rapidly wrinkles and darkens. With a muffled crackling sound, it rapidly cools and solidifies into a vast expanse of rugged, porous black volcanic rock. Despite solidification, a dark red residual glow still pulses within the deep cracks of the newly formed rock, creating a dramatic contrast with the intense glow of the liquid magma advancing behind it. The scene is illuminated solely by the magma itself, casting dancing shadows amidst the steaming heat and thin sulfurous fumes, creating an apocalyptic atmosphere imbued with primal power, both majestic and awe-inspiring.", + "seed": 21358, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_3638_seed_21358_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": -0.04357395367696881, + "vlm_logprob": -0.04357395367696881, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04357395367696881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11160483956336975 + ], + "opposite_logprob_tokens": [ + -2.736604928970337 + ], + "logprob_reward": -0.11160483956336975, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.11160483956336975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20917527377605438, + 0.07697729021310806, + 0.0340685099363327, + 0.125935360789299, + 0.13786223530769348, + 0.08427892625331879, + 0.14542965590953827, + 0.03494451195001602, + 0.1169319599866867, + 0.03439625725150108 + ], + "credit_map": [ + 0.023841798305511475, + 0.03409693017601967, + 0.03700408712029457, + 0.02198377437889576, + 0.02155136875808239, + 0.02344321459531784, + 0.027172163128852844, + 0.020081927999854088, + 0.011476002633571625, + 0.00865644495934248, + 0.007975451648235321, + 0.008935883641242981, + 0.010553814470767975, + 0.015310646966099739, + 0.007208844181150198, + 0.006860206834971905, + 0.004375114105641842, + 0.006148114334791899, + 0.003749289084225893, + 0.00382285937666893, + 0.00459402147680521, + 0.004078433848917484, + 0.003921576775610447, + 0.0033791009336709976, + 0.014301224611699581, + 0.0174721609801054, + 0.015695422887802124, + 0.016068870201706886, + 0.011603014543652534, + 0.016873648390173912, + 0.014547346159815788, + 0.019373681396245956, + 0.012291181832551956, + 0.03258858993649483, + 0.009839392267167568, + 0.011698046699166298, + 0.03383234888315201, + 0.015378120355308056, + 0.008936851285398006, + 0.013297717086970806, + 0.011659200303256512, + 0.011721530929207802, + 0.007426932454109192, + 0.0068433829583227634, + 0.01573427952826023, + 0.015569007955491543, + 0.006323853507637978, + 0.009000740945339203, + 0.018836023285984993, + 0.018293626606464386, + 0.017073996365070343, + 0.022670038044452667, + 0.02022928185760975, + 0.017174188047647476, + 0.017776843160390854, + 0.013375667855143547, + 0.006296580657362938, + 0.009212477132678032, + 0.0030356619972735643, + 0.002174707129597664, + 0.004381230100989342, + 0.003439377760514617, + 0.0030310151632875204, + 0.0033734608441591263, + 0.011404664255678654, + 0.01280076615512371, + 0.01604747772216797, + 0.014232604764401913, + 0.014148483984172344, + 0.016886334866285324, + 0.017363300547003746, + 0.014048327691853046, + 0.0034341549035161734, + 0.00426169577986002, + 0.005646157544106245, + 0.00452349754050374, + 0.004552627447992563, + 0.003359912196174264, + 0.004966748878359795, + 0.0036514613311737776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.11160483956336975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054973140358924866 + ], + "opposite_logprob_tokens": [ + -2.9299731254577637 + ], + "logprob_reward": -0.054973140358924866, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054973140358924866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16308708488941193, + 0.06506465375423431, + 0.16773992776870728, + 0.13842174410820007, + 0.08217089623212814, + 0.08024264127016068, + 0.04321014881134033, + 0.0402136966586113, + 0.16531769931316376, + 0.05453145131468773 + ], + "credit_map": [ + 0.012669109739363194, + 0.05088537931442261, + 0.011288226582109928, + 0.007645328529179096, + 0.04688877612352371, + 0.012146318331360817, + 0.011479531414806843, + 0.010084409266710281, + 0.00775787141174078, + 0.007117791101336479, + 0.007337940391153097, + 0.010551665909588337, + 0.0077591342851519585, + 0.00774119608104229, + 0.008837925270199776, + 0.007961121387779713, + 0.021347399801015854, + 0.03257809206843376, + 0.013149327598512173, + 0.011305282823741436, + 0.04274362698197365, + 0.020930549129843712, + 0.009839572012424469, + 0.01584607921540737, + 0.01562824845314026, + 0.02000250481069088, + 0.012179672718048096, + 0.01319094654172659, + 0.013543007895350456, + 0.026265213266015053, + 0.01764107309281826, + 0.019971074536442757, + 0.006782091688364744, + 0.021244565024971962, + 0.005138295236974955, + 0.007116605993360281, + 0.020795712247490883, + 0.009576876647770405, + 0.004439519252628088, + 0.0070772250182926655, + 0.00991018582135439, + 0.010428640991449356, + 0.007421907503157854, + 0.007193037308752537, + 0.01358422264456749, + 0.01544209849089384, + 0.007055374328047037, + 0.00920717604458332, + 0.005968680139631033, + 0.004098939243704081, + 0.0041944184340536594, + 0.004640903323888779, + 0.005795061122626066, + 0.007796584162861109, + 0.006040956359356642, + 0.004674605559557676, + 0.005787283647805452, + 0.00915012788027525, + 0.003461663844063878, + 0.0049742660485208035, + 0.004257918335497379, + 0.005809946451336145, + 0.003958523739129305, + 0.0028139683417975903, + 0.017035365104675293, + 0.018810303881764412, + 0.018873047083616257, + 0.018259266391396523, + 0.02206544764339924, + 0.02842298336327076, + 0.021300703287124634, + 0.020550580695271492, + 0.01004793867468834, + 0.005201208870857954, + 0.006195524241775274, + 0.004507746547460556, + 0.004804964642971754, + 0.013947699218988419, + 0.004795702174305916, + 0.00503066461533308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.054973140358924866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10319815576076508, + 0.09491849690675735, + 0.10642390698194504, + 0.13327831029891968, + 0.13598588109016418, + 0.11509054899215698, + 0.07906533032655716, + 0.057430703192949295, + 0.0870000496506691, + 0.08760859817266464 + ], + "credit_map": [ + 0.014101188629865646, + 0.012312229722738266, + 0.00948342401534319, + 0.012975496239960194, + 0.01308873575180769, + 0.015425138175487518, + 0.01622004620730877, + 0.009591897018253803, + 0.013857003301382065, + 0.009123281575739384, + 0.008814938366413116, + 0.011895177885890007, + 0.012453917413949966, + 0.018315725028514862, + 0.010622291825711727, + 0.009836162440478802, + 0.014641092158854008, + 0.021919095888733864, + 0.009364352561533451, + 0.008126964792609215, + 0.021862603724002838, + 0.01591959223151207, + 0.0063041518442332745, + 0.008286049589514732, + 0.015428824350237846, + 0.01458701305091381, + 0.01583849824965, + 0.014301990158855915, + 0.011832230724394321, + 0.021001366898417473, + 0.017307618632912636, + 0.022980768233537674, + 0.0088960574939847, + 0.04000648111104965, + 0.006836652290076017, + 0.006840369198471308, + 0.04771092161536217, + 0.009864299558103085, + 0.004818015731871128, + 0.011013098992407322, + 0.013000297360122204, + 0.014060442335903645, + 0.01116750854998827, + 0.013589857146143913, + 0.01637403666973114, + 0.021963639184832573, + 0.010781161487102509, + 0.014153610914945602, + 0.01177249662578106, + 0.007748851086944342, + 0.006255723536014557, + 0.010537763126194477, + 0.01250450685620308, + 0.010922419838607311, + 0.012445695698261261, + 0.006877867970615625, + 0.00788621511310339, + 0.008959920145571232, + 0.006835445761680603, + 0.005531465169042349, + 0.005811805371195078, + 0.010552550666034222, + 0.005164384841918945, + 0.006688917521387339, + 0.01026072446256876, + 0.011354316957294941, + 0.010523466393351555, + 0.01132111344486475, + 0.011149069294333458, + 0.010482591576874256, + 0.011504181660711765, + 0.010404586791992188, + 0.009223625995218754, + 0.010280606336891651, + 0.011072508059442043, + 0.011877215467393398, + 0.011532044969499111, + 0.010009191930294037, + 0.012026836164295673, + 0.0115865683183074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.0067449514754116535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347124949097633 + ], + "opposite_logprob_tokens": [ + -3.6513471603393555 + ], + "logprob_reward": -0.026347124949097633, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347124949097633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17906838655471802, + 0.12127655744552612, + 0.08803244680166245, + 0.1318484991788864, + 0.10122658312320709, + 0.05893431231379509, + 0.04096754267811775, + 0.10418369621038437, + 0.114865742623806, + 0.059596188366413116 + ], + "credit_map": [ + 0.01632664166390896, + 0.027913128957152367, + 0.02433587796986103, + 0.021564383059740067, + 0.021238544955849648, + 0.013798483647406101, + 0.03383249044418335, + 0.020058831200003624, + 0.00959971733391285, + 0.009597381576895714, + 0.015941085293889046, + 0.03217300400137901, + 0.008387571200728416, + 0.008940739557147026, + 0.02014787122607231, + 0.016489187255501747, + 0.010105205699801445, + 0.01325822900980711, + 0.0073912665247917175, + 0.008092697709798813, + 0.022163843736052513, + 0.009906955994665623, + 0.006955897435545921, + 0.010158354416489601, + 0.01517039630562067, + 0.01810792088508606, + 0.013724702410399914, + 0.01402067206799984, + 0.011935621500015259, + 0.02334517054259777, + 0.014248078688979149, + 0.021295925602316856, + 0.006918579805642366, + 0.015780430287122726, + 0.0066666072234511375, + 0.008017665706574917, + 0.025690825656056404, + 0.016465673223137856, + 0.007050259038805962, + 0.01463654451072216, + 0.008280948735773563, + 0.007518296130001545, + 0.005315060261636972, + 0.005260859150439501, + 0.010190610773861408, + 0.011166197247803211, + 0.005857951007783413, + 0.0053443945944309235, + 0.007172343786805868, + 0.004912929609417915, + 0.00431540422141552, + 0.003597304690629244, + 0.006266487296670675, + 0.006312686018645763, + 0.005038291681557894, + 0.003352097002789378, + 0.01928888075053692, + 0.02188131771981716, + 0.01095656119287014, + 0.017377853393554688, + 0.009728457778692245, + 0.007957310415804386, + 0.009323183447122574, + 0.0076701282523572445, + 0.013895442709326744, + 0.01525532640516758, + 0.011788870207965374, + 0.01154246460646391, + 0.013805314898490906, + 0.015007763169705868, + 0.014881502836942673, + 0.018689053133130074, + 0.004653330892324448, + 0.005405236501246691, + 0.008733483962714672, + 0.008320470340549946, + 0.007725327275693417, + 0.007308239582926035, + 0.008660270832479, + 0.00878983549773693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.026347124949097633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1983291655778885, + 0.1946420967578888, + 0.09928324073553085, + 0.05027049779891968, + 0.14901314675807953, + 0.04652062803506851, + 0.04379473999142647, + 0.046097513288259506, + 0.08985896408557892, + 0.08219001442193985 + ], + "credit_map": [ + 0.01757417619228363, + 0.03269847109913826, + 0.032719433307647705, + 0.015750227496027946, + 0.03767644986510277, + 0.020518003031611443, + 0.02242315001785755, + 0.018969273194670677, + 0.020576119422912598, + 0.014487829059362411, + 0.024160608649253845, + 0.0428694523870945, + 0.02059188298881054, + 0.023211734369397163, + 0.026149563491344452, + 0.022594928741455078, + 0.011978814378380775, + 0.01572747528553009, + 0.006476379465311766, + 0.006251112557947636, + 0.030967852100729942, + 0.01010201033204794, + 0.006033733952790499, + 0.011745858006179333, + 0.005492119118571281, + 0.00797553826123476, + 0.0042721303179860115, + 0.0049815732054412365, + 0.004739663098007441, + 0.006616438273340464, + 0.008203264325857162, + 0.007989768870174885, + 0.009185533039271832, + 0.03810545429587364, + 0.008158579468727112, + 0.009631572291254997, + 0.04829254001379013, + 0.016982119530439377, + 0.007088690996170044, + 0.011568664573132992, + 0.005446913652122021, + 0.005729570519179106, + 0.006263210438191891, + 0.0058593652211129665, + 0.006544264033436775, + 0.00802956335246563, + 0.0043246629647910595, + 0.004323075991123915, + 0.00540919927880168, + 0.007287897635251284, + 0.0033599582966417074, + 0.006411333568394184, + 0.007378304842859507, + 0.005530415568500757, + 0.004750384949147701, + 0.0036672467831522226, + 0.009477785788476467, + 0.01266887690871954, + 0.004242751281708479, + 0.002887164941057563, + 0.004554928280413151, + 0.004133597481995821, + 0.003826684085652232, + 0.004305725451558828, + 0.00993350800126791, + 0.011366537772119045, + 0.013378439471125603, + 0.00834515132009983, + 0.015135315246880054, + 0.012880443595349789, + 0.010412652976810932, + 0.008406919427216053, + 0.007445366121828556, + 0.004504866898059845, + 0.012692003510892391, + 0.01393609493970871, + 0.006575579755008221, + 0.007851524278521538, + 0.015579171478748322, + 0.01360540371388197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01819971203804016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, low-angle wide shot, a viscous, unstoppable torrent of incandescent magma slowly but surely intrudes from the right side of the frame, engulfing the dry, cracked, black earth in the foreground. The core of this magma glows a fiery white, while its surface is a dazzling orange-red fluid, covered by a thin, dark red crust that constantly tears and heals as it flows. As the magma front encounters the cold air, its shiny surface rapidly wrinkles and darkens. With a muffled crackling sound, it rapidly cools and solidifies into a vast expanse of rugged, porous black volcanic rock. Despite solidification, a dark red residual glow still pulses within the deep cracks of the newly formed rock, creating a dramatic contrast with the intense glow of the liquid magma advancing behind it. The scene is illuminated solely by the magma itself, casting dancing shadows amidst the steaming heat and thin sulfurous fumes, creating an apocalyptic atmosphere imbued with primal power, both majestic and awe-inspiring.", + "seed": 21359, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_3638_seed_21359_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": -0.04052754594013095, + "vlm_logprob": -0.04052754594013095, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04052754594013095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1116289421916008 + ], + "opposite_logprob_tokens": [ + -2.736629009246826 + ], + "logprob_reward": -0.1116289421916008, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.1116289421916008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13466215133666992, + 0.11533030867576599, + 0.055189941078424454, + 0.14337466657161713, + 0.22686180472373962, + 0.07657600194215775, + 0.06056569889187813, + 0.09837701171636581, + 0.059714362025260925, + 0.02934798039495945 + ], + "credit_map": [ + 0.01808799058198929, + 0.01384200993925333, + 0.017884012311697006, + 0.0091435257345438, + 0.01923638954758644, + 0.02645101770758629, + 0.016198543831706047, + 0.013818659819662571, + 0.01192746963351965, + 0.009988737292587757, + 0.027148941531777382, + 0.007547926623374224, + 0.009581195190548897, + 0.014275114983320236, + 0.0075463904067873955, + 0.027314534410834312, + 0.007184206508100033, + 0.011829433962702751, + 0.008535566739737988, + 0.005116276443004608, + 0.00830047857016325, + 0.004085847642272711, + 0.0046277353540062904, + 0.005510398652404547, + 0.014081337489187717, + 0.015817109495401382, + 0.018391167744994164, + 0.011404062621295452, + 0.019735608249902725, + 0.012744207866489887, + 0.01579255796968937, + 0.03540860489010811, + 0.02093602903187275, + 0.026309749111533165, + 0.048548996448516846, + 0.023601150140166283, + 0.030217088758945465, + 0.024219943210482597, + 0.020926205441355705, + 0.03210263326764107, + 0.0069153159856796265, + 0.009366834536194801, + 0.01765712909400463, + 0.007448683958500624, + 0.015653008595108986, + 0.004737138282507658, + 0.005991908721625805, + 0.008805984631180763, + 0.009872179478406906, + 0.019754555076360703, + 0.0038304035551846027, + 0.005492157302796841, + 0.007908668369054794, + 0.0058030616492033005, + 0.003526477375999093, + 0.004378196317702532, + 0.015185384079813957, + 0.019046099856495857, + 0.012074223719537258, + 0.010681981220841408, + 0.012906698510050774, + 0.011336388066411018, + 0.009011208079755306, + 0.008135027252137661, + 0.008884199894964695, + 0.005146068055182695, + 0.009167962707579136, + 0.006228859070688486, + 0.006976497825235128, + 0.007334395777434111, + 0.00659731263294816, + 0.009379063732922077, + 0.0041052596643567085, + 0.0035222407896071672, + 0.004311444237828255, + 0.0033583263866603374, + 0.0033545673359185457, + 0.004060501232743263, + 0.0027004224248230457, + 0.003935218323022127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.1116289421916008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2019704133272171, + 0.22608277201652527, + 0.09982505440711975, + 0.06352254748344421, + 0.0744401142001152, + 0.06846795976161957, + 0.03529168665409088, + 0.13461294770240784, + 0.03862403705716133, + 0.05716247484087944 + ], + "credit_map": [ + 0.02776079811155796, + 0.021021820604801178, + 0.030311239883303642, + 0.014479569159448147, + 0.0239133108407259, + 0.039152007550001144, + 0.02053770236670971, + 0.024793948978185654, + 0.02316250652074814, + 0.023043286055326462, + 0.051333893090486526, + 0.02230992540717125, + 0.022639768198132515, + 0.020174603909254074, + 0.01739203929901123, + 0.046026747673749924, + 0.009298251010477543, + 0.013126284815371037, + 0.011158660054206848, + 0.016359586268663406, + 0.01619594730436802, + 0.007892752066254616, + 0.014270229265093803, + 0.011523332446813583, + 0.009189054369926453, + 0.009888673201203346, + 0.010800614953041077, + 0.005525953136384487, + 0.012505985796451569, + 0.004792028572410345, + 0.004474639426916838, + 0.006345592439174652, + 0.008445270359516144, + 0.01260286197066307, + 0.01180305890738964, + 0.006885772570967674, + 0.007986976765096188, + 0.008258940652012825, + 0.00574831385165453, + 0.01270891074091196, + 0.006575800944119692, + 0.009851943701505661, + 0.01045923586934805, + 0.008264702744781971, + 0.013943036086857319, + 0.0067168171517550945, + 0.006320783868432045, + 0.006335628218948841, + 0.004711179528385401, + 0.005463118199259043, + 0.004618179518729448, + 0.004227784462273121, + 0.00650276942178607, + 0.0030102315358817577, + 0.002887108363211155, + 0.0038713118992745876, + 0.020311439409852028, + 0.01618693768978119, + 0.01934255100786686, + 0.011705626733601093, + 0.014276115223765373, + 0.02265963889658451, + 0.012598895467817783, + 0.017531733959913254, + 0.0050874268636107445, + 0.004627133719623089, + 0.005918202456086874, + 0.004181774333119392, + 0.004077444318681955, + 0.004836698062717915, + 0.0038894619792699814, + 0.00600589020177722, + 0.009345415979623795, + 0.008173913694918156, + 0.0063492427580058575, + 0.005412584636360407, + 0.010281245224177837, + 0.005351720843464136, + 0.0042577278800308704, + 0.007990620099008083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02059445157647133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.12385323643684387, + 0.14663803577423096, + 0.04946422949433327, + 0.10701863467693329, + 0.32561206817626953, + 0.07020384818315506, + 0.05775085836648941, + 0.0766245499253273, + 0.027374722063541412, + 0.015459820628166199 + ], + "credit_map": [ + 0.010666093789041042, + 0.015176690183579922, + 0.018400879576802254, + 0.012523447163403034, + 0.01752614974975586, + 0.015123351477086544, + 0.01395344827324152, + 0.020483165979385376, + 0.015335903503000736, + 0.015218192711472511, + 0.026924101635813713, + 0.018169619143009186, + 0.012141077779233456, + 0.016598928719758987, + 0.012796294875442982, + 0.029453899711370468, + 0.0037121574860066175, + 0.0050543127581477165, + 0.0065959785133600235, + 0.00682990113273263, + 0.005886351689696312, + 0.003443272551521659, + 0.007200667634606361, + 0.010741585865616798, + 0.012885472737252712, + 0.013374421745538712, + 0.01828291080892086, + 0.010827855207026005, + 0.018092840909957886, + 0.01042850036174059, + 0.010917111299932003, + 0.012209514155983925, + 0.03635714203119278, + 0.053946979343891144, + 0.06032157316803932, + 0.042784612625837326, + 0.03756183758378029, + 0.032026998698711395, + 0.03190344199538231, + 0.030709458515048027, + 0.008330633863806725, + 0.010895974934101105, + 0.012012504041194916, + 0.007598249241709709, + 0.014438287355005741, + 0.004893339704722166, + 0.006558327469974756, + 0.005476522725075483, + 0.006783098913729191, + 0.006888205651193857, + 0.008275169879198074, + 0.006828280631452799, + 0.008919277228415012, + 0.004293052479624748, + 0.005472321528941393, + 0.010291446931660175, + 0.00879439152777195, + 0.00972219742834568, + 0.010110718198120594, + 0.010737321339547634, + 0.010502047836780548, + 0.010106648318469524, + 0.008871969766914845, + 0.0077792443335056305, + 0.004606819711625576, + 0.003895170520991087, + 0.0026235459372401237, + 0.0028437627479434013, + 0.003748370800167322, + 0.0033260390628129244, + 0.0024476100225001574, + 0.00388340069912374, + 0.001795001677237451, + 0.002257924061268568, + 0.0020777652971446514, + 0.0014360587811097503, + 0.0020123543217778206, + 0.0022012738045305014, + 0.0013541877269744873, + 0.002325252862647176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.005950947757810354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06941226124763489, + 0.1264830380678177, + 0.03467484936118126, + 0.10143101215362549, + 0.13706724345684052, + 0.18718814849853516, + 0.12465540319681168, + 0.1205274686217308, + 0.07305652648210526, + 0.02550407312810421 + ], + "credit_map": [ + 0.007600291632115841, + 0.00725281098857522, + 0.011259356513619423, + 0.007845431566238403, + 0.008584251627326012, + 0.00879882276058197, + 0.007915769703686237, + 0.010155527852475643, + 0.014864434488117695, + 0.012599931098520756, + 0.032795075327157974, + 0.010724284686148167, + 0.008753160014748573, + 0.008791814558207989, + 0.010552908293902874, + 0.027401424944400787, + 0.003566760802641511, + 0.004204283002763987, + 0.007229601498693228, + 0.004673130344599485, + 0.0035170684568583965, + 0.0021269195713102818, + 0.0044770087115466595, + 0.004880073480308056, + 0.008311820216476917, + 0.01345817930996418, + 0.015967415645718575, + 0.010318842716515064, + 0.02062755823135376, + 0.010679754428565502, + 0.012400157749652863, + 0.009667283855378628, + 0.013836597092449665, + 0.022261125966906548, + 0.01977708376944065, + 0.01482982188463211, + 0.014452587813138962, + 0.012230266816914082, + 0.00999523140490055, + 0.02968451753258705, + 0.017901312559843063, + 0.027832351624965668, + 0.03154290094971657, + 0.022724322974681854, + 0.0384170338511467, + 0.015009491704404354, + 0.016641300171613693, + 0.01711944118142128, + 0.014980655163526535, + 0.02919701673090458, + 0.011337616480886936, + 0.013869189657270908, + 0.019654659554362297, + 0.011286523193120956, + 0.012344085611402988, + 0.011985665187239647, + 0.015442242845892906, + 0.011507968418300152, + 0.019592102617025375, + 0.015271130949258804, + 0.012867707759141922, + 0.01630924642086029, + 0.013847245834767818, + 0.01568981632590294, + 0.015259183011949062, + 0.008373701013624668, + 0.01017357874661684, + 0.0075745051726698875, + 0.007586035877466202, + 0.009050160646438599, + 0.0053386082872748375, + 0.009700758382678032, + 0.0038427768740803003, + 0.0029807924292981625, + 0.0038744588382542133, + 0.0030577746219933033, + 0.004377377685159445, + 0.00224692584015429, + 0.0023513699416071177, + 0.002772596664726734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 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 camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359546929597855 + ], + "opposite_logprob_tokens": [ + -3.6513595581054688 + ], + "logprob_reward": -0.026359546929597855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359546929597855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09260742366313934, + 0.10462518781423569, + 0.06776729971170425, + 0.06700393557548523, + 0.13888561725616455, + 0.17367714643478394, + 0.150993213057518, + 0.11558333784341812, + 0.0689590722322464, + 0.019897809252142906 + ], + "credit_map": [ + 0.009839074686169624, + 0.010945194400846958, + 0.014970939606428146, + 0.016112590208649635, + 0.013883800245821476, + 0.009478829801082611, + 0.009013267233967781, + 0.008363720029592514, + 0.011206770315766335, + 0.009883481077849865, + 0.027253136038780212, + 0.011536644771695137, + 0.007427753880620003, + 0.010833606123924255, + 0.007114826235920191, + 0.019368963316082954, + 0.00678071565926075, + 0.009396780282258987, + 0.011845670640468597, + 0.012403752654790878, + 0.006651710718870163, + 0.004731372464448214, + 0.008782152086496353, + 0.0071751419454813, + 0.0075878263451159, + 0.007987112738192081, + 0.010256444104015827, + 0.009064551442861557, + 0.010025249794125557, + 0.006046999711543322, + 0.007027048617601395, + 0.00900870468467474, + 0.012326622381806374, + 0.01688377931714058, + 0.02538185939192772, + 0.018865853548049927, + 0.019323229789733887, + 0.017172252759337425, + 0.01232029590755701, + 0.016611719503998756, + 0.01687542349100113, + 0.02508421801030636, + 0.03633204475045204, + 0.017079660668969154, + 0.03540815785527229, + 0.012146771885454655, + 0.013400668278336525, + 0.0173502080142498, + 0.015896128490567207, + 0.01673378422856331, + 0.027386292815208435, + 0.020608948543667793, + 0.02389395236968994, + 0.011121967807412148, + 0.018084099516272545, + 0.017268028110265732, + 0.012782598845660686, + 0.028628438711166382, + 0.007850917056202888, + 0.008051788434386253, + 0.01267095934599638, + 0.02811351977288723, + 0.009157796390354633, + 0.008327321149408817, + 0.016266893595457077, + 0.008249148726463318, + 0.007982427254319191, + 0.006177536677569151, + 0.007155855186283588, + 0.010120682418346405, + 0.003795410506427288, + 0.009211115539073944, + 0.0023610186763107777, + 0.0023673218674957752, + 0.0035383827053010464, + 0.002151565160602331, + 0.002471239771693945, + 0.0026671828236430883, + 0.0019084976520389318, + 0.0024326008278876543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.026359546929597855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, low-angle wide shot, a viscous, unstoppable torrent of incandescent magma slowly but surely intrudes from the right side of the frame, engulfing the dry, cracked, black earth in the foreground. The core of this magma glows a fiery white, while its surface is a dazzling orange-red fluid, covered by a thin, dark red crust that constantly tears and heals as it flows. As the magma front encounters the cold air, its shiny surface rapidly wrinkles and darkens. With a muffled crackling sound, it rapidly cools and solidifies into a vast expanse of rugged, porous black volcanic rock. Despite solidification, a dark red residual glow still pulses within the deep cracks of the newly formed rock, creating a dramatic contrast with the intense glow of the liquid magma advancing behind it. The scene is illuminated solely by the magma itself, casting dancing shadows amidst the steaming heat and thin sulfurous fumes, creating an apocalyptic atmosphere imbued with primal power, both majestic and awe-inspiring.", + "seed": 21360, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_12/index_3638_seed_21360_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": -0.06333755012601613, + "vlm_logprob": -0.06333755012601613, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.06333755012601613, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1932993084192276 + ], + "opposite_logprob_tokens": [ + -2.0682992935180664 + ], + "logprob_reward": -0.1932993084192276, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1932993084192276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20987899601459503, + 0.04904121160507202, + 0.14249484241008759, + 0.022941723465919495, + 0.037260446697473526, + 0.1247226893901825, + 0.08876136690378189, + 0.04774339869618416, + 0.15440787374973297, + 0.12274745106697083 + ], + "credit_map": [ + 0.020301425829529762, + 0.023655971512198448, + 0.037247899919748306, + 0.041077472269535065, + 0.02101431041955948, + 0.02693183347582817, + 0.014092223718762398, + 0.025557851418852806, + 0.00519234873354435, + 0.005080945324152708, + 0.006527077406644821, + 0.008517236448824406, + 0.0056661381386220455, + 0.004760066978633404, + 0.006575717590749264, + 0.00672168331220746, + 0.006929727271199226, + 0.034503571689128876, + 0.0073379711247980595, + 0.008371190167963505, + 0.020169585943222046, + 0.052037354558706284, + 0.0074699814431369305, + 0.005675449036061764, + 0.0016466035740450025, + 0.0019902882631868124, + 0.00261277356185019, + 0.004621247760951519, + 0.002634792821481824, + 0.003749644150957465, + 0.00220790458843112, + 0.0034784667659550905, + 0.004586668219417334, + 0.004819014575332403, + 0.004874588456004858, + 0.004071819595992565, + 0.005923967808485031, + 0.005479258019477129, + 0.004198059439659119, + 0.003307072911411524, + 0.007653738837689161, + 0.020384518429636955, + 0.016216475516557693, + 0.02103939838707447, + 0.008120827376842499, + 0.015762077644467354, + 0.01839875616133213, + 0.017146892845630646, + 0.0074228025041520596, + 0.00895944144576788, + 0.016378384083509445, + 0.012542786076664925, + 0.009257880970835686, + 0.011530175805091858, + 0.011234127916395664, + 0.011435778811573982, + 0.007643531076610088, + 0.0066819326020777225, + 0.005132861901074648, + 0.00824501272290945, + 0.004790704697370529, + 0.004832782316952944, + 0.004916069563478231, + 0.00550050288438797, + 0.02029598131775856, + 0.01956585980951786, + 0.023518286645412445, + 0.018061818554997444, + 0.018510613590478897, + 0.02465691603720188, + 0.013053865171968937, + 0.01674453169107437, + 0.008898444473743439, + 0.01501601841300726, + 0.014859740622341633, + 0.014292014762759209, + 0.018210966140031815, + 0.020065147429704666, + 0.018802717328071594, + 0.012602396309375763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.1932993084192276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498318374156952 + ], + "opposite_logprob_tokens": [ + -2.929983139038086 + ], + "logprob_reward": -0.05498318374156952, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498318374156952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061781980097293854, + 0.03126276656985283, + 0.49343833327293396, + 0.031326767057180405, + 0.10083438456058502, + 0.0968874916434288, + 0.043415069580078125, + 0.03569542616605759, + 0.02921323850750923, + 0.07614448666572571 + ], + "credit_map": [ + 0.005100168753415346, + 0.011130941100418568, + 0.008676322177052498, + 0.009298749268054962, + 0.008151126094162464, + 0.008742790669202805, + 0.00521300220862031, + 0.0054688830859959126, + 0.0032813893631100655, + 0.0033988154027611017, + 0.005368856713175774, + 0.005076263099908829, + 0.003426777431741357, + 0.0033557794522494078, + 0.003977985121309757, + 0.0033769013825803995, + 0.017220323905348778, + 0.1485457569360733, + 0.01970699615776539, + 0.019227173179388046, + 0.060576021671295166, + 0.20200353860855103, + 0.015399634838104248, + 0.010758915916085243, + 0.003003045218065381, + 0.0029182389844208956, + 0.004174324218183756, + 0.004808944184333086, + 0.004502500873059034, + 0.004520386923104525, + 0.0039453026838600636, + 0.003454029094427824, + 0.009916217997670174, + 0.011873982846736908, + 0.018029170110821724, + 0.01094574760645628, + 0.01172960177063942, + 0.013440784066915512, + 0.014004153199493885, + 0.01089473720639944, + 0.011881253682076931, + 0.02144348807632923, + 0.009554184973239899, + 0.00640180753543973, + 0.013353909365832806, + 0.016377458348870277, + 0.010265029035508633, + 0.007610364351421595, + 0.0056045507080852985, + 0.0053526065312325954, + 0.004170939326286316, + 0.00331302871927619, + 0.006505920086055994, + 0.007537513040006161, + 0.005837101023644209, + 0.005093416664749384, + 0.00481233187019825, + 0.00480360072106123, + 0.0048296344466507435, + 0.005169272888451815, + 0.0034210190642625093, + 0.004062320105731487, + 0.003842872567474842, + 0.0047543770633637905, + 0.0031744048465043306, + 0.005107917357236147, + 0.004055890254676342, + 0.0030598288867622614, + 0.00395650789141655, + 0.004072428215295076, + 0.003084559226408601, + 0.0027017020620405674, + 0.005646184086799622, + 0.008728476241230965, + 0.00714526791125536, + 0.0070342724211514, + 0.01317133754491806, + 0.015768077224493027, + 0.010459164157509804, + 0.008191712200641632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 5 + ], + "selected_reward": -0.05498318374156952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080567613244057 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080567613244057, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080567613244057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19406668841838837, + 0.048432108014822006, + 0.1836153268814087, + 0.0845520943403244, + 0.18772368133068085, + 0.0834825336933136, + 0.06233040988445282, + 0.05389818176627159, + 0.04912777990102768, + 0.052771132439374924 + ], + "credit_map": [ + 0.014957955107092857, + 0.036648593842983246, + 0.027371391654014587, + 0.02614511176943779, + 0.0234847255051136, + 0.03080740012228489, + 0.01763259433209896, + 0.017018916085362434, + 0.006109219044446945, + 0.005063326563686132, + 0.00688096322119236, + 0.006657441612333059, + 0.005105894524604082, + 0.006547565571963787, + 0.006692912895232439, + 0.005374784115701914, + 0.010354647412896156, + 0.037182439118623734, + 0.016434114426374435, + 0.018049880862236023, + 0.02343728579580784, + 0.0552879273891449, + 0.012840471230447292, + 0.010028566233813763, + 0.007074762135744095, + 0.0057201203890144825, + 0.014572065323591232, + 0.014390246942639351, + 0.008069008588790894, + 0.011132283136248589, + 0.009033415466547012, + 0.01456019002944231, + 0.0187180507928133, + 0.018479015678167343, + 0.03635271638631821, + 0.025076938793063164, + 0.018723655492067337, + 0.021663237363100052, + 0.023766031488776207, + 0.024944033473730087, + 0.009920581243932247, + 0.009601750411093235, + 0.011599486693739891, + 0.008053286001086235, + 0.016439730301499367, + 0.012292783707380295, + 0.008567918092012405, + 0.007006998173892498, + 0.005034164991229773, + 0.007226086687296629, + 0.00982465036213398, + 0.0051647815853357315, + 0.009254583157598972, + 0.00893923919647932, + 0.010028881020843983, + 0.0068580214865505695, + 0.007673138752579689, + 0.006301774177700281, + 0.006023661699146032, + 0.007517725229263306, + 0.004709766246378422, + 0.006548114120960236, + 0.005546288099139929, + 0.009577710181474686, + 0.0039724381640553474, + 0.0052788276225328445, + 0.007135579828172922, + 0.00835948996245861, + 0.00476553849875927, + 0.00571404816582799, + 0.007206223905086517, + 0.006695634685456753, + 0.004171811975538731, + 0.006231875624507666, + 0.0062200650572776794, + 0.006686090491712093, + 0.007432141341269016, + 0.009233830496668816, + 0.00727851502597332, + 0.005516799166798592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.016080567613244057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381164625287056 + ], + "opposite_logprob_tokens": [ + -3.288116455078125 + ], + "logprob_reward": -0.0381164625287056, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381164625287056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2461642175912857, + 0.047584131360054016, + 0.08511969447135925, + 0.0846874862909317, + 0.26981908082962036, + 0.09184444695711136, + 0.037966638803482056, + 0.029766755178570747, + 0.03475108742713928, + 0.07229644060134888 + ], + "credit_map": [ + 0.0329052098095417, + 0.04243701696395874, + 0.04001617059111595, + 0.03058977983891964, + 0.02469482272863388, + 0.02895391918718815, + 0.021095436066389084, + 0.025471849367022514, + 0.0057251788675785065, + 0.004600378684699535, + 0.005684498231858015, + 0.007300126366317272, + 0.005295796785503626, + 0.005116535350680351, + 0.006439576391130686, + 0.007422036025673151, + 0.004190580453723669, + 0.030534276738762856, + 0.004529010504484177, + 0.005102965980768204, + 0.006419505458325148, + 0.026838084682822227, + 0.003368867328390479, + 0.0041364035569131374, + 0.005777785088866949, + 0.007400604896247387, + 0.009297766722738743, + 0.013658938929438591, + 0.012972671538591385, + 0.013447857461869717, + 0.01313355378806591, + 0.008998305536806583, + 0.02579495869576931, + 0.025018997490406036, + 0.05224209651350975, + 0.037792839109897614, + 0.023585515096783638, + 0.027525275945663452, + 0.03641834855079651, + 0.04144103825092316, + 0.00873842928558588, + 0.01612776145339012, + 0.011189167387783527, + 0.011075693182647228, + 0.007921316660940647, + 0.013440318405628204, + 0.013232461176812649, + 0.010119295679032803, + 0.003289928426966071, + 0.0036684793885797262, + 0.007420722860842943, + 0.004247993230819702, + 0.004791466053575277, + 0.005398726090788841, + 0.004002726171165705, + 0.0051465947180986404, + 0.002608019392937422, + 0.0036328821443021297, + 0.0037674715276807547, + 0.004899398889392614, + 0.0032562180422246456, + 0.0029955795034766197, + 0.0045326766557991505, + 0.004074507392942905, + 0.0038615374360233545, + 0.004313548561185598, + 0.00419393926858902, + 0.004414957482367754, + 0.004112670663744211, + 0.0050121331587433815, + 0.004661012440919876, + 0.004181285854429007, + 0.0058482736349105835, + 0.008284414187073708, + 0.010816213674843311, + 0.009585881605744362, + 0.009224767796695232, + 0.008502674289047718, + 0.010004440322518349, + 0.010029769502580166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0381164625287056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot static, low-angle, and wide?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a viscous torrent of incandescent magma intrude from the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the magma core glowing fiery white with a surface of dazzling orange-red fluid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the magma rapidly cool and solidify into rugged, porous black volcanic rock with a dark red residual glow in the cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated solely by the magma, casting dancing shadows amidst steaming heat and thin sulfurous fumes?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28834402561187744, + 0.06520407646894455, + 0.19554820656776428, + 0.10583458840847015, + 0.08002466708421707, + 0.07355406135320663, + 0.036559488624334335, + 0.045631229877471924, + 0.03854687511920929, + 0.07075279951095581 + ], + "credit_map": [ + 0.03112158551812172, + 0.0458841547369957, + 0.05537281185388565, + 0.0419071801006794, + 0.028216112405061722, + 0.032489851117134094, + 0.02481137029826641, + 0.028540953993797302, + 0.0060263704508543015, + 0.0069688246585428715, + 0.010870655998587608, + 0.006804285570979118, + 0.004820168484002352, + 0.006460070144385099, + 0.006346958689391613, + 0.016906749457120895, + 0.010210930369794369, + 0.04704033583402634, + 0.018294557929039, + 0.014494938775897026, + 0.02306857705116272, + 0.06630987673997879, + 0.007580856792628765, + 0.008548139594495296, + 0.008500030264258385, + 0.008523814380168915, + 0.02378152683377266, + 0.018032612279057503, + 0.010389069095253944, + 0.014110753312706947, + 0.010025059804320335, + 0.012471734546124935, + 0.006565125193446875, + 0.009903225116431713, + 0.01086705643683672, + 0.010186946019530296, + 0.011309384368360043, + 0.01038193516433239, + 0.011914795264601707, + 0.008896208368241787, + 0.005537484772503376, + 0.009956053458154202, + 0.009541195817291737, + 0.00882486067712307, + 0.008137295953929424, + 0.009800312109291553, + 0.012297353707253933, + 0.009459513239562511, + 0.0038961449172347784, + 0.003706237766891718, + 0.005130085628479719, + 0.00386023405008018, + 0.004126632586121559, + 0.004923863802105188, + 0.0053239730186760426, + 0.00559232197701931, + 0.003981985151767731, + 0.005564179737120867, + 0.0071990774013102055, + 0.005704887211322784, + 0.005906091537326574, + 0.005166447255760431, + 0.00668421620503068, + 0.00542434910312295, + 0.0042221518233418465, + 0.005271356087177992, + 0.004813007544726133, + 0.005961012095212936, + 0.004353586118668318, + 0.004424585495144129, + 0.004979439079761505, + 0.0045217410661280155, + 0.006570430938154459, + 0.010265957564115524, + 0.007758034393191338, + 0.006809580139815807, + 0.01006745919585228, + 0.009929546155035496, + 0.009384327568113804, + 0.009967470541596413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.014208228327333927 + } + ], + "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/0000022/group_13/index_tensor([7742, 7742])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_tensor([7742, 7742])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a6bdc05259478a65abe10f75e0ecdcfdffb44494 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_tensor([7742, 7742])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, one-shot, long take begins with a slowly ascending overhead angle, then shifts to a steady, circling motion around the subject. In the frame, the Monkey King, clad in an intricately carved suit of golden chain mail, intricately engraved with dragon and auspicious cloud motifs, hovers majestically above a swirling sea of clouds. He wears a phoenix-winged, purple-gold crown, two slender pheasant feathers trembling in the celestial winds. His scarlet cape flutters majestically behind him like a battle banner. His face is defiant, his fiery eyes gleaming with an unyielding radiance. His tightly pursed lips betray supreme majesty and confidence. In his powerful hands, he clutches the legendary Ruyi Jingu Bang. The black iron shaft reflects a deep metallic texture in the light, while the golden hoops at each end reflect a dazzling brilliance. As he slowly and leisurely circles, the camera moves in sync, revealing the boundless, layered clouds beneath his feet and the magnificent celestial palace faintly visible deep within them—its jade-like buildings, soaring eaves, and angled corners looming through the mist, appearing incredibly majestic and distant. Most strikingly, the thousand-foot-long golden rays of light, like divine swords, pierce the thick clouds, forming distinct, Tyndall-effect beams imbued with dust and spiritual particles, illuminating the Monkey King's golden armor with a gleaming brilliance. The interplay of light and shadow imbues the entire scene with an epic atmosphere of sacredness, aloofness, and immense power.", + "seed": 32494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32494_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": -0.04371602712199092, + "vlm_logprob": -0.04371602712199092, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.04371602712199092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.010446391999721527 + ], + "opposite_logprob_tokens": [ + -4.885446548461914 + ], + "logprob_reward": -0.010446391999721527, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010446391999721527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11295534670352936, + 0.18089459836483002, + 0.05756225436925888, + 0.034991148859262466, + 0.04496132209897041, + 0.05536726489663124, + 0.11504779011011124, + 0.1998218148946762, + 0.08757776767015457, + 0.11082074791193008 + ], + "credit_map": [ + 0.011635958217084408, + 0.019802145659923553, + 0.019403105601668358, + 0.015838412567973137, + 0.007879753597080708, + 0.010645702481269836, + 0.014870161190629005, + 0.012880109250545502, + 0.012507722713053226, + 0.05988369882106781, + 0.05764376372098923, + 0.00943381804972887, + 0.009122400544583797, + 0.008544132113456726, + 0.011934983544051647, + 0.01182408444583416, + 0.008589815348386765, + 0.008747757412493229, + 0.007577167358249426, + 0.005571666639298201, + 0.007345881313085556, + 0.008104778826236725, + 0.007348115090280771, + 0.004277073312550783, + 0.004102497827261686, + 0.004388059955090284, + 0.004617059137672186, + 0.0053399852477014065, + 0.0038710462395101786, + 0.004817028064280748, + 0.004624710418283939, + 0.0032307615038007498, + 0.006612687837332487, + 0.009504913352429867, + 0.005991501267999411, + 0.0044592078775167465, + 0.0061076609417796135, + 0.005128209479153156, + 0.003901990130543709, + 0.0032551540061831474, + 0.009608486667275429, + 0.007270127534866333, + 0.007393485400825739, + 0.0069217984564602375, + 0.007384869735687971, + 0.005448725074529648, + 0.006283510476350784, + 0.005056262016296387, + 0.012944428250193596, + 0.01573924906551838, + 0.019071951508522034, + 0.015133943408727646, + 0.008801965042948723, + 0.016166280955076218, + 0.009616462513804436, + 0.017573513090610504, + 0.023577159270644188, + 0.013578195124864578, + 0.010504063218832016, + 0.02318563684821129, + 0.01118963211774826, + 0.0528930127620697, + 0.013240218162536621, + 0.051653895527124405, + 0.01039245817810297, + 0.006650734227150679, + 0.008468281477689743, + 0.0070245699025690556, + 0.01218192558735609, + 0.018047191202640533, + 0.0076567563228309155, + 0.017155855894088745, + 0.013781999237835407, + 0.013564305379986763, + 0.014154622331261635, + 0.017689457163214684, + 0.016444697976112366, + 0.013575758785009384, + 0.00997471809387207, + 0.011635185219347477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.010446391999721527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.007634746376425028 + ], + "opposite_logprob_tokens": [ + -4.88263463973999 + ], + "logprob_reward": -0.007634746376425028, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007634746376425028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07841142266988754, + 0.11009945720434189, + 0.08932521939277649, + 0.058151114732027054, + 0.02724972739815712, + 0.027980880811810493, + 0.0642869770526886, + 0.16518372297286987, + 0.16428928077220917, + 0.21502216160297394 + ], + "credit_map": [ + 0.009518844075500965, + 0.011959773488342762, + 0.016907591372728348, + 0.006748327985405922, + 0.005953934974968433, + 0.009200363419950008, + 0.00815765280276537, + 0.009964946657419205, + 0.013215138576924801, + 0.02686590515077114, + 0.023491892963647842, + 0.008343130350112915, + 0.009408429265022278, + 0.011149575933814049, + 0.008647114038467407, + 0.008978274650871754, + 0.018577249720692635, + 0.016873138025403023, + 0.012801581993699074, + 0.009712433442473412, + 0.009626345708966255, + 0.00898397620767355, + 0.007686345838010311, + 0.005064160097390413, + 0.006256123073399067, + 0.004286929965019226, + 0.007857692427933216, + 0.008586660958826542, + 0.007878360338509083, + 0.008233912289142609, + 0.007445533759891987, + 0.007605903781950474, + 0.004146169871091843, + 0.004356097429990768, + 0.003642978146672249, + 0.004454716108739376, + 0.002743491670116782, + 0.004388214088976383, + 0.0019602752290666103, + 0.0015577870653942227, + 0.0041798814199864864, + 0.003494423581287265, + 0.003559412667527795, + 0.003704698523506522, + 0.003462263848632574, + 0.004070708528161049, + 0.0022763630840927362, + 0.003233130555599928, + 0.008208398707211018, + 0.008780538104474545, + 0.007512588053941727, + 0.009682774543762207, + 0.005708782467991114, + 0.010314328595995903, + 0.0071433233097195625, + 0.006936250254511833, + 0.01653069071471691, + 0.014751404523849487, + 0.010168420150876045, + 0.018797915428876877, + 0.008789864368736744, + 0.04411110281944275, + 0.008235790766775608, + 0.0437985435128212, + 0.01281952764838934, + 0.013039479963481426, + 0.011144053190946579, + 0.014888439327478409, + 0.015514276921749115, + 0.04391384497284889, + 0.010607034899294376, + 0.04236263409256935, + 0.017774321138858795, + 0.010653041303157806, + 0.01789962872862816, + 0.02620423026382923, + 0.027951663359999657, + 0.03765113651752472, + 0.012029004283249378, + 0.06485916674137115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.007634746376425028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.07895011454820633 + ], + "opposite_logprob_tokens": [ + -2.5789501667022705 + ], + "logprob_reward": -0.07895011454820633, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895011454820633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12105704098939896, + 0.07515299320220947, + 0.18284748494625092, + 0.07104377448558807, + 0.10264191776514053, + 0.07916482537984848, + 0.11168273538351059, + 0.09164908528327942, + 0.09062313288450241, + 0.07413706183433533 + ], + "credit_map": [ + 0.012775704264640808, + 0.016864854842424393, + 0.03074468858540058, + 0.009202429093420506, + 0.006731342524290085, + 0.011950439773499966, + 0.010644646361470222, + 0.0221429243683815, + 0.01163332536816597, + 0.013300027698278427, + 0.01718256250023842, + 0.008719075471162796, + 0.005258406512439251, + 0.006698718760162592, + 0.004638206213712692, + 0.007722664624452591, + 0.03455197811126709, + 0.03486743941903114, + 0.024285754188895226, + 0.01988859288394451, + 0.026147058233618736, + 0.015294504351913929, + 0.013989773578941822, + 0.013822365552186966, + 0.006642740685492754, + 0.006567773409187794, + 0.00934798363596201, + 0.009661449119448662, + 0.009273365139961243, + 0.009497993625700474, + 0.012326392345130444, + 0.007726073730736971, + 0.014360226690769196, + 0.014510562643408775, + 0.012602140195667744, + 0.010337239131331444, + 0.01622931659221649, + 0.0128916185349226, + 0.00702074496075511, + 0.01469005923718214, + 0.008531124331057072, + 0.014311274513602257, + 0.00953883770853281, + 0.011217961087822914, + 0.010484767146408558, + 0.009390282444655895, + 0.0067851790226995945, + 0.008905400522053242, + 0.01642627641558647, + 0.016017209738492966, + 0.01622588373720646, + 0.013680865988135338, + 0.006735124625265598, + 0.011947574093937874, + 0.008022364228963852, + 0.022627433761954308, + 0.011687785387039185, + 0.016759736463427544, + 0.009061002172529697, + 0.011408675462007523, + 0.01001876313239336, + 0.011867151595652103, + 0.007478479761630297, + 0.01336748618632555, + 0.005829634610563517, + 0.004250481724739075, + 0.004627208225429058, + 0.003999243024736643, + 0.008465726859867573, + 0.030337750911712646, + 0.003864733502268791, + 0.029248343780636787, + 0.007066672667860985, + 0.004333862569183111, + 0.011336457915604115, + 0.011724362149834633, + 0.0074336412362754345, + 0.010647757910192013, + 0.004334889352321625, + 0.017259415239095688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.07895011454820633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24042360484600067, + 0.1312471479177475, + 0.06860507279634476, + 0.050888147205114365, + 0.08384127914905548, + 0.0509430430829525, + 0.10089676827192307, + 0.09132517129182816, + 0.05647624656558037, + 0.12535350024700165 + ], + "credit_map": [ + 0.020650207996368408, + 0.031621236354112625, + 0.06778416782617569, + 0.021250201389193535, + 0.014324727468192577, + 0.017624644562602043, + 0.018095973879098892, + 0.04907245934009552, + 0.010555252432823181, + 0.031657904386520386, + 0.027892867103219032, + 0.012831959873437881, + 0.008214997127652168, + 0.011722618713974953, + 0.010441443882882595, + 0.01793012022972107, + 0.010188398882746696, + 0.009720535017549992, + 0.00835677981376648, + 0.006618117447942495, + 0.009212910197675228, + 0.010812932625412941, + 0.008406360633671284, + 0.005289048887789249, + 0.0069371676072478294, + 0.004086267203092575, + 0.0073785060085356236, + 0.010224384255707264, + 0.0060668326914310455, + 0.00586722930893302, + 0.005557181779295206, + 0.004770580213516951, + 0.01421714574098587, + 0.013775266706943512, + 0.012486173771321774, + 0.01050860807299614, + 0.012281260453164577, + 0.009017350152134895, + 0.0056423647329211235, + 0.005913118366152048, + 0.008231788873672485, + 0.00790899433195591, + 0.007545663043856621, + 0.009026801213622093, + 0.005384123884141445, + 0.00432047201320529, + 0.005016060080379248, + 0.003509142203256488, + 0.020261364057660103, + 0.017251502722501755, + 0.012699533253908157, + 0.010860958136618137, + 0.00837021879851818, + 0.010065189562737942, + 0.008835799060761929, + 0.012552207335829735, + 0.010634574107825756, + 0.014062164351344109, + 0.012013430707156658, + 0.007569952867925167, + 0.00857181753963232, + 0.017999468371272087, + 0.006519044283777475, + 0.013954726047813892, + 0.006137666758149862, + 0.006199137773364782, + 0.007552316877990961, + 0.005531118251383305, + 0.01014743372797966, + 0.008155972696840763, + 0.0037823354359716177, + 0.008970269002020359, + 0.01146761979907751, + 0.010358618572354317, + 0.012439762242138386, + 0.012387373484671116, + 0.013751208782196045, + 0.019345082342624664, + 0.010046232491731644, + 0.03555761277675629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.008634849451482296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.11291403323411942 + ], + "opposite_logprob_tokens": [ + -2.2379140853881836 + ], + "logprob_reward": -0.11291403323411942, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291403323411942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13192515075206757, + 0.23410873115062714, + 0.06913694739341736, + 0.030769020318984985, + 0.108161561191082, + 0.05440004542469978, + 0.23455078899860382, + 0.060506731271743774, + 0.03174273669719696, + 0.0446983240544796 + ], + "credit_map": [ + 0.01404642965644598, + 0.017228461802005768, + 0.02350292168557644, + 0.018191931769251823, + 0.011252665892243385, + 0.01039741188287735, + 0.018630286678671837, + 0.018675044178962708, + 0.015068111009895802, + 0.08879502862691879, + 0.0849723368883133, + 0.009777574799954891, + 0.010572144761681557, + 0.007776836398988962, + 0.006773053668439388, + 0.010373659431934357, + 0.010237744078040123, + 0.01264221966266632, + 0.009276947937905788, + 0.006830743048340082, + 0.009831368923187256, + 0.009585649706423283, + 0.006391208153218031, + 0.00434107007458806, + 0.00428785989060998, + 0.0023479005321860313, + 0.004155439790338278, + 0.005891557317227125, + 0.0035194791853427887, + 0.0032413944136351347, + 0.0035066690761595964, + 0.003818720346316695, + 0.01724390685558319, + 0.01942971721291542, + 0.016231343150138855, + 0.01417219452559948, + 0.017031054943799973, + 0.011088320054113865, + 0.006891653873026371, + 0.006073369644582272, + 0.00724522490054369, + 0.009140629321336746, + 0.007896003313362598, + 0.011117085814476013, + 0.004971708171069622, + 0.005781062413007021, + 0.004763532429933548, + 0.0034847971983253956, + 0.03172330930829048, + 0.04784862697124481, + 0.028194719925522804, + 0.03745478391647339, + 0.0205236729234457, + 0.030917156487703323, + 0.017862556502223015, + 0.020025968551635742, + 0.006569790653884411, + 0.011104658246040344, + 0.007138457149267197, + 0.006287610623985529, + 0.006028426345437765, + 0.009521422907710075, + 0.005312878638505936, + 0.008543485775589943, + 0.002509607933461666, + 0.002705420134589076, + 0.0025167064741253853, + 0.0021770873572677374, + 0.004050029441714287, + 0.008414610289037228, + 0.00154600921086967, + 0.007823266088962555, + 0.0037429346702992916, + 0.003092297585681081, + 0.005969509948045015, + 0.0060900175012648106, + 0.005682684946805239, + 0.006670739036053419, + 0.0023424371611326933, + 0.011107702739536762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.11291403323411942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, one-shot, long take begins with a slowly ascending overhead angle, then shifts to a steady, circling motion around the subject. In the frame, the Monkey King, clad in an intricately carved suit of golden chain mail, intricately engraved with dragon and auspicious cloud motifs, hovers majestically above a swirling sea of clouds. He wears a phoenix-winged, purple-gold crown, two slender pheasant feathers trembling in the celestial winds. His scarlet cape flutters majestically behind him like a battle banner. His face is defiant, his fiery eyes gleaming with an unyielding radiance. His tightly pursed lips betray supreme majesty and confidence. In his powerful hands, he clutches the legendary Ruyi Jingu Bang. The black iron shaft reflects a deep metallic texture in the light, while the golden hoops at each end reflect a dazzling brilliance. As he slowly and leisurely circles, the camera moves in sync, revealing the boundless, layered clouds beneath his feet and the magnificent celestial palace faintly visible deep within them—its jade-like buildings, soaring eaves, and angled corners looming through the mist, appearing incredibly majestic and distant. Most strikingly, the thousand-foot-long golden rays of light, like divine swords, pierce the thick clouds, forming distinct, Tyndall-effect beams imbued with dust and spiritual particles, illuminating the Monkey King's golden armor with a gleaming brilliance. The interplay of light and shadow imbues the entire scene with an epic atmosphere of sacredness, aloofness, and immense power.", + "seed": 32495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32495_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": -0.11940772458910942, + "vlm_logprob": -0.11940772458910942, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.11940772458910942, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.012255130335688591 + ], + "opposite_logprob_tokens": [ + -4.7622551918029785 + ], + "logprob_reward": -0.012255130335688591, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012255130335688591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16687269508838654, + 0.10239891707897186, + 0.12249038368463516, + 0.12612123787403107, + 0.07716517895460129, + 0.04481237754225731, + 0.06435523927211761, + 0.10483650118112564, + 0.10540773719549179, + 0.08553969115018845 + ], + "credit_map": [ + 0.034193601459264755, + 0.01929176226258278, + 0.018373247236013412, + 0.01539663691073656, + 0.018452582880854607, + 0.021456623449921608, + 0.015630362555384636, + 0.0240778885781765, + 0.024313485249876976, + 0.010709361173212528, + 0.018916858360171318, + 0.006642613094300032, + 0.008687307126820087, + 0.018784793093800545, + 0.00922140758484602, + 0.005123092792928219, + 0.020585181191563606, + 0.014420967549085617, + 0.023603156208992004, + 0.013051597401499748, + 0.016040032729506493, + 0.01886937953531742, + 0.011801866814494133, + 0.004118212498724461, + 0.02281203493475914, + 0.02033194899559021, + 0.011076304130256176, + 0.013905213214457035, + 0.021214216947555542, + 0.015020848251879215, + 0.013174697756767273, + 0.008585971780121326, + 0.010819582268595695, + 0.010259940288960934, + 0.005621050950139761, + 0.008368754759430885, + 0.012014912441372871, + 0.008929396979510784, + 0.01404795702546835, + 0.007103584706783295, + 0.004747060593217611, + 0.005423829425126314, + 0.00616569584235549, + 0.0050166333094239235, + 0.005974385421723127, + 0.006629696115851402, + 0.005138043314218521, + 0.005717034451663494, + 0.007267339620739222, + 0.0057602617889642715, + 0.004249049816280603, + 0.004790807142853737, + 0.008052119053900242, + 0.016597485169768333, + 0.007631269283592701, + 0.010006903670728207, + 0.01004109438508749, + 0.007986677810549736, + 0.007877795957028866, + 0.009331956505775452, + 0.009708084166049957, + 0.014829396270215511, + 0.027150988578796387, + 0.017910512164235115, + 0.006318467669188976, + 0.005927537567913532, + 0.00626741349697113, + 0.00540469354018569, + 0.007195889484137297, + 0.024805530905723572, + 0.009587716311216354, + 0.03990048170089722, + 0.009063255973160267, + 0.007263582199811935, + 0.007365054916590452, + 0.008361006155610085, + 0.005708834622055292, + 0.015216217376291752, + 0.008942561224102974, + 0.02361917868256569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.012255130335688591 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.0067444778978824615 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.0067444778978824615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067444778978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09101471304893494, + 0.12443889677524567, + 0.07321549206972122, + 0.06729583442211151, + 0.026592247188091278, + 0.03443196788430214, + 0.06624113768339157, + 0.10674667358398438, + 0.3277066946029663, + 0.0823163241147995 + ], + "credit_map": [ + 0.010887837037444115, + 0.017991622909903526, + 0.007803275715559721, + 0.008969577960669994, + 0.012601880356669426, + 0.011041722260415554, + 0.011518809013068676, + 0.010199996642768383, + 0.024082083255052567, + 0.008515133522450924, + 0.031217778101563454, + 0.00791692640632391, + 0.007037019822746515, + 0.030518746003508568, + 0.009187255054712296, + 0.005963963456451893, + 0.013843175023794174, + 0.00851961225271225, + 0.01034454070031643, + 0.007501020561903715, + 0.011854690499603748, + 0.008144225925207138, + 0.009972665458917618, + 0.0030355642084032297, + 0.01703096367418766, + 0.009139403700828552, + 0.011580180376768112, + 0.004884500987827778, + 0.008607671596109867, + 0.007108277641236782, + 0.005081695504486561, + 0.003863144665956497, + 0.0029310164973139763, + 0.003997025545686483, + 0.0017669865628704429, + 0.0027366357389837503, + 0.005231549963355064, + 0.004168293438851833, + 0.0030082184821367264, + 0.0027525208424776793, + 0.004874031059443951, + 0.0020088069140911102, + 0.0034691374748945236, + 0.0033318912610411644, + 0.0034718303941190243, + 0.007714380044490099, + 0.005026538856327534, + 0.004535357002168894, + 0.009529336355626583, + 0.013638255186378956, + 0.005285259801894426, + 0.004759208299219608, + 0.007636406924575567, + 0.010817422531545162, + 0.006395725067704916, + 0.008179526776075363, + 0.012796378694474697, + 0.013097656890749931, + 0.013644689694046974, + 0.009251882322132587, + 0.01217799074947834, + 0.014201590791344643, + 0.014912606216967106, + 0.016663886606693268, + 0.02308955416083336, + 0.023806266486644745, + 0.024774722754955292, + 0.020807988941669464, + 0.03409599885344505, + 0.06568814814090729, + 0.03255215659737587, + 0.10289186984300613, + 0.008901679888367653, + 0.006472191773355007, + 0.005836108233779669, + 0.007053885143250227, + 0.005782044492661953, + 0.015470552258193493, + 0.007887030951678753, + 0.024912841618061066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.0067444778978824615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.25205764174461365 + ], + "opposite_logprob_tokens": [ + -1.502057671546936 + ], + "logprob_reward": -0.25205764174461365, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205764174461365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2078494131565094, + 0.09852935373783112, + 0.18117249011993408, + 0.15540891885757446, + 0.04119765758514404, + 0.1043926477432251, + 0.10582451522350311, + 0.03168781101703644, + 0.047984883189201355, + 0.025952311232686043 + ], + "credit_map": [ + 0.027389192953705788, + 0.0330984890460968, + 0.01938965730369091, + 0.019380690529942513, + 0.025296609848737717, + 0.028026994317770004, + 0.02870742604136467, + 0.026560377329587936, + 0.013572306372225285, + 0.01109025627374649, + 0.015460005961358547, + 0.006589054130017757, + 0.0094295060262084, + 0.02739504538476467, + 0.008541852235794067, + 0.006451325491070747, + 0.03491934388875961, + 0.022290078923106194, + 0.024946346879005432, + 0.01801385171711445, + 0.02926342561841011, + 0.020142020657658577, + 0.022989341989159584, + 0.008608066476881504, + 0.03295862674713135, + 0.021183567121624947, + 0.02842857874929905, + 0.012907134369015694, + 0.01943092793226242, + 0.02035784348845482, + 0.011753571219742298, + 0.008388661779463291, + 0.005535302683711052, + 0.010354761965572834, + 0.0027778921648859978, + 0.0030123512260615826, + 0.006873878184705973, + 0.0055025480687618256, + 0.003759963670745492, + 0.003380959387868643, + 0.01159075926989317, + 0.009279118850827217, + 0.010008941404521465, + 0.007915278896689415, + 0.008937194012105465, + 0.024656914174556732, + 0.009418707340955734, + 0.022585738450288773, + 0.012823869474232197, + 0.015356287360191345, + 0.008233312517404556, + 0.007596347015351057, + 0.013163424097001553, + 0.02154143899679184, + 0.01487745251506567, + 0.012232381850481033, + 0.003778933547437191, + 0.0038408918771892786, + 0.004027916584163904, + 0.0024302364327013493, + 0.003488432615995407, + 0.004708503372967243, + 0.003752992022782564, + 0.005659904796630144, + 0.004128856118768454, + 0.0047896686010062695, + 0.004925877321511507, + 0.0036765714175999165, + 0.006818174384534359, + 0.007806861307471991, + 0.005589914508163929, + 0.010248956270515919, + 0.0023543706629425287, + 0.0024059442803263664, + 0.0026126501616090536, + 0.0034914074931293726, + 0.002771146362647414, + 0.0032816827297210693, + 0.0043962132185697556, + 0.00463889678940177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.25205764174461365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.01253500021994114 + ], + "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": [ + 0.1433144360780716, + 0.13570831716060638, + 0.15472674369812012, + 0.06057923287153244, + 0.04203953966498375, + 0.06596076488494873, + 0.08245222270488739, + 0.056339215487241745, + 0.10193683952093124, + 0.15694278478622437 + ], + "credit_map": [ + 0.02318909578025341, + 0.02144099771976471, + 0.014366013929247856, + 0.0128706656396389, + 0.013508453033864498, + 0.01581820286810398, + 0.018492870032787323, + 0.02362811379134655, + 0.02165820822119713, + 0.014903387054800987, + 0.03323149308562279, + 0.009754975326359272, + 0.011223381385207176, + 0.02872064709663391, + 0.009269255213439465, + 0.006946953944861889, + 0.02959291636943817, + 0.019820602610707283, + 0.020186567679047585, + 0.015061679296195507, + 0.026200538501143456, + 0.01723339781165123, + 0.020285416394472122, + 0.006345607340335846, + 0.02036220207810402, + 0.0065790386870503426, + 0.015529201366007328, + 0.0038464320823550224, + 0.004604544956237078, + 0.004521939437836409, + 0.002477873582392931, + 0.002657999051734805, + 0.00618757214397192, + 0.00967821478843689, + 0.004538020119071007, + 0.00405274098739028, + 0.005220665596425533, + 0.004888285882771015, + 0.0034032377880066633, + 0.004070799332112074, + 0.010511337779462337, + 0.0046240463852882385, + 0.006516035180538893, + 0.005600085016340017, + 0.006026249844580889, + 0.0129080880433321, + 0.009580002166330814, + 0.010194914415478706, + 0.01198951993137598, + 0.016607951372861862, + 0.005367589183151722, + 0.005967042408883572, + 0.008874061517417431, + 0.015493092127144337, + 0.007523613981902599, + 0.010629341937601566, + 0.006114036776125431, + 0.00616932800039649, + 0.004882826469838619, + 0.004991683177649975, + 0.005267712287604809, + 0.007916158065199852, + 0.011353610083460808, + 0.00964385736733675, + 0.007947990670800209, + 0.0082905488088727, + 0.007450272794812918, + 0.0066759525798261166, + 0.009823095053434372, + 0.020323513075709343, + 0.009853137657046318, + 0.03157231584191322, + 0.012677723541855812, + 0.013962950557470322, + 0.013835110701620579, + 0.01372302882373333, + 0.011232671327888966, + 0.02990979515016079, + 0.014738271944224834, + 0.04686322435736656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.01253500021994114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.31344637274742126 + ], + "opposite_logprob_tokens": [ + -1.3134464025497437 + ], + "logprob_reward": -0.31344637274742126, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344637274742126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1936926692724228, + 0.11796092242002487, + 0.08264149725437164, + 0.05052550137042999, + 0.031054845079779625, + 0.03041074052453041, + 0.08197081089019775, + 0.19340774416923523, + 0.13908100128173828, + 0.07925426214933395 + ], + "credit_map": [ + 0.024625329300761223, + 0.025613611564040184, + 0.018589602783322334, + 0.016793591901659966, + 0.023518765345215797, + 0.02471742406487465, + 0.026323536410927773, + 0.03351081535220146, + 0.024853216484189034, + 0.01019609346985817, + 0.028352612629532814, + 0.006789987441152334, + 0.0070152305997908115, + 0.028405876830220222, + 0.006687162443995476, + 0.005660736933350563, + 0.01712734065949917, + 0.010530783794820309, + 0.015517338179051876, + 0.006537255365401506, + 0.010698630474507809, + 0.012225514277815819, + 0.0060997274704277515, + 0.0039049112237989902, + 0.014304461888968945, + 0.003964164759963751, + 0.01472040917724371, + 0.0038666538894176483, + 0.0042166500352323055, + 0.003728857496753335, + 0.002964004874229431, + 0.0027602987829595804, + 0.0047058165073394775, + 0.008726807311177254, + 0.0020467606373131275, + 0.0021805947180837393, + 0.004839329048991203, + 0.003940184600651264, + 0.00259740580804646, + 0.0020179469138383865, + 0.004351595416665077, + 0.0024303914979100227, + 0.0037146597169339657, + 0.0027930522337555885, + 0.003231959417462349, + 0.004932906478643417, + 0.004387905355542898, + 0.0045682708732783794, + 0.010855318047106266, + 0.014936406165361404, + 0.005921213421970606, + 0.006266891956329346, + 0.008731857873499393, + 0.01634482480585575, + 0.008686581626534462, + 0.010227723978459835, + 0.02649787999689579, + 0.031358372420072556, + 0.01669515296816826, + 0.014339681714773178, + 0.017872152850031853, + 0.026410868391394615, + 0.02756669372320175, + 0.032666947692632675, + 0.010430562309920788, + 0.010461809113621712, + 0.009452822618186474, + 0.008528269827365875, + 0.012952975928783417, + 0.028569692745804787, + 0.013784494251012802, + 0.04490037262439728, + 0.009786651469767094, + 0.0076323458924889565, + 0.005816071759909391, + 0.005476654507219791, + 0.006067367270588875, + 0.014787477441132069, + 0.006807081401348114, + 0.022880610078573227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.31344637274742126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, one-shot, long take begins with a slowly ascending overhead angle, then shifts to a steady, circling motion around the subject. In the frame, the Monkey King, clad in an intricately carved suit of golden chain mail, intricately engraved with dragon and auspicious cloud motifs, hovers majestically above a swirling sea of clouds. He wears a phoenix-winged, purple-gold crown, two slender pheasant feathers trembling in the celestial winds. His scarlet cape flutters majestically behind him like a battle banner. His face is defiant, his fiery eyes gleaming with an unyielding radiance. His tightly pursed lips betray supreme majesty and confidence. In his powerful hands, he clutches the legendary Ruyi Jingu Bang. The black iron shaft reflects a deep metallic texture in the light, while the golden hoops at each end reflect a dazzling brilliance. As he slowly and leisurely circles, the camera moves in sync, revealing the boundless, layered clouds beneath his feet and the magnificent celestial palace faintly visible deep within them—its jade-like buildings, soaring eaves, and angled corners looming through the mist, appearing incredibly majestic and distant. Most strikingly, the thousand-foot-long golden rays of light, like divine swords, pierce the thick clouds, forming distinct, Tyndall-effect beams imbued with dust and spiritual particles, illuminating the Monkey King's golden armor with a gleaming brilliance. The interplay of light and shadow imbues the entire scene with an epic atmosphere of sacredness, aloofness, and immense power.", + "seed": 32496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32496_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": -0.06272356049157679, + "vlm_logprob": -0.06272356049157679, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.06272356049157679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.014420812949538231 + ], + "opposite_logprob_tokens": [ + -4.639420986175537 + ], + "logprob_reward": -0.014420812949538231, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.014420812949538231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057782214134931564, + 0.11730210483074188, + 0.029737401753664017, + 0.2623518705368042, + 0.037346985191106796, + 0.026998139917850494, + 0.13426853716373444, + 0.2465345412492752, + 0.03497990220785141, + 0.05269832909107208 + ], + "credit_map": [ + 0.007591386791318655, + 0.006608237512409687, + 0.006429999601095915, + 0.005209323484450579, + 0.007970614358782768, + 0.008481253869831562, + 0.006362840533256531, + 0.009128556586802006, + 0.023190341889858246, + 0.04124283045530319, + 0.009044711478054523, + 0.005110014695674181, + 0.009309115819633007, + 0.012258008122444153, + 0.005777311511337757, + 0.011369762942194939, + 0.005243096500635147, + 0.0026487892027944326, + 0.0035228002816438675, + 0.0041596717201173306, + 0.0025087986141443253, + 0.004123627208173275, + 0.003024570643901825, + 0.004506047815084457, + 0.03136920556426048, + 0.01312440074980259, + 0.06563644856214523, + 0.016236789524555206, + 0.013532960787415504, + 0.04116377979516983, + 0.01644999347627163, + 0.06483832001686096, + 0.004223683383315802, + 0.0032544774003326893, + 0.006419104058295488, + 0.002891949377954006, + 0.003060547634959221, + 0.005231473129242659, + 0.004512138199061155, + 0.007753612473607063, + 0.0024848829489201307, + 0.0026615483220666647, + 0.003319229232147336, + 0.0029776289593428373, + 0.005294979549944401, + 0.0035739694721996784, + 0.002906257752329111, + 0.0037796441465616226, + 0.03443393111228943, + 0.006174338050186634, + 0.008999318815767765, + 0.008343885652720928, + 0.008647887036204338, + 0.042910385876894, + 0.009706792421638966, + 0.01505199819803238, + 0.042470287531614304, + 0.034035373479127884, + 0.03103196620941162, + 0.02498672530055046, + 0.027769260108470917, + 0.03200426325201988, + 0.024456581100821495, + 0.029780084267258644, + 0.004786140751093626, + 0.003713109064847231, + 0.003965490031987429, + 0.00480402447283268, + 0.005154274869710207, + 0.005465068854391575, + 0.0037731369957327843, + 0.0033186557702720165, + 0.008487796410918236, + 0.007395761087536812, + 0.004852159880101681, + 0.00476471334695816, + 0.009009482339024544, + 0.006798418238759041, + 0.006618089508265257, + 0.0047719138674438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.014420812949538231 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.0032024311367422342 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.0032024311367422342, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032024311367422342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061348412185907364, + 0.051983751356601715, + 0.06097216159105301, + 0.5145115852355957, + 0.05544133484363556, + 0.0215185284614563, + 0.07843008637428284, + 0.1096230000257492, + 0.029927928000688553, + 0.016243143007159233 + ], + "credit_map": [ + 0.007290516514331102, + 0.007633344270288944, + 0.008314011618494987, + 0.007714997977018356, + 0.007955100387334824, + 0.0066554248332977295, + 0.008791576139628887, + 0.006993442308157682, + 0.005497063044458628, + 0.012090533971786499, + 0.007311587687581778, + 0.004222329705953598, + 0.003983066417276859, + 0.006539963651448488, + 0.005014292895793915, + 0.007324918173253536, + 0.007579646538943052, + 0.0058760750107467175, + 0.01264949794858694, + 0.006891871336847544, + 0.00401137163862586, + 0.006938302423804998, + 0.006093801464885473, + 0.010931599885225296, + 0.06460274755954742, + 0.021523138508200645, + 0.1329590529203415, + 0.030512157827615738, + 0.020794691517949104, + 0.08251785486936569, + 0.02835124172270298, + 0.13325075805187225, + 0.0063758594915270805, + 0.0039423685520887375, + 0.010304775089025497, + 0.0038962617982178926, + 0.005250629503279924, + 0.008414865471422672, + 0.005010997410863638, + 0.012245584279298782, + 0.0020183883607387543, + 0.0022284258157014847, + 0.0024659521877765656, + 0.0022742662113159895, + 0.003898837137967348, + 0.0029109136667102575, + 0.002563627203926444, + 0.0031581188086420298, + 0.020431745797395706, + 0.00387589237652719, + 0.005459831561893225, + 0.0052587431855499744, + 0.00523510854691267, + 0.025556854903697968, + 0.004107086919248104, + 0.008504834957420826, + 0.01900002546608448, + 0.01455952599644661, + 0.014109686948359013, + 0.011289093643426895, + 0.011919758282601833, + 0.014678760431706905, + 0.010696685872972012, + 0.013369472697377205, + 0.004640020430088043, + 0.005035384092479944, + 0.0037432117387652397, + 0.003685158211737871, + 0.003546393010765314, + 0.0036099546123296022, + 0.003201213898137212, + 0.002466592937707901, + 0.002447795821353793, + 0.0025934912264347076, + 0.0013653305359184742, + 0.0011956720845773816, + 0.00183325435500592, + 0.003884667996317148, + 0.0015161808114498854, + 0.0014067500596866012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.0032024311367422342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.1603224128484726 + ], + "opposite_logprob_tokens": [ + -1.9103224277496338 + ], + "logprob_reward": -0.1603224128484726, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603224128484726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07511623948812485, + 0.16185642778873444, + 0.03849928081035614, + 0.4398074448108673, + 0.02680795267224312, + 0.03949959576129913, + 0.05667068064212799, + 0.04301263764500618, + 0.06256415694952011, + 0.05616563558578491 + ], + "credit_map": [ + 0.009387023746967316, + 0.009135628119111061, + 0.006391837727278471, + 0.006908737123012543, + 0.011349310167133808, + 0.009779403917491436, + 0.01062061544507742, + 0.011543679982423782, + 0.05068980157375336, + 0.058957625180482864, + 0.01137675903737545, + 0.007629748899489641, + 0.00659777969121933, + 0.009415053762495518, + 0.006219778209924698, + 0.010969866067171097, + 0.005884483456611633, + 0.003517962759360671, + 0.005438348278403282, + 0.005070358514785767, + 0.0029027422424405813, + 0.004493979271501303, + 0.002950429916381836, + 0.008240973576903343, + 0.05205647647380829, + 0.02066485770046711, + 0.11037090420722961, + 0.02521580457687378, + 0.0188696701079607, + 0.07045358419418335, + 0.02525380440056324, + 0.11692232638597488, + 0.0027225769590586424, + 0.003020948264747858, + 0.004490815568715334, + 0.0022596868220716715, + 0.00247437646612525, + 0.0037749907933175564, + 0.0030325527768582106, + 0.005032004788517952, + 0.0031208470463752747, + 0.0045433505438268185, + 0.00554864015430212, + 0.003102027578279376, + 0.005709425546228886, + 0.005636557471007109, + 0.0049021439626812935, + 0.006936599500477314, + 0.011909127235412598, + 0.00365302711725235, + 0.005168350413441658, + 0.004330629948526621, + 0.004506343510001898, + 0.01864047534763813, + 0.002954348921775818, + 0.005508375819772482, + 0.009077740833163261, + 0.009341312572360039, + 0.003893817774951458, + 0.003298674477264285, + 0.003891219850629568, + 0.006573632825165987, + 0.002970528556033969, + 0.003965707495808601, + 0.008680400438606739, + 0.009788664057850838, + 0.0050573404878377914, + 0.007754610851407051, + 0.006589427590370178, + 0.009619790129363537, + 0.007610237691551447, + 0.007463675923645496, + 0.008228947408497334, + 0.008952203206717968, + 0.004719431046396494, + 0.0033321522641927004, + 0.008267803117632866, + 0.010622123256325722, + 0.007958844304084778, + 0.004084128420799971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.1603224128484726 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.00863721314817667 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.00863721314817667, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863721314817667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11918758600950241, + 0.21891196072101593, + 0.08275029808282852, + 0.316313236951828, + 0.0532822348177433, + 0.04800231009721756, + 0.05743693187832832, + 0.02569042704999447, + 0.04305890202522278, + 0.03536606580018997 + ], + "credit_map": [ + 0.02021644450724125, + 0.016298742964863777, + 0.007700188085436821, + 0.008775102905929089, + 0.022155484184622765, + 0.01317116804420948, + 0.011688672006130219, + 0.01918179728090763, + 0.05566483736038208, + 0.08026042580604553, + 0.01759275235235691, + 0.009471388533711433, + 0.010125729255378246, + 0.011300127021968365, + 0.010851778090000153, + 0.023644939064979553, + 0.011352524161338806, + 0.008672727271914482, + 0.013821073807775974, + 0.010804813355207443, + 0.006793390028178692, + 0.010233971290290356, + 0.010310135781764984, + 0.010761667974293232, + 0.039901960641145706, + 0.013565467670559883, + 0.07928662747144699, + 0.019946230575442314, + 0.013267546892166138, + 0.04869851842522621, + 0.01863020658493042, + 0.08301668614149094, + 0.005624685902148485, + 0.002732452703639865, + 0.011041372083127499, + 0.005000523291528225, + 0.003243784885853529, + 0.007811000570654869, + 0.004369143862277269, + 0.013459279201924801, + 0.004386288113892078, + 0.003614452201873064, + 0.008953708223998547, + 0.004240557551383972, + 0.005670582409948111, + 0.008058049716055393, + 0.004365963861346245, + 0.008712711744010448, + 0.008347618393599987, + 0.004675903357565403, + 0.00798218883574009, + 0.007228505332022905, + 0.005924795754253864, + 0.011159748770296574, + 0.004093557130545378, + 0.00802462175488472, + 0.0041741481982171535, + 0.004725519102066755, + 0.003234356641769409, + 0.0021090591326355934, + 0.00257478398270905, + 0.003897016867995262, + 0.0025533554144203663, + 0.0024221891071647406, + 0.0077223326079547405, + 0.009952940978109837, + 0.00366420391947031, + 0.003694291226565838, + 0.005002318881452084, + 0.007234393619000912, + 0.002914926502853632, + 0.0028734994120895863, + 0.005213383585214615, + 0.006952086929231882, + 0.0033218455500900745, + 0.0028071755077689886, + 0.006310965400189161, + 0.003810172900557518, + 0.003929697442799807, + 0.003020743140950799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.00863721314817667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.12703493237495422 + ], + "opposite_logprob_tokens": [ + -2.127034902572632 + ], + "logprob_reward": -0.12703493237495422, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703493237495422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06181694567203522, + 0.1145973801612854, + 0.08361438661813736, + 0.5127366185188293, + 0.037386272102594376, + 0.02828947827219963, + 0.03764658421278, + 0.03040407784283161, + 0.07364721596240997, + 0.019861074164509773 + ], + "credit_map": [ + 0.009630536660552025, + 0.007863110862672329, + 0.005304929334670305, + 0.004548290278762579, + 0.008540719747543335, + 0.011693206615746021, + 0.005757184233516455, + 0.008478973060846329, + 0.023451175540685654, + 0.04153914004564285, + 0.014698995277285576, + 0.005566135048866272, + 0.005979122128337622, + 0.009046603925526142, + 0.0059919534251093864, + 0.008324244059622288, + 0.013918365351855755, + 0.00651730177924037, + 0.009493052028119564, + 0.012637896463274956, + 0.005702597554773092, + 0.00990541186183691, + 0.010900166817009449, + 0.01453959196805954, + 0.06420167535543442, + 0.021665019914507866, + 0.12921378016471863, + 0.031595807522535324, + 0.02325083315372467, + 0.08095026016235352, + 0.028715165331959724, + 0.1331440657377243, + 0.00531949894502759, + 0.0029040935914963484, + 0.006903721019625664, + 0.003278258955106139, + 0.002808134537190199, + 0.005207777488976717, + 0.003272318746894598, + 0.00769247068092227, + 0.002871257020160556, + 0.002603638218715787, + 0.004616425838321447, + 0.0026044452097266912, + 0.0030129707884043455, + 0.003948109690099955, + 0.003529723733663559, + 0.005102910101413727, + 0.006554095074534416, + 0.002936477307230234, + 0.004862920846790075, + 0.00415138341486454, + 0.0033529221545904875, + 0.00871611014008522, + 0.0025229942984879017, + 0.00454968074336648, + 0.004566500429064035, + 0.006424982100725174, + 0.0037305171135812998, + 0.002659736666828394, + 0.002881143242120743, + 0.004090751521289349, + 0.003056057495996356, + 0.00299438857473433, + 0.011985031887888908, + 0.01408299058675766, + 0.008546342141926289, + 0.007587216328829527, + 0.008901990950107574, + 0.011058822274208069, + 0.005291198380291462, + 0.006193624809384346, + 0.002599007450044155, + 0.0026668207719922066, + 0.0023940035607665777, + 0.0019344985485076904, + 0.0027680075727403164, + 0.0027680550701916218, + 0.002749254461377859, + 0.0019814276602119207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.12703493237495422 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, one-shot, long take begins with a slowly ascending overhead angle, then shifts to a steady, circling motion around the subject. In the frame, the Monkey King, clad in an intricately carved suit of golden chain mail, intricately engraved with dragon and auspicious cloud motifs, hovers majestically above a swirling sea of clouds. He wears a phoenix-winged, purple-gold crown, two slender pheasant feathers trembling in the celestial winds. His scarlet cape flutters majestically behind him like a battle banner. His face is defiant, his fiery eyes gleaming with an unyielding radiance. His tightly pursed lips betray supreme majesty and confidence. In his powerful hands, he clutches the legendary Ruyi Jingu Bang. The black iron shaft reflects a deep metallic texture in the light, while the golden hoops at each end reflect a dazzling brilliance. As he slowly and leisurely circles, the camera moves in sync, revealing the boundless, layered clouds beneath his feet and the magnificent celestial palace faintly visible deep within them—its jade-like buildings, soaring eaves, and angled corners looming through the mist, appearing incredibly majestic and distant. Most strikingly, the thousand-foot-long golden rays of light, like divine swords, pierce the thick clouds, forming distinct, Tyndall-effect beams imbued with dust and spiritual particles, illuminating the Monkey King's golden armor with a gleaming brilliance. The interplay of light and shadow imbues the entire scene with an epic atmosphere of sacredness, aloofness, and immense power.", + "seed": 32497, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_13/index_7742_seed_32497_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": -0.09152067261748016, + "vlm_logprob": -0.09152067261748016, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.09152067261748016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.010835618712008 + ], + "opposite_logprob_tokens": [ + -4.885835647583008 + ], + "logprob_reward": -0.010835618712008, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010835618712008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2835357189178467, + 0.052598267793655396, + 0.09627626836299896, + 0.3755500912666321, + 0.023734277114272118, + 0.033709924668073654, + 0.03096975013613701, + 0.03956970199942589, + 0.03346802666783333, + 0.030587954446673393 + ], + "credit_map": [ + 0.05812494084239006, + 0.02126222476363182, + 0.034586209803819656, + 0.030270474031567574, + 0.030628161504864693, + 0.03042931854724884, + 0.036416251212358475, + 0.041818153113126755, + 0.006798035465180874, + 0.005386570002883673, + 0.01044817641377449, + 0.0071839396841824055, + 0.006395296193659306, + 0.005459663923829794, + 0.006186656653881073, + 0.0047399322502315044, + 0.01532297395169735, + 0.016136616468429565, + 0.012773559428751469, + 0.012984841130673885, + 0.01258397102355957, + 0.010907167568802834, + 0.0082402303814888, + 0.0073269023559987545, + 0.014004229567945004, + 0.012693828903138638, + 0.01379039604216814, + 0.015283416025340557, + 0.04427720606327057, + 0.15213803946971893, + 0.009839101694524288, + 0.11352388560771942, + 0.0028839167207479477, + 0.002213152823969722, + 0.001854490372352302, + 0.0023424960672855377, + 0.0022622470278292894, + 0.006409152876585722, + 0.0022057960741221905, + 0.00356302410364151, + 0.00570607278496027, + 0.0027982371393591166, + 0.0076341875828802586, + 0.004822320770472288, + 0.0027189827524125576, + 0.0038575094658881426, + 0.002061396837234497, + 0.0041112145408988, + 0.003723711473867297, + 0.0043386854231357574, + 0.004571061581373215, + 0.0027858391404151917, + 0.002930008340626955, + 0.00430802209302783, + 0.0027388304006308317, + 0.005573589820414782, + 0.0058784629218280315, + 0.004397374577820301, + 0.003654773812741041, + 0.0038636564277112484, + 0.004175093956291676, + 0.005743802059441805, + 0.0037206050474196672, + 0.008135932497680187, + 0.004700078628957272, + 0.004002443514764309, + 0.004450123757123947, + 0.0030249457340687513, + 0.00492836395278573, + 0.005580219905823469, + 0.0025561796501278877, + 0.004225674085319042, + 0.0032721511088311672, + 0.00381851801648736, + 0.004326220601797104, + 0.0038193613290786743, + 0.0034644771367311478, + 0.004870730917900801, + 0.0035360080655664206, + 0.0034804861061275005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.010835618712008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20694205164909363, + 0.06279398500919342, + 0.11160879582166672, + 0.30830520391464233, + 0.05270862579345703, + 0.026307838037610054, + 0.05574384704232216, + 0.05940938740968704, + 0.05685724318027496, + 0.05932295694947243 + ], + "credit_map": [ + 0.058599162846803665, + 0.01833203248679638, + 0.021408947184681892, + 0.0148777412250638, + 0.017135068774223328, + 0.024208925664424896, + 0.018480809405446053, + 0.033899370580911636, + 0.010372643359005451, + 0.008564128540456295, + 0.011935760267078876, + 0.006833519786596298, + 0.006025385111570358, + 0.006604797672480345, + 0.005737748462706804, + 0.006720004603266716, + 0.01927960477769375, + 0.018336089327931404, + 0.014927913434803486, + 0.015309343114495277, + 0.012644490227103233, + 0.012100286781787872, + 0.009486904367804527, + 0.00952416006475687, + 0.010904989205300808, + 0.010604518465697765, + 0.006606000009924173, + 0.011249516159296036, + 0.037618037313222885, + 0.12903182208538055, + 0.00675781536847353, + 0.09553252905607224, + 0.007073018234223127, + 0.00472239637747407, + 0.0026601373683661222, + 0.0043218242935836315, + 0.005916604306548834, + 0.01444978453218937, + 0.0048409816808998585, + 0.008723878301680088, + 0.004414097405970097, + 0.002632021438330412, + 0.002985861850902438, + 0.004712433088570833, + 0.0030779193621128798, + 0.003181842854246497, + 0.0021701299119740725, + 0.003133530728518963, + 0.010262329131364822, + 0.007768028415739536, + 0.006580110173672438, + 0.004056931007653475, + 0.005419258493930101, + 0.00878428015857935, + 0.0037805584724992514, + 0.00909234955906868, + 0.007248343899846077, + 0.008138217031955719, + 0.004499821458011866, + 0.0045427000150084496, + 0.007702333386987448, + 0.009419549256563187, + 0.0060065858997404575, + 0.011851836927235126, + 0.0075453659519553185, + 0.007042133715003729, + 0.007466858718544245, + 0.005022954195737839, + 0.006876394618302584, + 0.00869788508862257, + 0.00474990950897336, + 0.009455744177103043, + 0.011006142012774944, + 0.007751015014946461, + 0.004488629288971424, + 0.0034450071398168802, + 0.009971778839826584, + 0.009060616604983807, + 0.00803419016301632, + 0.00556558184325695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.005958176217973232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.3485891819000244 + ], + "opposite_logprob_tokens": [ + -1.2235891819000244 + ], + "logprob_reward": -0.3485891819000244, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3485891819000244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23638814687728882, + 0.037798598408699036, + 0.07372415065765381, + 0.2761878967285156, + 0.13514655828475952, + 0.043025970458984375, + 0.03833778575062752, + 0.06542646139860153, + 0.04641354829072952, + 0.047550857067108154 + ], + "credit_map": [ + 0.04782039672136307, + 0.018606165423989296, + 0.03076780214905739, + 0.02326294779777527, + 0.027933351695537567, + 0.027004139497876167, + 0.025401415303349495, + 0.03559194132685661, + 0.005302335135638714, + 0.004451468121260405, + 0.006484131794422865, + 0.004642345476895571, + 0.003774421289563179, + 0.0040323915891349316, + 0.004807214718312025, + 0.004304292146116495, + 0.012822815217077732, + 0.012536713853478432, + 0.009341630153357983, + 0.010439107194542885, + 0.008234060369431973, + 0.00808411743491888, + 0.0062413946725428104, + 0.006024310365319252, + 0.01017149630934, + 0.009548772126436234, + 0.006557203363627195, + 0.010455360636115074, + 0.03406205028295517, + 0.11400523036718369, + 0.007569127716124058, + 0.08381865173578262, + 0.022546133026480675, + 0.010982731357216835, + 0.012692310847342014, + 0.013590042479336262, + 0.012681802734732628, + 0.0329328253865242, + 0.010525965131819248, + 0.019194750115275383, + 0.007719907443970442, + 0.0051790690049529076, + 0.006910675670951605, + 0.006667861249297857, + 0.004740586504340172, + 0.004874073900282383, + 0.0027511161752045155, + 0.004182680509984493, + 0.00558442622423172, + 0.00631316751241684, + 0.0036319312639534473, + 0.003550108755007386, + 0.005149138625711203, + 0.005494122393429279, + 0.0037644109688699245, + 0.004850481636822224, + 0.008671767078340054, + 0.011780736967921257, + 0.005089377518743277, + 0.0053158472292125225, + 0.008342516608536243, + 0.008820384740829468, + 0.0071117510087788105, + 0.01029407698661089, + 0.006581986788660288, + 0.00569827388972044, + 0.007027218118309975, + 0.0035398919135332108, + 0.005763113033026457, + 0.007574988529086113, + 0.003463596338406205, + 0.006764479447156191, + 0.006007516290992498, + 0.00693053612485528, + 0.005082306917756796, + 0.004194468259811401, + 0.005672796629369259, + 0.009775360114872456, + 0.004599260166287422, + 0.005288613028824329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.3485891819000244 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09603900462388992, + 0.05471320450305939, + 0.08962184190750122, + 0.26588279008865356, + 0.08807473629713058, + 0.025340748950839043, + 0.03226640447974205, + 0.061052195727825165, + 0.1604829579591751, + 0.1265260875225067 + ], + "credit_map": [ + 0.01567484252154827, + 0.008002767339348793, + 0.012819081544876099, + 0.010993778705596924, + 0.01206705067306757, + 0.01048994529992342, + 0.01219191960990429, + 0.01379961147904396, + 0.009636376984417439, + 0.005647027865052223, + 0.008963325060904026, + 0.006167978513985872, + 0.007191999815404415, + 0.005989172495901585, + 0.005659544840455055, + 0.005457780323922634, + 0.014601009897887707, + 0.012286740355193615, + 0.011929042637348175, + 0.01362619549036026, + 0.010271228849887848, + 0.01063591055572033, + 0.009498526342213154, + 0.006773189641535282, + 0.008618329651653767, + 0.009080218151211739, + 0.006030216347426176, + 0.010105099529027939, + 0.03192233294248581, + 0.11108435690402985, + 0.006147337146103382, + 0.08289488404989243, + 0.01971251703798771, + 0.0076399692334234715, + 0.00900440476834774, + 0.0104683181270957, + 0.008195862174034119, + 0.01465091947466135, + 0.007897467352449894, + 0.010505280457437038, + 0.00359689025208354, + 0.0027959132567048073, + 0.003408168675377965, + 0.0028202892281115055, + 0.0033707045949995518, + 0.00379934418015182, + 0.0019684615544974804, + 0.0035809739492833614, + 0.004551379010081291, + 0.005567534361034632, + 0.002120658289641142, + 0.0032099720556288958, + 0.0039001060649752617, + 0.004614265635609627, + 0.0027400003746151924, + 0.005562491714954376, + 0.00831116084009409, + 0.008190622553229332, + 0.007211313582956791, + 0.007350125815719366, + 0.009508521296083927, + 0.007633454166352749, + 0.005753559060394764, + 0.007093437481671572, + 0.024697640910744667, + 0.016293160617351532, + 0.016397830098867416, + 0.010652929544448853, + 0.018602615222334862, + 0.03303666412830353, + 0.013921201229095459, + 0.026880914345383644, + 0.0154651440680027, + 0.014393373392522335, + 0.008684512227773666, + 0.008987992070615292, + 0.014221257530152798, + 0.029867151752114296, + 0.018152590841054916, + 0.01675407588481903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.0031914988067001104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the Monkey King wearing a golden chain mail suit with dragon and cloud motifs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the Monkey King hovering above a swirling sea of clouds while holding the Ruyi Jingu Bang?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a slowly ascending overhead angle and then shift to a steady circling motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are golden rays of light visible piercing the clouds to form distinct beams with dust and spiritual particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a celestial palace with jade-like buildings and soaring eaves faintly visible 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": [ + -0.08902888745069504 + ], + "opposite_logprob_tokens": [ + -2.464028835296631 + ], + "logprob_reward": -0.08902888745069504, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902888745069504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14689937233924866, + 0.02837183326482773, + 0.06789809465408325, + 0.30198734998703003, + 0.06666216254234314, + 0.043284088373184204, + 0.07340360432863235, + 0.20642636716365814, + 0.030633600428700447, + 0.034433573484420776 + ], + "credit_map": [ + 0.03107396513223648, + 0.011166226118803024, + 0.017609577625989914, + 0.014955907128751278, + 0.01567726954817772, + 0.016619926318526268, + 0.017430627718567848, + 0.022365881130099297, + 0.003335730405524373, + 0.0029585841111838818, + 0.0049526505172252655, + 0.0037016787100583315, + 0.003296814626082778, + 0.002862948691472411, + 0.0043330262415111065, + 0.0029304020572453737, + 0.010942406952381134, + 0.010010874830186367, + 0.010354622267186642, + 0.00989886187016964, + 0.007100839167833328, + 0.007715669460594654, + 0.006266400683671236, + 0.005608424078673124, + 0.010412360541522503, + 0.011140471324324608, + 0.00839726161211729, + 0.01196887344121933, + 0.035407789051532745, + 0.12402796745300293, + 0.006899867672473192, + 0.09373272955417633, + 0.012961506843566895, + 0.007032008841633797, + 0.007909921929240227, + 0.0076499939896166325, + 0.008294598199427128, + 0.00930720567703247, + 0.005587904714047909, + 0.007919027470052242, + 0.0075891208834946156, + 0.00499767204746604, + 0.005852873437106609, + 0.004746613092720509, + 0.005171770695596933, + 0.005645053461194038, + 0.003986066672950983, + 0.0052949171513319016, + 0.014786464162170887, + 0.013416585512459278, + 0.007289591245353222, + 0.0056788125075399876, + 0.008898243308067322, + 0.011160495691001415, + 0.004308497533202171, + 0.007864915765821934, + 0.03140106424689293, + 0.02167464606463909, + 0.013167502358555794, + 0.014550908468663692, + 0.031153125688433647, + 0.03509685397148132, + 0.020146004855632782, + 0.03923623636364937, + 0.0043159350752830505, + 0.0037380445282906294, + 0.003518054960295558, + 0.002214365638792515, + 0.004999673925340176, + 0.004315172787755728, + 0.002899889135733247, + 0.004632463678717613, + 0.0041050296276807785, + 0.003909635357558727, + 0.0036162803880870342, + 0.003812916576862335, + 0.004438275936990976, + 0.007257882039994001, + 0.0035169266629964113, + 0.003776626195758581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.08902888745069504 + } + ], + "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/0000022/group_14/index_tensor([7417, 7417])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_tensor([7417, 7417])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..16747923a16bd3c81b15891e1df96b8ffabaee76 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_tensor([7417, 7417])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot, starting from a low angle and then gently tilting upward, captured a vast vortex of incandescent gas, swirling violently in the depths of space, illuminated by the hyper-realistic lighting and shadows rendered by Octane. These nebulous gases entwine and collide in complex hydrodynamic paths, forming a dynamic center of energy. Suddenly, a blue dragon, its scales gleaming with cosmic light, its majestic head bursts through the vortex's core, its golden eyes blazing with starlight scanning the void. Then, its slender, sinuous body rises from the vortex in a graceful and powerful gesture. Each deep emerald scale seems to envelop a miniature galaxy, aglow with a constantly shifting glow of deep purple, royal blue, and gold stardust. When it fully displays its sacred and majestic appearance, the surrounding nebula gas is stirred by its powerful energy, forming countless tiny, glowing energy threads and particle streams that entwine around its body, while the cosmic light on its scales interacts with the blazing light of the nebula, casting complex and ever-changing light and shadows on its smooth body. The whole picture presents a sense of grandeur and sacredness of the epic creation of the universe.", + "seed": 16577, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_7417_seed_16577_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": -0.11627187454141677, + "vlm_logprob": -0.11627187454141677, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.11627187454141677, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16563855111598969 + ], + "opposite_logprob_tokens": [ + -1.9156385660171509 + ], + "logprob_reward": -0.16563855111598969, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16563855111598969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42072731256484985, + 0.08191076666116714, + 0.0704251304268837, + 0.06488573551177979, + 0.16662156581878662, + 0.08082294464111328, + 0.0367886982858181, + 0.012616308405995369, + 0.02502480521798134, + 0.04017670452594757 + ], + "credit_map": [ + 0.038902610540390015, + 0.076435886323452, + 0.05472157150506973, + 0.05083607882261276, + 0.048961084336042404, + 0.03776853159070015, + 0.08179985731840134, + 0.03130169212818146, + 0.008146414533257484, + 0.01036885753273964, + 0.0072794570587575436, + 0.01115515548735857, + 0.012380912899971008, + 0.009664743207395077, + 0.013193288818001747, + 0.009721933864057064, + 0.00401433277875185, + 0.012233193032443523, + 0.00929106306284666, + 0.008117791265249252, + 0.012345448136329651, + 0.008265222422778606, + 0.00785007607191801, + 0.008307999931275845, + 0.004505080170929432, + 0.012047132477164268, + 0.007572939619421959, + 0.007804852910339832, + 0.011445722542703152, + 0.007005561608821154, + 0.006857092492282391, + 0.007647353690117598, + 0.009476453997194767, + 0.019674669951200485, + 0.05537699908018112, + 0.016405494883656502, + 0.015166597440838814, + 0.011323795653879642, + 0.013357909396290779, + 0.02583964169025421, + 0.006644940469413996, + 0.007367987651377916, + 0.009051278233528137, + 0.01884729042649269, + 0.007878405973315239, + 0.010048279538750648, + 0.005972293671220541, + 0.015012471005320549, + 0.002555310260504484, + 0.0025196801871061325, + 0.009974251501262188, + 0.0032122423872351646, + 0.0023727479856461287, + 0.010663568042218685, + 0.0027552584651857615, + 0.002735638292506337, + 0.00226305122487247, + 0.0014172241790220141, + 0.0011869688751176, + 0.0017481351969763637, + 0.0013878822792321444, + 0.0020031752064824104, + 0.0015663396334275603, + 0.0010435318108648062, + 0.004947083070874214, + 0.0026364149525761604, + 0.0023408783599734306, + 0.0026772262062877417, + 0.002443100791424513, + 0.0031135499011725187, + 0.0022147037088871002, + 0.004651847295463085, + 0.0046169147826731205, + 0.005847877357155085, + 0.004895451944321394, + 0.004489533603191376, + 0.005124825052917004, + 0.00386377633549273, + 0.006563019007444382, + 0.004775304812937975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.16563855111598969 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018217505887150764 + ], + "opposite_logprob_tokens": [ + -4.01821756362915 + ], + "logprob_reward": -0.018217505887150764, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018217505887150764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34122830629348755, + 0.05289435759186745, + 0.11692396551370621, + 0.11373808234930038, + 0.04580690711736679, + 0.048174794763326645, + 0.10980899631977081, + 0.028788477182388306, + 0.0691199004650116, + 0.07351618260145187 + ], + "credit_map": [ + 0.026475343853235245, + 0.04843815043568611, + 0.07787219434976578, + 0.047959618270397186, + 0.03452755883336067, + 0.029819227755069733, + 0.05139843374490738, + 0.02473776787519455, + 0.00430811382830143, + 0.007579403463751078, + 0.005168849136680365, + 0.005811112001538277, + 0.007659997325390577, + 0.010255439206957817, + 0.00695791793987155, + 0.0051535265520215034, + 0.007953754626214504, + 0.018774086609482765, + 0.016369763761758804, + 0.014770284295082092, + 0.018305007368326187, + 0.01144031435251236, + 0.015355243347585201, + 0.013955502770841122, + 0.007483015768229961, + 0.019591066986322403, + 0.011990115977823734, + 0.016546474769711494, + 0.016284339129924774, + 0.010006873868405819, + 0.013998741284012794, + 0.01783745549619198, + 0.003130301833152771, + 0.010011387057602406, + 0.005846351385116577, + 0.0071169789880514145, + 0.0051121520809829235, + 0.003931929357349873, + 0.006299152504652739, + 0.004358653444796801, + 0.004302434157580137, + 0.005513377021998167, + 0.006991065107285976, + 0.010197783820331097, + 0.005207685753703117, + 0.0046119969338178635, + 0.005730331875383854, + 0.005620123352855444, + 0.008795111440122128, + 0.009515121579170227, + 0.0214641522616148, + 0.013909999281167984, + 0.007464738097041845, + 0.025159835815429688, + 0.009865155443549156, + 0.013634881936013699, + 0.003183053806424141, + 0.0026341190095990896, + 0.0072668083012104034, + 0.0035444663371890783, + 0.002104841871187091, + 0.004085161257535219, + 0.001705951290205121, + 0.004264075309038162, + 0.012160524725914001, + 0.0069982996210455894, + 0.011080462485551834, + 0.006451437249779701, + 0.007168420124799013, + 0.007927057333290577, + 0.006806043907999992, + 0.010527655482292175, + 0.008638524450361729, + 0.014554237946867943, + 0.008062743581831455, + 0.006118889432400465, + 0.013960388489067554, + 0.006660289131104946, + 0.007955656386911869, + 0.0075654443353414536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.018217505887150764 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252132773399353 + ], + "opposite_logprob_tokens": [ + -1.502132773399353 + ], + "logprob_reward": -0.252132773399353, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252132773399353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4746769368648529, + 0.055333249270915985, + 0.03266553580760956, + 0.12370423227548599, + 0.07681919634342194, + 0.03221472352743149, + 0.04532115161418915, + 0.02028864622116089, + 0.03482853248715401, + 0.10414785891771317 + ], + "credit_map": [ + 0.046136680990457535, + 0.09890192002058029, + 0.05348985642194748, + 0.051707081496715546, + 0.05803059786558151, + 0.036690063774585724, + 0.09043072164058685, + 0.03928997367620468, + 0.0045717195607721806, + 0.008256852626800537, + 0.005282978992909193, + 0.00721192080527544, + 0.007295946590602398, + 0.009342355653643608, + 0.00727192172780633, + 0.0060995458625257015, + 0.003511459333822131, + 0.0055867466144263744, + 0.004074077121913433, + 0.004773872438818216, + 0.004136443603783846, + 0.0030100757721811533, + 0.004474090412259102, + 0.003098763059824705, + 0.007567910943180323, + 0.015291155315935612, + 0.015628201887011528, + 0.01331156212836504, + 0.03005121275782585, + 0.014459335245192051, + 0.016142480075359344, + 0.011252356693148613, + 0.005786473862826824, + 0.008270935155451298, + 0.02673965133726597, + 0.007964310236275196, + 0.006641267798841, + 0.006973638199269772, + 0.0050107804127037525, + 0.009432130493223667, + 0.002677026903256774, + 0.0040577007457613945, + 0.004839934874325991, + 0.007211452815681696, + 0.00295206718146801, + 0.0045107207261025906, + 0.002908133901655674, + 0.0030576814897358418, + 0.004095714073628187, + 0.0037540739867836237, + 0.006749541964381933, + 0.005419652443379164, + 0.0034190414007753134, + 0.009117108769714832, + 0.005166909191757441, + 0.0075990986078977585, + 0.00239306828007102, + 0.0026808984111994505, + 0.0035716828424483538, + 0.0029549342580139637, + 0.001716529717668891, + 0.0017041313694790006, + 0.0022203281987458467, + 0.0030470690689980984, + 0.006933723110705614, + 0.006636994890868664, + 0.003256340976804495, + 0.003895498812198639, + 0.003745446680113673, + 0.003802936291322112, + 0.003689546836540103, + 0.0028680411633104086, + 0.012107418850064278, + 0.02752532809972763, + 0.011636552400887012, + 0.007779106497764587, + 0.01689819060266018, + 0.009339755401015282, + 0.010909799486398697, + 0.007951698265969753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.252132773399353 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428815722465515 + ], + "opposite_logprob_tokens": [ + -2.0178816318511963 + ], + "logprob_reward": -0.1428815722465515, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428815722465515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24266327917575836, + 0.047658711671829224, + 0.09990581125020981, + 0.14453281462192535, + 0.0525067038834095, + 0.04288071021437645, + 0.21644888818264008, + 0.018541259691119194, + 0.055383190512657166, + 0.07947860658168793 + ], + "credit_map": [ + 0.025798074901103973, + 0.05308195576071739, + 0.033449120819568634, + 0.02760307304561138, + 0.028452111408114433, + 0.014908093959093094, + 0.04113307595252991, + 0.018237795680761337, + 0.0038018953055143356, + 0.007101931143552065, + 0.004780307877808809, + 0.00706084119156003, + 0.006141854915767908, + 0.009114987216889858, + 0.005720585118979216, + 0.003936311695724726, + 0.005986504722386599, + 0.015164998359978199, + 0.013394300825893879, + 0.01290816068649292, + 0.01585058867931366, + 0.010429018177092075, + 0.012976697646081448, + 0.01319554541260004, + 0.010021831840276718, + 0.028621137142181396, + 0.01565980166196823, + 0.020224468782544136, + 0.02043188363313675, + 0.01232454739511013, + 0.017678329721093178, + 0.01957082748413086, + 0.004140657372772694, + 0.003983154892921448, + 0.013044682331383228, + 0.0040077888406813145, + 0.004500097129493952, + 0.00586638692766428, + 0.0035883730743080378, + 0.01337556540966034, + 0.00413529435172677, + 0.0057267495431005955, + 0.005610012914985418, + 0.010626029223203659, + 0.0032204787712544203, + 0.0055617839097976685, + 0.0034971218556165695, + 0.00450324360281229, + 0.01722961850464344, + 0.01881970651447773, + 0.041993897408246994, + 0.025617308914661407, + 0.013763023540377617, + 0.05022444576025009, + 0.022275175899267197, + 0.0265257116407156, + 0.002190864644944668, + 0.0018783686682581902, + 0.0020376904867589474, + 0.0019166705897077918, + 0.002155468799173832, + 0.002330966992303729, + 0.003679610090330243, + 0.002351620467379689, + 0.013617899268865585, + 0.008256121538579464, + 0.007264560554176569, + 0.0038843548391014338, + 0.005113772116601467, + 0.006477581802755594, + 0.0039133308455348015, + 0.006855576299130917, + 0.00983826071023941, + 0.022730926051735878, + 0.009519129060208797, + 0.005862850230187178, + 0.010596276260912418, + 0.006211908999830484, + 0.007004215382039547, + 0.007715045008808374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.1428815722465515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3227611780166626, + 0.056443966925144196, + 0.06272347271442413, + 0.06530947983264923, + 0.07331940531730652, + 0.031437184661626816, + 0.15011271834373474, + 0.04041920602321625, + 0.06747330725193024, + 0.13000011444091797 + ], + "credit_map": [ + 0.028162728995084763, + 0.062015313655138016, + 0.04442450776696205, + 0.039519086480140686, + 0.03728379309177399, + 0.025959832593798637, + 0.05978012457489967, + 0.025615787133574486, + 0.005370997358113527, + 0.009311416186392307, + 0.004794997163116932, + 0.007562031503766775, + 0.008200832642614841, + 0.008064191788434982, + 0.006608074996620417, + 0.006531425751745701, + 0.004434752278029919, + 0.009411930106580257, + 0.009158575907349586, + 0.007855960167944431, + 0.009067895822227001, + 0.006988477427512407, + 0.008157312870025635, + 0.0076485625468194485, + 0.00522661954164505, + 0.008762197569012642, + 0.007752889767289162, + 0.006736064795404673, + 0.013502806425094604, + 0.00639937212690711, + 0.007421397138386965, + 0.00950813852250576, + 0.004702514037489891, + 0.00804566964507103, + 0.024320393800735474, + 0.007000400219112635, + 0.0067823706194758415, + 0.0061500887386500835, + 0.00499690230935812, + 0.01132106315344572, + 0.0034424567129462957, + 0.0043647438287734985, + 0.003790911752730608, + 0.006555644795298576, + 0.003478097263723612, + 0.0032720468007028103, + 0.003159035462886095, + 0.0033742478117346764, + 0.011619873344898224, + 0.012568501755595207, + 0.029830805957317352, + 0.01755804754793644, + 0.010458214208483696, + 0.03529853746294975, + 0.012446138076484203, + 0.020332608371973038, + 0.005500605795532465, + 0.005490552168339491, + 0.005094762425869703, + 0.0060807508416473866, + 0.0037100936751812696, + 0.0034497997257858515, + 0.0045578680001199245, + 0.006534775253385305, + 0.017566757276654243, + 0.00646592304110527, + 0.008960649371147156, + 0.0052258968353271484, + 0.006144201382994652, + 0.008335300721228123, + 0.007260136306285858, + 0.007514436729252338, + 0.01352014672011137, + 0.022594748064875603, + 0.014564900659024715, + 0.012858646921813488, + 0.027019500732421875, + 0.013690007850527763, + 0.018265238031744957, + 0.0074869184754788876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0024889700580388308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, starting from a low angle and then gently tilting upward, captured a vast vortex of incandescent gas, swirling violently in the depths of space, illuminated by the hyper-realistic lighting and shadows rendered by Octane. These nebulous gases entwine and collide in complex hydrodynamic paths, forming a dynamic center of energy. Suddenly, a blue dragon, its scales gleaming with cosmic light, its majestic head bursts through the vortex's core, its golden eyes blazing with starlight scanning the void. Then, its slender, sinuous body rises from the vortex in a graceful and powerful gesture. Each deep emerald scale seems to envelop a miniature galaxy, aglow with a constantly shifting glow of deep purple, royal blue, and gold stardust. When it fully displays its sacred and majestic appearance, the surrounding nebula gas is stirred by its powerful energy, forming countless tiny, glowing energy threads and particle streams that entwine around its body, while the cosmic light on its scales interacts with the blazing light of the nebula, casting complex and ever-changing light and shadows on its smooth body. The whole picture presents a sense of grandeur and sacredness of the epic creation of the universe.", + "seed": 16578, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_7417_seed_16578_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": -0.14296141834929585, + "vlm_logprob": -0.14296141834929585, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.14296141834929585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28575074672698975 + ], + "opposite_logprob_tokens": [ + -1.4107507467269897 + ], + "logprob_reward": -0.28575074672698975, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28575074672698975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2830016016960144, + 0.09783771634101868, + 0.12173619866371155, + 0.07101339101791382, + 0.0593787282705307, + 0.20045238733291626, + 0.051991477608680725, + 0.036223188042640686, + 0.05097758769989014, + 0.02738776057958603 + ], + "credit_map": [ + 0.03185860440135002, + 0.026128161698579788, + 0.01800057291984558, + 0.025793831795454025, + 0.03184041753411293, + 0.09060065448284149, + 0.02623571828007698, + 0.03254363313317299, + 0.011575998738408089, + 0.0070602321065962315, + 0.012763313949108124, + 0.012656382285058498, + 0.012246601283550262, + 0.016549715772271156, + 0.008535745553672314, + 0.016449717804789543, + 0.02569829858839512, + 0.013256405480206013, + 0.022170105949044228, + 0.014623994007706642, + 0.009278317913413048, + 0.010358912870287895, + 0.01121180783957243, + 0.0151383550837636, + 0.008992854505777359, + 0.00831129215657711, + 0.007190880365669727, + 0.012252525426447392, + 0.009690530598163605, + 0.0067762876860797405, + 0.012130843475461006, + 0.005668174475431442, + 0.00750932889059186, + 0.008319003507494926, + 0.008324874565005302, + 0.0063611953519284725, + 0.009850715287029743, + 0.004967924673110247, + 0.007379500661045313, + 0.006666178815066814, + 0.030065113678574562, + 0.039825111627578735, + 0.01556494552642107, + 0.020827315747737885, + 0.023686379194259644, + 0.023032506927847862, + 0.021374881267547607, + 0.02607612870633602, + 0.0080908527597785, + 0.010135230608284473, + 0.004405176732689142, + 0.004527867306023836, + 0.006633181590586901, + 0.004503967240452766, + 0.008076757192611694, + 0.005618439055979252, + 0.003467835020273924, + 0.003985861781984568, + 0.003520348807796836, + 0.003998983185738325, + 0.004015302285552025, + 0.005079945083707571, + 0.006316830404102802, + 0.0058380779810249805, + 0.006398539990186691, + 0.0060264673084020615, + 0.005049944389611483, + 0.004582073073834181, + 0.005741997621953487, + 0.011210399679839611, + 0.00687149865552783, + 0.005096668843179941, + 0.0034253918565809727, + 0.0025943692307919264, + 0.00245204521343112, + 0.00284580048173666, + 0.0038867266848683357, + 0.0034167757257819176, + 0.0036282395012676716, + 0.005138410720974207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.28575074672698975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029837431386113167 + ], + "opposite_logprob_tokens": [ + -3.5298373699188232 + ], + "logprob_reward": -0.029837431386113167, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029837431386113167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2808704376220703, + 0.06663036346435547, + 0.10241612046957016, + 0.053566377609968185, + 0.06493014842271805, + 0.0903727263212204, + 0.051708467304706573, + 0.02622891776263714, + 0.13516545295715332, + 0.12811103463172913 + ], + "credit_map": [ + 0.040080469101667404, + 0.020760172978043556, + 0.013653657399117947, + 0.014561714604496956, + 0.02669355273246765, + 0.09681613743305206, + 0.027175651863217354, + 0.04112905636429787, + 0.00649077445268631, + 0.007348714396357536, + 0.006687069311738014, + 0.011812247335910797, + 0.013249030336737633, + 0.006748611107468605, + 0.006696566939353943, + 0.007597346790134907, + 0.026290863752365112, + 0.00782373920083046, + 0.013430643826723099, + 0.006931129842996597, + 0.006262662820518017, + 0.006672096438705921, + 0.010787196457386017, + 0.024217786267399788, + 0.006627559196203947, + 0.004607658367604017, + 0.011824298650026321, + 0.007408106233924627, + 0.00497831916436553, + 0.0032600860577076674, + 0.011482791043817997, + 0.003377554239705205, + 0.007161562796682119, + 0.009199273772537708, + 0.00988147221505642, + 0.010138572193682194, + 0.010096656158566475, + 0.005412530153989792, + 0.007151545491069555, + 0.005888537038117647, + 0.008196085691452026, + 0.01580139994621277, + 0.010127944871783257, + 0.011150242760777473, + 0.01600273698568344, + 0.007859332486987114, + 0.0116334930062294, + 0.009601478464901447, + 0.005242742132395506, + 0.005859377793967724, + 0.005124578718096018, + 0.004992886912077665, + 0.008558576926589012, + 0.005008414387702942, + 0.009989611804485321, + 0.006932276301085949, + 0.001610640319995582, + 0.0030235927551984787, + 0.002328141825273633, + 0.003601404605433345, + 0.004309442359954119, + 0.0015165226068347692, + 0.006278972141444683, + 0.003560200799256563, + 0.012538240291178226, + 0.016699492931365967, + 0.010843691416084766, + 0.015289979055523872, + 0.026604318991303444, + 0.020165400579571724, + 0.017293766140937805, + 0.01573055237531662, + 0.01164132822304964, + 0.01480933465063572, + 0.011406667530536652, + 0.014206736348569393, + 0.025661008432507515, + 0.010682843625545502, + 0.01668398082256317, + 0.023019134998321533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.029837431386113167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428428292274475 + ], + "opposite_logprob_tokens": [ + -2.0178427696228027 + ], + "logprob_reward": -0.1428428292274475, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428428292274475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24303340911865234, + 0.11420457065105438, + 0.14137618243694305, + 0.06333070993423462, + 0.08939700573682785, + 0.0625840350985527, + 0.06396618485450745, + 0.026882965117692947, + 0.11475399136543274, + 0.08047091215848923 + ], + "credit_map": [ + 0.028862403705716133, + 0.018057389184832573, + 0.011597839184105396, + 0.016893060877919197, + 0.02956535294651985, + 0.08056267350912094, + 0.041561685502529144, + 0.015933003276586533, + 0.011027628555893898, + 0.008283794857561588, + 0.012814894318580627, + 0.010067910887300968, + 0.013500670902431011, + 0.022594671696424484, + 0.007323997560888529, + 0.028591012582182884, + 0.03600414842367172, + 0.01038861833512783, + 0.012099155224859715, + 0.008247224614024162, + 0.00789890717715025, + 0.010047761723399162, + 0.015594136901199818, + 0.04109623283147812, + 0.007031631655991077, + 0.005992457270622253, + 0.012043453752994537, + 0.009917302057147026, + 0.007816310040652752, + 0.0037393756210803986, + 0.01300316583365202, + 0.003787011606618762, + 0.009449642151594162, + 0.013405781239271164, + 0.015062485821545124, + 0.008865365758538246, + 0.01374670211225748, + 0.009561203420162201, + 0.011557997204363346, + 0.007747830357402563, + 0.006013078615069389, + 0.008876456879079342, + 0.006534948945045471, + 0.0064035640098154545, + 0.011075658723711967, + 0.006314438302069902, + 0.008254745975136757, + 0.009111144579946995, + 0.007554164156317711, + 0.009008830413222313, + 0.006077529396861792, + 0.00498255155980587, + 0.010800953954458237, + 0.00678250240162015, + 0.01130725722759962, + 0.007452399469912052, + 0.003223808715119958, + 0.0035680707078427076, + 0.00252264435403049, + 0.0031405382324010134, + 0.0031802160665392876, + 0.0039020872209221125, + 0.0031338052358478308, + 0.004211793188005686, + 0.017609337344765663, + 0.010281560942530632, + 0.011997575871646404, + 0.01102315541356802, + 0.012833822518587112, + 0.021684177219867706, + 0.01873309351503849, + 0.01059127226471901, + 0.008147457614541054, + 0.01053712796419859, + 0.0078648766502738, + 0.010120336897671223, + 0.012566905468702316, + 0.010699253529310226, + 0.008717712946236134, + 0.011817236430943012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.1428428292274475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25227832794189453 + ], + "opposite_logprob_tokens": [ + -1.5022783279418945 + ], + "logprob_reward": -0.25227832794189453, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25227832794189453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2740758955478668, + 0.12948361039161682, + 0.11296256631612778, + 0.11076987534761429, + 0.07843077927827835, + 0.05829790234565735, + 0.04259892553091049, + 0.04856716841459274, + 0.0487438403069973, + 0.09606941044330597 + ], + "credit_map": [ + 0.03878067061305046, + 0.018338702619075775, + 0.015345849096775055, + 0.01757434383034706, + 0.032166413962841034, + 0.09465794265270233, + 0.029591333121061325, + 0.02762061543762684, + 0.01358070783317089, + 0.010365284979343414, + 0.01226123794913292, + 0.016404567286372185, + 0.01658495143055916, + 0.020557748153805733, + 0.010955234058201313, + 0.028773874044418335, + 0.020789939910173416, + 0.014873973093926907, + 0.017776181921362877, + 0.012579845264554024, + 0.009050407446920872, + 0.007066147867590189, + 0.014824507758021355, + 0.016001567244529724, + 0.012714101932942867, + 0.009891131892800331, + 0.022691775113344193, + 0.017741356045007706, + 0.011586079373955727, + 0.006367682944983244, + 0.02259421907365322, + 0.007183531299233437, + 0.007651430554687977, + 0.009443648159503937, + 0.013484831899404526, + 0.00870576873421669, + 0.01241040788590908, + 0.009745219722390175, + 0.01046908088028431, + 0.006520390510559082, + 0.00585705554112792, + 0.007069396786391735, + 0.00670900521799922, + 0.0060866777785122395, + 0.009408003650605679, + 0.007213167380541563, + 0.007728849072009325, + 0.008225744590163231, + 0.005940073169767857, + 0.005270854569971561, + 0.003656015731394291, + 0.004644913133233786, + 0.004327023401856422, + 0.00385775207541883, + 0.009139925241470337, + 0.005762370768934488, + 0.006829148158431053, + 0.004304228350520134, + 0.003687325632199645, + 0.005435790866613388, + 0.00472696078941226, + 0.006101774051785469, + 0.008962600491940975, + 0.008519344963133335, + 0.003351801075041294, + 0.005607951432466507, + 0.004459107760339975, + 0.006098051089793444, + 0.010428126901388168, + 0.004989979322999716, + 0.008632069453597069, + 0.005176757927983999, + 0.009754982776939869, + 0.01183908898383379, + 0.008794323541224003, + 0.011527488008141518, + 0.016489649191498756, + 0.012826300226151943, + 0.010338600724935532, + 0.014498980715870857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.25227832794189453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32568490505218506, + 0.09209033846855164, + 0.07980428636074066, + 0.07230471074581146, + 0.08139931410551071, + 0.07155753672122955, + 0.07455594837665558, + 0.03669610619544983, + 0.11399303376674652, + 0.05191386118531227 + ], + "credit_map": [ + 0.04185432195663452, + 0.02809959277510643, + 0.01812123879790306, + 0.01763860508799553, + 0.038134243339300156, + 0.10660328716039658, + 0.042899832129478455, + 0.032333798706531525, + 0.008000979200005531, + 0.006510763429105282, + 0.011473487131297588, + 0.011173014529049397, + 0.011025619693100452, + 0.02002454549074173, + 0.007868051528930664, + 0.01601388305425644, + 0.023113584145903587, + 0.005640121642500162, + 0.011428515426814556, + 0.005366398487240076, + 0.004134170711040497, + 0.004608249291777611, + 0.00541864987462759, + 0.020094597712159157, + 0.0082717165350914, + 0.007515120320022106, + 0.0138918561860919, + 0.013018330559134483, + 0.00796473864465952, + 0.004371489863842726, + 0.012648182921111584, + 0.004623271059244871, + 0.008182257413864136, + 0.009529604576528072, + 0.014046913012862206, + 0.015310604125261307, + 0.01025892049074173, + 0.007284344639629126, + 0.010832749307155609, + 0.0059539200738072395, + 0.005987999029457569, + 0.010554147884249687, + 0.007652487605810165, + 0.008205519057810307, + 0.01360086165368557, + 0.005927043966948986, + 0.008917993865907192, + 0.010711479932069778, + 0.006301394663751125, + 0.009596046060323715, + 0.008610835298895836, + 0.006794732064008713, + 0.011740397661924362, + 0.006428518798202276, + 0.01660098508000374, + 0.008483040146529675, + 0.004010723438113928, + 0.003554780036211014, + 0.0034468378871679306, + 0.0039489055052399635, + 0.0043297335505485535, + 0.00584652042016387, + 0.006466899067163467, + 0.005091708619147539, + 0.010480495169758797, + 0.014166329056024551, + 0.009701995179057121, + 0.013991321437060833, + 0.02487899735569954, + 0.010480422526597977, + 0.016721457242965698, + 0.01357201673090458, + 0.004582670051604509, + 0.005410284735262394, + 0.004040983039885759, + 0.0055236187763512135, + 0.009555792436003685, + 0.003746339352801442, + 0.008993429131805897, + 0.010060741566121578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.004097756464034319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, starting from a low angle and then gently tilting upward, captured a vast vortex of incandescent gas, swirling violently in the depths of space, illuminated by the hyper-realistic lighting and shadows rendered by Octane. These nebulous gases entwine and collide in complex hydrodynamic paths, forming a dynamic center of energy. Suddenly, a blue dragon, its scales gleaming with cosmic light, its majestic head bursts through the vortex's core, its golden eyes blazing with starlight scanning the void. Then, its slender, sinuous body rises from the vortex in a graceful and powerful gesture. Each deep emerald scale seems to envelop a miniature galaxy, aglow with a constantly shifting glow of deep purple, royal blue, and gold stardust. When it fully displays its sacred and majestic appearance, the surrounding nebula gas is stirred by its powerful energy, forming countless tiny, glowing energy threads and particle streams that entwine around its body, while the cosmic light on its scales interacts with the blazing light of the nebula, casting complex and ever-changing light and shadows on its smooth body. The whole picture presents a sense of grandeur and sacredness of the epic creation of the universe.", + "seed": 16579, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_7417_seed_16579_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": -0.14385435497388244, + "vlm_logprob": -0.14385435497388244, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.14385435497388244, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28689831495285034 + ], + "opposite_logprob_tokens": [ + -1.4118983745574951 + ], + "logprob_reward": -0.28689831495285034, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28689831495285034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4377397894859314, + 0.08291400969028473, + 0.1652594953775406, + 0.055063359439373016, + 0.0775735154747963, + 0.026289675384759903, + 0.03155077248811722, + 0.029551910236477852, + 0.061067719012498856, + 0.03298972547054291 + ], + "credit_map": [ + 0.056416116654872894, + 0.05560492351651192, + 0.05205916613340378, + 0.03941130265593529, + 0.054101020097732544, + 0.07480020076036453, + 0.052196383476257324, + 0.053150683641433716, + 0.011594047769904137, + 0.00929067563265562, + 0.013147414661943913, + 0.010132644325494766, + 0.008068110793828964, + 0.010598724707961082, + 0.006798539776355028, + 0.013283859007060528, + 0.011358309537172318, + 0.01737806759774685, + 0.01456441543996334, + 0.014967062510550022, + 0.03883738815784454, + 0.022219935432076454, + 0.02071276679635048, + 0.025221548974514008, + 0.005521694663912058, + 0.009198806248605251, + 0.008007364347577095, + 0.008429338224232197, + 0.005567688960582018, + 0.0055060917511582375, + 0.0064248922280967236, + 0.006407479755580425, + 0.010075146332383156, + 0.009233295917510986, + 0.005981655791401863, + 0.00558461481705308, + 0.022328412160277367, + 0.004841870162636042, + 0.008127493783831596, + 0.011401022784411907, + 0.004150861408561468, + 0.003247444983571768, + 0.00376683147624135, + 0.0029545053839683533, + 0.003032504813745618, + 0.001992557430639863, + 0.0037372622173279524, + 0.0034077090676873922, + 0.00329931010492146, + 0.0034278484527021646, + 0.005453863181173801, + 0.004544787108898163, + 0.003509582718834281, + 0.002354757161810994, + 0.003282519755885005, + 0.005678102839738131, + 0.006093289237469435, + 0.003678394015878439, + 0.004918166436254978, + 0.0038868868723511696, + 0.0034810954239219427, + 0.0024432006757706404, + 0.002392078982666135, + 0.0026587974280118942, + 0.005915830377489328, + 0.008983553387224674, + 0.00857485644519329, + 0.00837935321033001, + 0.011022524908185005, + 0.0053915646858513355, + 0.006151529494673014, + 0.006648501381278038, + 0.004289657808840275, + 0.005389745347201824, + 0.0030240900814533234, + 0.004254419822245836, + 0.005046045407652855, + 0.0038244007155299187, + 0.003976354841142893, + 0.003185013774782419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.28689831495285034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225504361093044 + ], + "opposite_logprob_tokens": [ + -4.264225482940674 + ], + "logprob_reward": -0.014225504361093044, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225504361093044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1056821420788765, + 0.09731940180063248, + 0.24423399567604065, + 0.10975928604602814, + 0.09330177307128906, + 0.03998669236898422, + 0.05126836895942688, + 0.04418165609240532, + 0.12464561313390732, + 0.08962107449769974 + ], + "credit_map": [ + 0.013044455088675022, + 0.008184564299881458, + 0.015559786930680275, + 0.010562415234744549, + 0.012710748240351677, + 0.020176153630018234, + 0.011175880208611488, + 0.014268138445913792, + 0.008563479408621788, + 0.020128343254327774, + 0.009900794364511967, + 0.007851035334169865, + 0.00937364250421524, + 0.013882795348763466, + 0.005507409106940031, + 0.022111903876066208, + 0.012984338216483593, + 0.01738768257200718, + 0.007259218487888575, + 0.014117350801825523, + 0.03823758289217949, + 0.08604727685451508, + 0.011297928169369698, + 0.05690261349081993, + 0.010738594457507133, + 0.018045689910650253, + 0.008097108453512192, + 0.019099105149507523, + 0.017099156975746155, + 0.010724791325628757, + 0.013540569692850113, + 0.01241427380591631, + 0.014883337542414665, + 0.012624022550880909, + 0.0069862231612205505, + 0.006249601021409035, + 0.029201563447713852, + 0.005261252634227276, + 0.009041950106620789, + 0.009053824469447136, + 0.00537051260471344, + 0.0056134541518986225, + 0.004148073494434357, + 0.005338301416486502, + 0.004495586734265089, + 0.002097499091178179, + 0.00929353479295969, + 0.0036297314800322056, + 0.0065896580927073956, + 0.0056917546316981316, + 0.011136434972286224, + 0.006912854965776205, + 0.005078835878521204, + 0.00406251335516572, + 0.004769555758684874, + 0.0070267608389258385, + 0.006500727962702513, + 0.005329909268766642, + 0.005811918061226606, + 0.006118251010775566, + 0.005720582790672779, + 0.005046155769377947, + 0.0030785591807216406, + 0.006575551349669695, + 0.012881835922598839, + 0.01681402139365673, + 0.018820103257894516, + 0.019164176657795906, + 0.020186468958854675, + 0.010932894423604012, + 0.012532401829957962, + 0.013313704170286655, + 0.011555342003703117, + 0.017404770478606224, + 0.007385562174022198, + 0.014447641558945179, + 0.010876726359128952, + 0.009350908920168877, + 0.012007588520646095, + 0.006592533551156521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.014225504361093044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10036412626504898 + ], + "opposite_logprob_tokens": [ + -2.3503642082214355 + ], + "logprob_reward": -0.10036412626504898, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10036412626504898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09732576459646225, + 0.12822017073631287, + 0.08383041620254517, + 0.10872909426689148, + 0.2533193528652191, + 0.07575022429227829, + 0.08147076517343521, + 0.04599315673112869, + 0.08312221616506577, + 0.04223890230059624 + ], + "credit_map": [ + 0.013783549889922142, + 0.009338105097413063, + 0.01499202474951744, + 0.009973120875656605, + 0.010370030999183655, + 0.016595348715782166, + 0.01126913446933031, + 0.011004445143043995, + 0.01206616684794426, + 0.021095316857099533, + 0.015042895451188087, + 0.007612105458974838, + 0.01017997320741415, + 0.016099052503705025, + 0.009775981307029724, + 0.03634866699576378, + 0.0060761673375964165, + 0.009436949156224728, + 0.007030643057078123, + 0.01104265172034502, + 0.019062425941228867, + 0.011948684230446815, + 0.010188163258135319, + 0.009044725447893143, + 0.009426895529031754, + 0.01637338660657406, + 0.009514554403722286, + 0.015531825833022594, + 0.018984531983733177, + 0.010820664465427399, + 0.013589957728981972, + 0.014487269334495068, + 0.048411495983600616, + 0.03144988790154457, + 0.01681087166070938, + 0.014037101529538631, + 0.08381767570972443, + 0.012713159434497356, + 0.023162558674812317, + 0.02291659079492092, + 0.011137301102280617, + 0.014210465364158154, + 0.009109762497246265, + 0.008699617348611355, + 0.008909555152058601, + 0.004018286243081093, + 0.01459533255547285, + 0.005069897044450045, + 0.009370114654302597, + 0.010518375784158707, + 0.011847848072648048, + 0.014875500462949276, + 0.00989675335586071, + 0.005976258311420679, + 0.008740494027733803, + 0.010245411656796932, + 0.01028626598417759, + 0.00584057392552495, + 0.005572046618908644, + 0.006398875266313553, + 0.005525786429643631, + 0.003743563313037157, + 0.003201681887730956, + 0.005424359813332558, + 0.010132343508303165, + 0.011981927789747715, + 0.01250054594129324, + 0.010683714412152767, + 0.010221877135336399, + 0.0079183429479599, + 0.0071299416013062, + 0.012553523294627666, + 0.005163033492863178, + 0.00790619757026434, + 0.0033539871219545603, + 0.005326931830495596, + 0.007087817881256342, + 0.0044686803594231606, + 0.00582365645095706, + 0.0031085945665836334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.10036412626504898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3136860728263855 + ], + "opposite_logprob_tokens": [ + -1.3136861324310303 + ], + "logprob_reward": -0.3136860728263855, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3136860728263855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18802212178707123, + 0.17127978801727295, + 0.11129041016101837, + 0.09422191977500916, + 0.08120976388454437, + 0.04867187887430191, + 0.052990447729825974, + 0.030124951153993607, + 0.1598430573940277, + 0.062345702201128006 + ], + "credit_map": [ + 0.021265428513288498, + 0.02307121641933918, + 0.020266564562916756, + 0.018029311671853065, + 0.026325516402721405, + 0.032279618084430695, + 0.02279827557504177, + 0.02398619055747986, + 0.014068396762013435, + 0.032410942018032074, + 0.019144302234053612, + 0.01144692674279213, + 0.014771842397749424, + 0.02250998467206955, + 0.012522608041763306, + 0.0444047711789608, + 0.008445612154901028, + 0.01083285827189684, + 0.007156748790293932, + 0.010616053827106953, + 0.012296787463128567, + 0.02618294768035412, + 0.01027726661413908, + 0.02548213303089142, + 0.00967690534889698, + 0.020439578220248222, + 0.010952115058898926, + 0.019842790439724922, + 0.008635378442704678, + 0.005571932531893253, + 0.012238463386893272, + 0.0068647549487650394, + 0.01277693547308445, + 0.011067650280892849, + 0.005974683444947004, + 0.007469623349606991, + 0.022762037813663483, + 0.004605095833539963, + 0.008887728676199913, + 0.007666011806577444, + 0.005767632741481066, + 0.007713425438851118, + 0.005816163029521704, + 0.0069220708683133125, + 0.006910874508321285, + 0.002636190503835678, + 0.009499890729784966, + 0.0034056282602250576, + 0.00735481595620513, + 0.00646985275670886, + 0.011317073367536068, + 0.008528390899300575, + 0.005547250621020794, + 0.0038974490016698837, + 0.005054746754467487, + 0.004820867907255888, + 0.0050613596104085445, + 0.0033933776430785656, + 0.003380958689376712, + 0.004314864985644817, + 0.0038923751562833786, + 0.002881623338907957, + 0.002809845842421055, + 0.00439054798334837, + 0.015452900901436806, + 0.021300483494997025, + 0.024821054190397263, + 0.02510906383395195, + 0.024343855679035187, + 0.013104666024446487, + 0.017432454973459244, + 0.018278582021594048, + 0.008472895249724388, + 0.013205152004957199, + 0.005802525207400322, + 0.0070663499645888805, + 0.009214473888278008, + 0.006187901832163334, + 0.00769729632884264, + 0.004699104465544224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.3136860728263855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097756464034319 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004097756464034319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097756464034319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09503153711557388, + 0.10222110897302628, + 0.08215446025133133, + 0.08347800374031067, + 0.15397128462791443, + 0.1170864924788475, + 0.06609752029180527, + 0.058664098381996155, + 0.11124921590089798, + 0.13004626333713531 + ], + "credit_map": [ + 0.011787793599069118, + 0.00683145597577095, + 0.01715707965195179, + 0.012013995088636875, + 0.010295061394572258, + 0.01498924195766449, + 0.012341910041868687, + 0.009615001268684864, + 0.011836269870400429, + 0.008867085911333561, + 0.015005460008978844, + 0.008254197426140308, + 0.009887025691568851, + 0.01282084733247757, + 0.010871093720197678, + 0.02467912621796131, + 0.007694660685956478, + 0.007784139830619097, + 0.006828548852354288, + 0.007513006217777729, + 0.015280363149940968, + 0.012882079929113388, + 0.008497623726725578, + 0.015674035996198654, + 0.009697653353214264, + 0.013772184029221535, + 0.00783571694046259, + 0.014319514855742455, + 0.013661473989486694, + 0.005893238354474306, + 0.01075963955372572, + 0.0075385854579508305, + 0.031679846346378326, + 0.019575418904423714, + 0.008772672154009342, + 0.007685027085244656, + 0.05245108902454376, + 0.0069802384823560715, + 0.013622695580124855, + 0.01320429053157568, + 0.016536042094230652, + 0.019209975376725197, + 0.015940308570861816, + 0.017390109598636627, + 0.013050241395831108, + 0.007364312186837196, + 0.01626817137002945, + 0.011327334679663181, + 0.007581322453916073, + 0.008020187728106976, + 0.015188152901828289, + 0.010823857970535755, + 0.005789712071418762, + 0.003802536753937602, + 0.006779022049158812, + 0.00811273418366909, + 0.007494267541915178, + 0.007368798833340406, + 0.009894494898617268, + 0.00984791200608015, + 0.006761648692190647, + 0.004642082843929529, + 0.005478668492287397, + 0.007176228333264589, + 0.012306060642004013, + 0.018799616023898125, + 0.009799963794648647, + 0.015409151092171669, + 0.01581106334924698, + 0.009566821157932281, + 0.010917982086539268, + 0.01863856054842472, + 0.020084591582417488, + 0.020428616553544998, + 0.017420025542378426, + 0.015327158384025097, + 0.010321839712560177, + 0.015931280329823494, + 0.014259745366871357, + 0.01627301797270775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.004097756464034319 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, starting from a low angle and then gently tilting upward, captured a vast vortex of incandescent gas, swirling violently in the depths of space, illuminated by the hyper-realistic lighting and shadows rendered by Octane. These nebulous gases entwine and collide in complex hydrodynamic paths, forming a dynamic center of energy. Suddenly, a blue dragon, its scales gleaming with cosmic light, its majestic head bursts through the vortex's core, its golden eyes blazing with starlight scanning the void. Then, its slender, sinuous body rises from the vortex in a graceful and powerful gesture. Each deep emerald scale seems to envelop a miniature galaxy, aglow with a constantly shifting glow of deep purple, royal blue, and gold stardust. When it fully displays its sacred and majestic appearance, the surrounding nebula gas is stirred by its powerful energy, forming countless tiny, glowing energy threads and particle streams that entwine around its body, while the cosmic light on its scales interacts with the blazing light of the nebula, casting complex and ever-changing light and shadows on its smooth body. The whole picture presents a sense of grandeur and sacredness of the epic creation of the universe.", + "seed": 16580, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_14/index_7417_seed_16580_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": -0.16348761143162846, + "vlm_logprob": -0.16348761143162846, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.16348761143162846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5291131138801575 + ], + "opposite_logprob_tokens": [ + -0.9041131138801575 + ], + "logprob_reward": -0.5291131138801575, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5291131138801575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21975791454315186, + 0.05274059250950813, + 0.07494943588972092, + 0.0937250554561615, + 0.09784304350614548, + 0.12069274485111237, + 0.21677576005458832, + 0.033884987235069275, + 0.06034413352608681, + 0.02928629331290722 + ], + "credit_map": [ + 0.023789357393980026, + 0.020595291629433632, + 0.014800108037889004, + 0.01942426897585392, + 0.03300236165523529, + 0.023635143414139748, + 0.03443706035614014, + 0.05007433146238327, + 0.006320653483271599, + 0.007557344157248735, + 0.006104639265686274, + 0.0061826445162296295, + 0.00922326184809208, + 0.005122967530041933, + 0.005687509197741747, + 0.00654157530516386, + 0.008931144140660763, + 0.010491393506526947, + 0.018779657781124115, + 0.015477805398404598, + 0.004621792584657669, + 0.004840492736548185, + 0.0071720858104527, + 0.004635067190974951, + 0.009048724547028542, + 0.01743585243821144, + 0.011499510146677494, + 0.012925512157380581, + 0.014230439439415932, + 0.00970712210983038, + 0.010320596396923065, + 0.008557296358048916, + 0.012713080272078514, + 0.013380363583564758, + 0.01679679937660694, + 0.009402181953191757, + 0.013268849812448025, + 0.006493785884231329, + 0.012837529182434082, + 0.012950455769896507, + 0.007938763126730919, + 0.04263841360807419, + 0.03809020295739174, + 0.007332524750381708, + 0.006780008785426617, + 0.00288963015191257, + 0.00810733251273632, + 0.006915870122611523, + 0.016227448359131813, + 0.0826752781867981, + 0.019841518253087997, + 0.05344779044389725, + 0.015084334649145603, + 0.0074562751688063145, + 0.011555596254765987, + 0.010487505234777927, + 0.004343804437667131, + 0.002912396565079689, + 0.007105038035660982, + 0.0036357673816382885, + 0.0030501799192279577, + 0.004719694145023823, + 0.002568817464634776, + 0.005549288354814053, + 0.006057192571461201, + 0.008879669941961765, + 0.009438181295990944, + 0.00666130380704999, + 0.009958377107977867, + 0.008307013660669327, + 0.007409518584609032, + 0.0036328767891973257, + 0.003817445831373334, + 0.003230402246117592, + 0.004897437058389187, + 0.0033043522853404284, + 0.0037934212014079094, + 0.005019668489694595, + 0.0025430843234062195, + 0.0026804793160408735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.5291131138801575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014231615699827671 + ], + "opposite_logprob_tokens": [ + -4.2642316818237305 + ], + "logprob_reward": -0.014231615699827671, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014231615699827671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1561974287033081, + 0.05996141582727432, + 0.058100372552871704, + 0.14695291221141815, + 0.04797590896487236, + 0.0856630876660347, + 0.09360640496015549, + 0.09338365495204926, + 0.1574915200471878, + 0.10066724568605423 + ], + "credit_map": [ + 0.015370707958936691, + 0.014075962826609612, + 0.01002295408397913, + 0.012224327772855759, + 0.019869869574904442, + 0.017924003303050995, + 0.02402319386601448, + 0.04268641024827957, + 0.0051505062729120255, + 0.007046403829008341, + 0.007170452270656824, + 0.008043716661632061, + 0.009243051521480083, + 0.006866049021482468, + 0.008126460947096348, + 0.008314778096973896, + 0.006005981005728245, + 0.009563690982758999, + 0.01108157355338335, + 0.00950294267386198, + 0.0056197382509708405, + 0.005924387369304895, + 0.005601375829428434, + 0.004800684750080109, + 0.0124591626226902, + 0.018246280029416084, + 0.01799107901751995, + 0.013706713914871216, + 0.02519778534770012, + 0.02682877518236637, + 0.018433162942528725, + 0.014089962467551231, + 0.003346764948219061, + 0.01107872650027275, + 0.011848323047161102, + 0.003347051329910755, + 0.004856937099248171, + 0.003231215290725231, + 0.004757205955684185, + 0.005509683862328529, + 0.01125177089124918, + 0.013470153324306011, + 0.01814962364733219, + 0.00965957436710596, + 0.004940085578709841, + 0.0038741398602724075, + 0.012882459908723831, + 0.011435281485319138, + 0.01610684208571911, + 0.012374494224786758, + 0.009574533440172672, + 0.021932780742645264, + 0.0093934815376997, + 0.007813417352735996, + 0.007803723681718111, + 0.008607123978435993, + 0.0109372204169631, + 0.012093370780348778, + 0.014009580947458744, + 0.01191762462258339, + 0.016039134934544563, + 0.009402248077094555, + 0.010599326342344284, + 0.008385148830711842, + 0.015837697312235832, + 0.02327509969472885, + 0.022174248471856117, + 0.019539866596460342, + 0.02376657910645008, + 0.020462438464164734, + 0.02121591754257679, + 0.011219672858715057, + 0.010964772664010525, + 0.012673486024141312, + 0.012375233694911003, + 0.013957057148218155, + 0.01478313747793436, + 0.013761790469288826, + 0.015150283463299274, + 0.007001478690654039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.014231615699827671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710222601890564 + ], + "opposite_logprob_tokens": [ + -2.1271021366119385 + ], + "logprob_reward": -0.12710222601890564, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12710222601890564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10397277027368546, + 0.03433515504002571, + 0.04063063859939575, + 0.08631490916013718, + 0.10990006476640701, + 0.19245044887065887, + 0.1451849490404129, + 0.06480515003204346, + 0.09422548860311508, + 0.12818042933940887 + ], + "credit_map": [ + 0.012998947873711586, + 0.008496533147990704, + 0.008201755583286285, + 0.009676050394773483, + 0.01296449638903141, + 0.011776715517044067, + 0.012887349352240562, + 0.026970919221639633, + 0.004248384851962328, + 0.003575591603294015, + 0.0045446958392858505, + 0.004163593519479036, + 0.003941298928111792, + 0.004924437031149864, + 0.003741632914170623, + 0.005195518024265766, + 0.004036932718008757, + 0.006217800546437502, + 0.010433116927742958, + 0.008212389424443245, + 0.0027185454964637756, + 0.0025236494839191437, + 0.0039825523272156715, + 0.0025056481827050447, + 0.008282223716378212, + 0.011465109884738922, + 0.010942953638732433, + 0.011128495447337627, + 0.009536733850836754, + 0.0123578617349267, + 0.011663759127259254, + 0.010937763378024101, + 0.01123478077352047, + 0.014057393185794353, + 0.011190131306648254, + 0.009526626206934452, + 0.020609261468052864, + 0.009549831971526146, + 0.01920275203883648, + 0.014529266394674778, + 0.018248604610562325, + 0.06007176265120506, + 0.05335618183016777, + 0.011419815011322498, + 0.01142969261854887, + 0.004742590244859457, + 0.014627019874751568, + 0.018554767593741417, + 0.01207263208925724, + 0.050202809274196625, + 0.015020961873233318, + 0.02939174696803093, + 0.011120134964585304, + 0.008204755373299122, + 0.009369800798594952, + 0.00980209931731224, + 0.007279952988028526, + 0.007380873430520296, + 0.010982239618897438, + 0.008777942508459091, + 0.010792813263833523, + 0.007018475793302059, + 0.005665136966854334, + 0.006907709874212742, + 0.014690730720758438, + 0.008819819428026676, + 0.012480713427066803, + 0.008262749761343002, + 0.008481430821120739, + 0.013939627446234226, + 0.010095097124576569, + 0.01745530590415001, + 0.012102887034416199, + 0.02304425835609436, + 0.012031176127493382, + 0.014713851734995842, + 0.02257874421775341, + 0.017825867980718613, + 0.016964618116617203, + 0.00891901832073927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.12710222601890564 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428934633731842 + ], + "opposite_logprob_tokens": [ + -2.0178935527801514 + ], + "logprob_reward": -0.1428934633731842, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428934633731842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13942714035511017, + 0.04668935760855675, + 0.07131963968276978, + 0.033832449465990067, + 0.07784707099199295, + 0.2444910705089569, + 0.04347647726535797, + 0.09602279216051102, + 0.09855133295059204, + 0.1483427733182907 + ], + "credit_map": [ + 0.015102099627256393, + 0.01343417726457119, + 0.009011678397655487, + 0.012852651067078114, + 0.018444878980517387, + 0.014762884937226772, + 0.017351578921079636, + 0.038467176258563995, + 0.004391447640955448, + 0.005823899060487747, + 0.004952997900545597, + 0.007304054219275713, + 0.007864377461373806, + 0.005113510414958, + 0.005311132874339819, + 0.005927931051701307, + 0.00943948794156313, + 0.011364196427166462, + 0.01654873788356781, + 0.013914293609559536, + 0.004625932779163122, + 0.004694173578172922, + 0.005662164185196161, + 0.005070640705525875, + 0.004030296113342047, + 0.004129726905375719, + 0.005847449880093336, + 0.004039531107991934, + 0.003718574997037649, + 0.004159438423812389, + 0.003585378173738718, + 0.004322051536291838, + 0.007432535290718079, + 0.01707923784852028, + 0.017950227484107018, + 0.006793754640966654, + 0.007535087410360575, + 0.0052504464983940125, + 0.009750286117196083, + 0.006055490579456091, + 0.020872583612799644, + 0.07465299218893051, + 0.07219811528921127, + 0.01596836745738983, + 0.012577400542795658, + 0.006922267377376556, + 0.01954367198050022, + 0.0217556431889534, + 0.006907492410391569, + 0.007364924065768719, + 0.003993401303887367, + 0.009482167661190033, + 0.0035918187350034714, + 0.00433286651968956, + 0.0036571407690644264, + 0.004146660678088665, + 0.00895470567047596, + 0.01117617916315794, + 0.025887420400977135, + 0.010253027081489563, + 0.010434217751026154, + 0.01207680907100439, + 0.008171016350388527, + 0.009069406427443027, + 0.014111240394413471, + 0.012527192942798138, + 0.013584605418145657, + 0.008601298555731773, + 0.013479883782565594, + 0.015005741268396378, + 0.009679289534687996, + 0.011562070809304714, + 0.014620441943407059, + 0.018814153969287872, + 0.024907367303967476, + 0.015501143410801888, + 0.025170831009745598, + 0.01991901360452175, + 0.017742952331900597, + 0.011666848324239254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.1428934633731842 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 start with a low angle shot that gently tilts upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a blue dragon visible bursting through the core of a swirling vortex of incandescent gas?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the dragon's scales appear deep emerald with a shifting glow of deep purple, royal blue, and gold stardust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the dragon's golden eyes blaze with starlight while scanning the void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do glowing energy threads and particle streams entwine around the dragon's body as the surrounding nebula gas is stirred?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097638186067343 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097638186067343, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097638186067343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16491875052452087, + 0.030747082084417343, + 0.06897325813770294, + 0.05929889529943466, + 0.05098939687013626, + 0.13889671862125397, + 0.2391951084136963, + 0.049452997744083405, + 0.08567590266466141, + 0.11185194551944733 + ], + "credit_map": [ + 0.016965510323643684, + 0.014331715181469917, + 0.011430731043219566, + 0.01220884919166565, + 0.017645923420786858, + 0.019740097224712372, + 0.027000760659575462, + 0.04559515416622162, + 0.003817314049229026, + 0.00414192583411932, + 0.0032207586336880922, + 0.005397663451731205, + 0.004227898083627224, + 0.0028073291759938, + 0.0039567891508340836, + 0.003177401376888156, + 0.005830444861203432, + 0.008434855379164219, + 0.021783648058772087, + 0.015360603109002113, + 0.005408274009823799, + 0.003774921875447035, + 0.005275703500956297, + 0.003104814328253269, + 0.005981937050819397, + 0.01074246782809496, + 0.008435560390353203, + 0.007992452941834927, + 0.00885026715695858, + 0.006666164845228195, + 0.005237980280071497, + 0.0053920624777674675, + 0.0061886305920779705, + 0.006345947738736868, + 0.006288791075348854, + 0.007622207514941692, + 0.007054734043776989, + 0.0034217063803225756, + 0.008361857384443283, + 0.00570552796125412, + 0.016104059293866158, + 0.02974838763475418, + 0.03182927891612053, + 0.01282835565507412, + 0.007288340479135513, + 0.005025904625654221, + 0.019655192270874977, + 0.016417207196354866, + 0.023817241191864014, + 0.10222592949867249, + 0.02127058617770672, + 0.03885757923126221, + 0.021382074803113937, + 0.00840703397989273, + 0.016060644760727882, + 0.007174025755375624, + 0.006220773793756962, + 0.006149572785943747, + 0.006190005224198103, + 0.00635159807279706, + 0.008943363092839718, + 0.006785204634070396, + 0.0050263251177966595, + 0.003786151297390461, + 0.008847573772072792, + 0.01160599198192358, + 0.014343268238008022, + 0.010686369612812996, + 0.012930063530802727, + 0.011876801960170269, + 0.010845877230167389, + 0.004539958667010069, + 0.01025758869946003, + 0.01606934703886509, + 0.01305640209466219, + 0.01612734980881214, + 0.012939773499965668, + 0.015583466738462448, + 0.01984228566288948, + 0.00797573383897543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.004097638186067343 + } + ], + "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/0000022/group_15/index_tensor([3082, 3082])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_tensor([3082, 3082])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a610c0f0486c6eb7b765de89e234c1e7f29b4f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_tensor([3082, 3082])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "majestic low-angle shot begins, the camera slowly and steadily tilting upward from the heart of Venezuela's dense, emerald jungle, following the breathtaking descent of Angel Falls. A powerful, relentless stream of white, foaming water cascades down from the flat-topped summit of a massive, ancient cliff face. The water cascades relentlessly down the sheer, weathered rock face, a tapestry of dark gray and ochre covered with damp green moss. As the water plummets through the air, it atomizes into a vast, billowing cloud of fine mist, catching the bright, direct sunlight and creating delicate, shimmering rainbows near the base. The massive waterfall is surrounded by lush, vibrant tropical vegetation on the lower portion of the cliff, forming a dense canopy below, providing a stark, beautiful contrast between life and stone. Above, the scene is crowned by a vast, clear, intense azure sky, completely devoid of cloud, amplifying the spectacular spectacle. The entire scene is filmed on an epic scale, encapsulating the raw, untouched and awe-inspiring beauty of this world-famous landmark, creating an atmosphere that is both serene and incredibly powerful.", + "seed": 14008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_3082_seed_14008_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": -0.10063628405332566, + "vlm_logprob": -0.10063628405332566, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.10063628405332566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3930879533290863 + ], + "opposite_logprob_tokens": [ + -1.1430879831314087 + ], + "logprob_reward": -0.3930879533290863, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3930879533290863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19751952588558197, + 0.2459055632352829, + 0.14854204654693604, + 0.0950443372130394, + 0.059490956366062164, + 0.07242301106452942, + 0.03638844937086105, + 0.033308565616607666, + 0.06322091817855835, + 0.04815660044550896 + ], + "credit_map": [ + 0.014437757432460785, + 0.018276691436767578, + 0.029480259865522385, + 0.028086865320801735, + 0.0252738855779171, + 0.024819539859890938, + 0.03346043825149536, + 0.02368411049246788, + 0.023277873173356056, + 0.023357897996902466, + 0.02485411986708641, + 0.04072904959321022, + 0.027108008041977882, + 0.04049135744571686, + 0.031054014340043068, + 0.03503326326608658, + 0.01581341214478016, + 0.013922804035246372, + 0.028124922886490822, + 0.03360635042190552, + 0.013094188645482063, + 0.011032783426344395, + 0.019920969381928444, + 0.013026626780629158, + 0.011899277567863464, + 0.011783218011260033, + 0.01672757975757122, + 0.019404899328947067, + 0.008703800849616528, + 0.007075357250869274, + 0.011355949565768242, + 0.008094258606433868, + 0.008308446034789085, + 0.005837494041770697, + 0.009756244719028473, + 0.010214055888354778, + 0.005465880502015352, + 0.005287938751280308, + 0.007796354126185179, + 0.006824549753218889, + 0.012127948924899101, + 0.012505149468779564, + 0.01476630475372076, + 0.010121109895408154, + 0.005802626255899668, + 0.005187585949897766, + 0.005765787325799465, + 0.006146506872028112, + 0.005817411467432976, + 0.0058158524334430695, + 0.008038964122533798, + 0.004794401116669178, + 0.0026083572302013636, + 0.0027945502661168575, + 0.002835396444424987, + 0.0036835221108049154, + 0.005389885976910591, + 0.00546899251639843, + 0.005291012581437826, + 0.004180694930255413, + 0.0030944233294576406, + 0.00330935837700963, + 0.003077350091189146, + 0.003496852470561862, + 0.007474903017282486, + 0.00704449275508523, + 0.012535064481198788, + 0.006901899352669716, + 0.004669800400733948, + 0.004318717401474714, + 0.014487635344266891, + 0.005788417533040047, + 0.004693036898970604, + 0.0035878976341336966, + 0.005350285675376654, + 0.005434371531009674, + 0.004509286489337683, + 0.004912855103611946, + 0.013305243104696274, + 0.006363630760461092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.3930879533290863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011112703010439873 + ], + "opposite_logprob_tokens": [ + -4.511112689971924 + ], + "logprob_reward": -0.011112703010439873, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011112703010439873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21704207360744476, + 0.22911354899406433, + 0.2069045454263687, + 0.1693003922700882, + 0.049119748175144196, + 0.022629542276263237, + 0.031627681106328964, + 0.020836567506194115, + 0.026865733787417412, + 0.026560217142105103 + ], + "credit_map": [ + 0.018233994022011757, + 0.022629795596003532, + 0.028035730123519897, + 0.03158260136842728, + 0.026585659012198448, + 0.02524550072848797, + 0.03631162270903587, + 0.028417158871889114, + 0.020606141537427902, + 0.015825383365154266, + 0.03265058249235153, + 0.025217102840542793, + 0.023284519091248512, + 0.03132119029760361, + 0.030716370791196823, + 0.0494922399520874, + 0.022514265030622482, + 0.017770029604434967, + 0.03627936542034149, + 0.041354384273290634, + 0.014697343111038208, + 0.019583459943532944, + 0.028004838153719902, + 0.02670084685087204, + 0.022444231435656548, + 0.018568605184555054, + 0.038171425461769104, + 0.03079862706363201, + 0.014493288472294807, + 0.01249640341848135, + 0.018121540546417236, + 0.014206264168024063, + 0.007131937425583601, + 0.0053901467472314835, + 0.009633434005081654, + 0.009251766838133335, + 0.003953418228775263, + 0.003994686529040337, + 0.004727692808955908, + 0.005036664195358753, + 0.0033129665534943342, + 0.0022129688877612352, + 0.005625579971820116, + 0.002967970911413431, + 0.0017010141164064407, + 0.001827591797336936, + 0.0027297246269881725, + 0.0022517237812280655, + 0.004197757691144943, + 0.003244970925152302, + 0.0047806245274841785, + 0.0034433151595294476, + 0.0031880380120128393, + 0.0026833058800548315, + 0.006447337567806244, + 0.003642328316345811, + 0.002982343314215541, + 0.002323694294318557, + 0.0035054157488048077, + 0.003150166943669319, + 0.0016717704711481929, + 0.0017809710698202252, + 0.003381638787686825, + 0.002040567109361291, + 0.0037460392341017723, + 0.003061151597648859, + 0.006277186330407858, + 0.002705383114516735, + 0.002569671021774411, + 0.0022628307342529297, + 0.004131762310862541, + 0.002111708512529731, + 0.003803675528615713, + 0.0037581666838377714, + 0.004430337343364954, + 0.0026046365965157747, + 0.003078691428527236, + 0.002585390582680702, + 0.004014081321656704, + 0.002285236492753029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.011112703010439873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420646533370018 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.01420646533370018, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420646533370018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14600691199302673, + 0.11241549998521805, + 0.07997601479291916, + 0.15432418882846832, + 0.10441860556602478, + 0.057699114084243774, + 0.043256908655166626, + 0.04109331965446472, + 0.024425774812698364, + 0.2363835722208023 + ], + "credit_map": [ + 0.009086288511753082, + 0.017286840826272964, + 0.02068164572119713, + 0.02623039111495018, + 0.01565053127706051, + 0.021530264988541603, + 0.017597746104002, + 0.01794322580099106, + 0.01051133032888174, + 0.01117015816271305, + 0.013074362650513649, + 0.016372593119740486, + 0.015955807641148567, + 0.012844324111938477, + 0.019725391641259193, + 0.01276154350489378, + 0.009745168499648571, + 0.01073506474494934, + 0.009711423888802528, + 0.012379907071590424, + 0.006468380801379681, + 0.009092537686228752, + 0.009307190775871277, + 0.012536345049738884, + 0.022112982347607613, + 0.02328764833509922, + 0.03592301905155182, + 0.021981552243232727, + 0.013779602013528347, + 0.011431409977376461, + 0.012280542403459549, + 0.013527445495128632, + 0.016040682792663574, + 0.01581060327589512, + 0.019284948706626892, + 0.017625384032726288, + 0.008907326497137547, + 0.007933595217764378, + 0.008955792523920536, + 0.009860283695161343, + 0.007621621713042259, + 0.005719571840018034, + 0.013505255803465843, + 0.006689171772450209, + 0.004762524738907814, + 0.004689068999141455, + 0.00956442765891552, + 0.0051474799402058125, + 0.006749333348125219, + 0.0063581690192222595, + 0.006590507458895445, + 0.005148991476744413, + 0.004153482615947723, + 0.0031289865728467703, + 0.007007143925875425, + 0.004120299592614174, + 0.005506162531673908, + 0.0038046028930693865, + 0.007271272595971823, + 0.005864012986421585, + 0.00433742068707943, + 0.003453044453635812, + 0.00647209445014596, + 0.004384717438369989, + 0.004006647504866123, + 0.003717531682923436, + 0.004551374353468418, + 0.002875552512705326, + 0.0029828199185431004, + 0.0019126860424876213, + 0.0026439970824867487, + 0.0017351680435240269, + 0.022854501381516457, + 0.0175534226000309, + 0.031733281910419464, + 0.024258527904748917, + 0.022857261821627617, + 0.02373933419585228, + 0.06626138091087341, + 0.02712589129805565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.01420646533370018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054975397884845734 + ], + "opposite_logprob_tokens": [ + -2.9299755096435547 + ], + "logprob_reward": -0.054975397884845734, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054975397884845734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12057198584079742, + 0.13018612563610077, + 0.19696299731731415, + 0.0869285836815834, + 0.188007652759552, + 0.08098422735929489, + 0.062496036291122437, + 0.059463150799274445, + 0.03290121257305145, + 0.041497983038425446 + ], + "credit_map": [ + 0.008552695624530315, + 0.012061700224876404, + 0.01589823141694069, + 0.022393275052309036, + 0.01173512451350689, + 0.017303353175520897, + 0.015588179230690002, + 0.017039436846971512, + 0.01221088133752346, + 0.012780970893800259, + 0.016334909945726395, + 0.0248001366853714, + 0.0160844586789608, + 0.01920933835208416, + 0.01674623042345047, + 0.012019192799925804, + 0.02226020209491253, + 0.016052333638072014, + 0.031926095485687256, + 0.04867423325777054, + 0.01816837303340435, + 0.016687599942088127, + 0.024577589705586433, + 0.018616579473018646, + 0.010820597410202026, + 0.009799336083233356, + 0.02000238560140133, + 0.016509607434272766, + 0.007713766768574715, + 0.005936078727245331, + 0.009533043019473553, + 0.006613766308873892, + 0.028857076540589333, + 0.036437515169382095, + 0.02721155621111393, + 0.027536114677786827, + 0.019265275448560715, + 0.013737517409026623, + 0.020325271412730217, + 0.014637320302426815, + 0.013322738930583, + 0.013248050585389137, + 0.011828472837805748, + 0.012118318118155003, + 0.00871819257736206, + 0.005759525578469038, + 0.010244346223771572, + 0.005744579713791609, + 0.00991581566631794, + 0.009792606346309185, + 0.010544775985181332, + 0.008458147756755352, + 0.006052649579942226, + 0.005802360363304615, + 0.0059382114559412, + 0.005991475190967321, + 0.008230999112129211, + 0.006524086929857731, + 0.011702769435942173, + 0.007672247942537069, + 0.005492730997502804, + 0.005480491556227207, + 0.008392809890210629, + 0.005967011675238609, + 0.004571369383484125, + 0.004463067278265953, + 0.00507015734910965, + 0.004397510550916195, + 0.0030935201793909073, + 0.003357961541041732, + 0.004908875096589327, + 0.0030387514270842075, + 0.006265691481530666, + 0.004673101007938385, + 0.0060525513254106045, + 0.005486497189849615, + 0.004603820387274027, + 0.0034988755360245705, + 0.007036328781396151, + 0.0038811159320175648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.054975397884845734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029798900708556175 + ], + "opposite_logprob_tokens": [ + -3.529798984527588 + ], + "logprob_reward": -0.029798900708556175, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029798900708556175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101110816001892, + 0.19349399209022522, + 0.11593307554721832, + 0.15465840697288513, + 0.13988375663757324, + 0.04533938318490982, + 0.040946513414382935, + 0.03621045872569084, + 0.01860155165195465, + 0.04482172802090645 + ], + "credit_map": [ + 0.013323883526027203, + 0.026607418432831764, + 0.02563193067908287, + 0.03477469086647034, + 0.026111822575330734, + 0.019830770790576935, + 0.033356595784425735, + 0.030473971739411354, + 0.015586036257445812, + 0.015812261030077934, + 0.027284516021609306, + 0.02958161197602749, + 0.023622356355190277, + 0.02454138733446598, + 0.024023424834012985, + 0.03304239362478256, + 0.012925563380122185, + 0.010878311470150948, + 0.024099163711071014, + 0.02837933413684368, + 0.006522964686155319, + 0.00873676035553217, + 0.012833230197429657, + 0.011557749472558498, + 0.019423160701990128, + 0.01898961141705513, + 0.030211053788661957, + 0.031076012179255486, + 0.013918490149080753, + 0.011340606957674026, + 0.016660010442137718, + 0.013039474375545979, + 0.021056722849607468, + 0.023418670520186424, + 0.01964448392391205, + 0.02767925336956978, + 0.012418815866112709, + 0.011344820261001587, + 0.012102894484996796, + 0.012218110263347626, + 0.006340202875435352, + 0.004308599513024092, + 0.007887247018516064, + 0.005321505945175886, + 0.004052001517266035, + 0.004916047677397728, + 0.007379763759672642, + 0.005134014878422022, + 0.006098659243434668, + 0.004914002493023872, + 0.00764611829072237, + 0.005675237625837326, + 0.004005045630037785, + 0.0030700063798576593, + 0.005685838870704174, + 0.003851603716611862, + 0.005285135004669428, + 0.004014528822153807, + 0.006720155011862516, + 0.004903794266283512, + 0.003169051371514797, + 0.003385458840057254, + 0.004574045073240995, + 0.0041582933627069, + 0.0029470594599843025, + 0.003296856302767992, + 0.0034208467695862055, + 0.0025732028298079967, + 0.0014014043845236301, + 0.0016789850778877735, + 0.0018373861676082015, + 0.0014458117075264454, + 0.004747195169329643, + 0.003859825897961855, + 0.005897145718336105, + 0.004325097892433405, + 0.0037771491333842278, + 0.003880059812217951, + 0.012555073015391827, + 0.00578018045052886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.029798900708556175 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle shot begins, the camera slowly and steadily tilting upward from the heart of Venezuela's dense, emerald jungle, following the breathtaking descent of Angel Falls. A powerful, relentless stream of white, foaming water cascades down from the flat-topped summit of a massive, ancient cliff face. The water cascades relentlessly down the sheer, weathered rock face, a tapestry of dark gray and ochre covered with damp green moss. As the water plummets through the air, it atomizes into a vast, billowing cloud of fine mist, catching the bright, direct sunlight and creating delicate, shimmering rainbows near the base. The massive waterfall is surrounded by lush, vibrant tropical vegetation on the lower portion of the cliff, forming a dense canopy below, providing a stark, beautiful contrast between life and stone. Above, the scene is crowned by a vast, clear, intense azure sky, completely devoid of cloud, amplifying the spectacular spectacle. The entire scene is filmed on an epic scale, encapsulating the raw, untouched and awe-inspiring beauty of this world-famous landmark, creating an atmosphere that is both serene and incredibly powerful.", + "seed": 14009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_3082_seed_14009_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": -0.1367040024138987, + "vlm_logprob": -0.1367040024138987, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.1367040024138987, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5810372233390808 + ], + "opposite_logprob_tokens": [ + -0.8310372233390808 + ], + "logprob_reward": -0.5810372233390808, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5810372233390808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14905239641666412, + 0.0723256766796112, + 0.047323886305093765, + 0.04025072976946831, + 0.09498948603868484, + 0.03434685617685318, + 0.14679019153118134, + 0.023603662848472595, + 0.055643558502197266, + 0.33567357063293457 + ], + "credit_map": [ + 0.022961728274822235, + 0.012422532774508, + 0.028138374909758568, + 0.01295908261090517, + 0.01762363873422146, + 0.01426446158438921, + 0.019111067056655884, + 0.021571515128016472, + 0.00816021766513586, + 0.004980366677045822, + 0.008982708677649498, + 0.006775875575840473, + 0.00462908111512661, + 0.009644330479204655, + 0.01819913648068905, + 0.010953963734209538, + 0.0051696086302399635, + 0.003208502195775509, + 0.0064947581849992275, + 0.0035056504420936108, + 0.0032658802811056376, + 0.01246489118784666, + 0.00523874768987298, + 0.007975850254297256, + 0.007202707231044769, + 0.0032833197619765997, + 0.006277901120483875, + 0.002758279675617814, + 0.00283288536593318, + 0.006439828313887119, + 0.0050802333280444145, + 0.006375574506819248, + 0.01393210794776678, + 0.01668633706867695, + 0.01205526478588581, + 0.005677407141774893, + 0.009389668703079224, + 0.010607842355966568, + 0.012694994919002056, + 0.013945858925580978, + 0.004895487800240517, + 0.0040184431709349155, + 0.006026309449225664, + 0.002163191558793187, + 0.002623774576932192, + 0.004957615397870541, + 0.00416216067969799, + 0.005499872379004955, + 0.016423894092440605, + 0.016473136842250824, + 0.01566113717854023, + 0.017834773287177086, + 0.010046105831861496, + 0.019431428983807564, + 0.029651029035449028, + 0.021268680691719055, + 0.0030496353283524513, + 0.0019236905500292778, + 0.0033246909733861685, + 0.002491509076207876, + 0.00195496273227036, + 0.002995196031406522, + 0.003830553963780403, + 0.004033427219837904, + 0.005405509378761053, + 0.005404660012573004, + 0.0076085845939815044, + 0.005339027848094702, + 0.005766470450907946, + 0.006959374528378248, + 0.01055824849754572, + 0.00860168319195509, + 0.047054264694452286, + 0.03906907141208649, + 0.05201149359345436, + 0.044003136456012726, + 0.032418664544820786, + 0.03401828184723854, + 0.040490441024303436, + 0.04660820960998535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.5810372233390808 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005288302432745695 + ], + "opposite_logprob_tokens": [ + -5.255288124084473 + ], + "logprob_reward": -0.005288302432745695, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005288302432745695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2658700942993164, + 0.13757829368114471, + 0.096281036734581, + 0.15225441753864288, + 0.06684935092926025, + 0.03919531777501106, + 0.03522244468331337, + 0.02857215330004692, + 0.1002446860074997, + 0.07793216407299042 + ], + "credit_map": [ + 0.03193734958767891, + 0.022170305252075195, + 0.03412742540240288, + 0.03328046202659607, + 0.032013654708862305, + 0.046874988824129105, + 0.031953733414411545, + 0.033512182533741, + 0.014877960085868835, + 0.009128003381192684, + 0.013793081045150757, + 0.010258986614644527, + 0.012103201821446419, + 0.03287791833281517, + 0.023011790588498116, + 0.021527355536818504, + 0.013271893374621868, + 0.012224304489791393, + 0.01619277149438858, + 0.008152981288731098, + 0.006605619564652443, + 0.017032640054821968, + 0.008211708627641201, + 0.014589112251996994, + 0.020631859079003334, + 0.021611299365758896, + 0.020864088088274002, + 0.01048431359231472, + 0.010888234712183475, + 0.02922416478395462, + 0.01686304621398449, + 0.021687421947717667, + 0.010100753046572208, + 0.008124522864818573, + 0.00869302824139595, + 0.0042805434204638, + 0.005055584479123354, + 0.008838539011776447, + 0.010623540729284286, + 0.01113283820450306, + 0.0052437856793403625, + 0.005029078107327223, + 0.006028787232935429, + 0.0031395365949720144, + 0.0024169008247554302, + 0.006043375935405493, + 0.005103896837681532, + 0.0061899591237306595, + 0.004336620215326548, + 0.002886053640395403, + 0.00375159177929163, + 0.003808405250310898, + 0.0025837006978690624, + 0.004437020514160395, + 0.007657712325453758, + 0.005761341191828251, + 0.003927269484847784, + 0.0029828506521880627, + 0.004428466781973839, + 0.002985502826049924, + 0.0024429215118288994, + 0.003888761391863227, + 0.0041783591732382774, + 0.003738020546734333, + 0.0111332843080163, + 0.01772182062268257, + 0.00945686362683773, + 0.00868638139218092, + 0.013384408317506313, + 0.010641111060976982, + 0.016726922243833542, + 0.012493899092078209, + 0.009857557713985443, + 0.006385134067386389, + 0.010466738604009151, + 0.011079005897045135, + 0.006023555528372526, + 0.007662512362003326, + 0.015414372086524963, + 0.011043285951018333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.005288302432745695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075758263468742 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.016075758263468742, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075758263468742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19223864376544952, + 0.09470069408416748, + 0.0979943796992302, + 0.10031285136938095, + 0.06997090578079224, + 0.09303673356771469, + 0.09763229638338089, + 0.056396108120679855, + 0.13360564410686493, + 0.06411172449588776 + ], + "credit_map": [ + 0.03117893449962139, + 0.013602430932223797, + 0.02617652714252472, + 0.018689148128032684, + 0.014294646680355072, + 0.03573039919137955, + 0.024877561256289482, + 0.027688998728990555, + 0.01083359494805336, + 0.005068450700491667, + 0.012096578255295753, + 0.005751978140324354, + 0.007558755110949278, + 0.021660365164279938, + 0.014495881274342537, + 0.01723508909344673, + 0.008282982744276524, + 0.004307290073484182, + 0.008870067074894905, + 0.00711796572431922, + 0.007682797499001026, + 0.030367417261004448, + 0.013587689027190208, + 0.017778171226382256, + 0.012596913613379002, + 0.01176446583122015, + 0.012231582775712013, + 0.0068082259967923164, + 0.008686775341629982, + 0.01936321146786213, + 0.014725243672728539, + 0.014136438257992268, + 0.0072274478152394295, + 0.013383530080318451, + 0.010408857837319374, + 0.005046863108873367, + 0.005641624331474304, + 0.00923596415668726, + 0.008681073784828186, + 0.01034554187208414, + 0.010363171808421612, + 0.007302634883671999, + 0.012532522901892662, + 0.007520774379372597, + 0.010051478631794453, + 0.017765015363693237, + 0.01341151725500822, + 0.014089612290263176, + 0.011293145827949047, + 0.009857099503278732, + 0.015011344105005264, + 0.008164779283106327, + 0.007805950939655304, + 0.011385013349354267, + 0.019818508997559547, + 0.014296445995569229, + 0.007674789056181908, + 0.004749517887830734, + 0.00895505491644144, + 0.004900816828012466, + 0.005482864566147327, + 0.009788489900529385, + 0.007760952692478895, + 0.0070836227387189865, + 0.012870759703218937, + 0.02646167017519474, + 0.012087724171578884, + 0.01091080904006958, + 0.020258665084838867, + 0.014909341000020504, + 0.02015296183526516, + 0.01595372147858143, + 0.0071705784648656845, + 0.007607628125697374, + 0.0105177853256464, + 0.010230974294245243, + 0.0052631087601184845, + 0.0064467135816812515, + 0.0085673239082098, + 0.008307614363729954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.016075758263468742 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003915309906006 + ], + "opposite_logprob_tokens": [ + -2.6950392723083496 + ], + "logprob_reward": -0.07003915309906006, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003915309906006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11542963236570358, + 0.07323452085256577, + 0.06779249757528305, + 0.06172904372215271, + 0.14057590067386627, + 0.14701387286186218, + 0.040606576949357986, + 0.04453557729721069, + 0.12285906076431274, + 0.1862233430147171 + ], + "credit_map": [ + 0.014134114608168602, + 0.009037825278937817, + 0.018619447946548462, + 0.008426929824054241, + 0.010048418305814266, + 0.023740503937005997, + 0.012719055637717247, + 0.01870332844555378, + 0.004464111756533384, + 0.0047189341858029366, + 0.00903038028627634, + 0.0054430426098406315, + 0.0058691417798399925, + 0.015713388100266457, + 0.015930719673633575, + 0.012064806185662746, + 0.010050229728221893, + 0.0066962591372430325, + 0.009675481356680393, + 0.005600509699434042, + 0.005293453577905893, + 0.012032482773065567, + 0.008165073581039906, + 0.010279005393385887, + 0.008453967049717903, + 0.00871512945741415, + 0.008115471340715885, + 0.0036748091224581003, + 0.004691458772867918, + 0.008953525684773922, + 0.009314309805631638, + 0.009810367599129677, + 0.018296321853995323, + 0.026503438130021095, + 0.022189544513821602, + 0.0104045569896698, + 0.010339937172830105, + 0.014645528979599476, + 0.018467584624886513, + 0.01972898840904236, + 0.021435420960187912, + 0.026315325871109962, + 0.02430659346282482, + 0.012599684298038483, + 0.008903591893613338, + 0.015334024094045162, + 0.01830422692000866, + 0.019814999774098396, + 0.005417613312602043, + 0.004361771512776613, + 0.004501889925450087, + 0.0037092596758157015, + 0.003086464712396264, + 0.006237954832613468, + 0.0077109322883188725, + 0.0055806939490139484, + 0.006021083798259497, + 0.004663357511162758, + 0.007761095184832811, + 0.003980944864451885, + 0.0037161807995289564, + 0.006012150086462498, + 0.006068232003599405, + 0.0063125332817435265, + 0.01242529321461916, + 0.026579681783914566, + 0.010440011508762836, + 0.009401165880262852, + 0.021559610962867737, + 0.012124244123697281, + 0.014924091286957264, + 0.015404967591166496, + 0.022199075669050217, + 0.023873964324593544, + 0.027011072263121605, + 0.023166269063949585, + 0.01833222061395645, + 0.01650780253112316, + 0.031195297837257385, + 0.023937642574310303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.07003915309906006 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1345008909702301, + 0.055372267961502075, + 0.13788963854312897, + 0.07192053645849228, + 0.23090296983718872, + 0.09056120365858078, + 0.0321684330701828, + 0.030959835276007652, + 0.09993600100278854, + 0.11578827351331711 + ], + "credit_map": [ + 0.023983115330338478, + 0.011277866549789906, + 0.01861373707652092, + 0.011402782052755356, + 0.007790449541062117, + 0.023706743493676186, + 0.01631230115890503, + 0.02141389064490795, + 0.005844897124916315, + 0.005346452817320824, + 0.005295126233249903, + 0.005747015587985516, + 0.00603571766987443, + 0.01088967639952898, + 0.00816887803375721, + 0.008044498972594738, + 0.018454497680068016, + 0.014230725355446339, + 0.014707885682582855, + 0.01242069061845541, + 0.009672881104052067, + 0.030083583667874336, + 0.017078042030334473, + 0.02124132588505745, + 0.008235516026616096, + 0.012548842467367649, + 0.0076584783382713795, + 0.007620964664965868, + 0.004832619801163673, + 0.011365093290805817, + 0.009753894992172718, + 0.00990512315183878, + 0.0291815847158432, + 0.051795706152915955, + 0.030225668102502823, + 0.018906380981206894, + 0.02050716057419777, + 0.0234301146119833, + 0.026190871372818947, + 0.030665472149848938, + 0.011289771646261215, + 0.021553292870521545, + 0.010552366264164448, + 0.006939890328794718, + 0.007210553623735905, + 0.010986138135194778, + 0.009996906854212284, + 0.012032276950776577, + 0.004647227935492992, + 0.0031302999705076218, + 0.004716621246188879, + 0.0028583158273249865, + 0.0036021030973643064, + 0.004416621755808592, + 0.004565616138279438, + 0.00423162616789341, + 0.003833792870864272, + 0.0031258147209882736, + 0.005182086024433374, + 0.0021868962794542313, + 0.0027307213749736547, + 0.005094578955322504, + 0.004299884662032127, + 0.0045060585252940655, + 0.009058281779289246, + 0.012941750697791576, + 0.011057254858314991, + 0.009483486413955688, + 0.010187330655753613, + 0.013283117674291134, + 0.020679935812950134, + 0.013244840316474438, + 0.014861015602946281, + 0.01311077456921339, + 0.017514053732156754, + 0.0150438342243433, + 0.011635280214250088, + 0.011854276992380619, + 0.014873678795993328, + 0.01689535565674305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.011079574935138226 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle shot begins, the camera slowly and steadily tilting upward from the heart of Venezuela's dense, emerald jungle, following the breathtaking descent of Angel Falls. A powerful, relentless stream of white, foaming water cascades down from the flat-topped summit of a massive, ancient cliff face. The water cascades relentlessly down the sheer, weathered rock face, a tapestry of dark gray and ochre covered with damp green moss. As the water plummets through the air, it atomizes into a vast, billowing cloud of fine mist, catching the bright, direct sunlight and creating delicate, shimmering rainbows near the base. The massive waterfall is surrounded by lush, vibrant tropical vegetation on the lower portion of the cliff, forming a dense canopy below, providing a stark, beautiful contrast between life and stone. Above, the scene is crowned by a vast, clear, intense azure sky, completely devoid of cloud, amplifying the spectacular spectacle. The entire scene is filmed on an epic scale, encapsulating the raw, untouched and awe-inspiring beauty of this world-famous landmark, creating an atmosphere that is both serene and incredibly powerful.", + "seed": 14010, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_3082_seed_14010_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": -0.20240913033485414, + "vlm_logprob": -0.20240913033485414, + "vlm_yes_no_margin": 2.874999976158142, + "vlm_token_credit": -0.20240913033485414, + "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\": \"Does the camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8320444226264954 + ], + "opposite_logprob_tokens": [ + -0.5820444226264954 + ], + "logprob_reward": -0.8320444226264954, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8320444226264954, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1921636015176773, + 0.11717718839645386, + 0.08893328160047531, + 0.1033533439040184, + 0.1523270606994629, + 0.08592996001243591, + 0.11437029391527176, + 0.037413835525512695, + 0.05255395174026489, + 0.05577753111720085 + ], + "credit_map": [ + 0.022123200818896294, + 0.01570907048881054, + 0.032840266823768616, + 0.019807523116469383, + 0.02939422056078911, + 0.02671915665268898, + 0.01857161335647106, + 0.026998549699783325, + 0.013523447327315807, + 0.01022670604288578, + 0.022440828382968903, + 0.014792059548199177, + 0.012385484762489796, + 0.013621695339679718, + 0.013840672560036182, + 0.016346290707588196, + 0.013044477440416813, + 0.012373460456728935, + 0.016768457368016243, + 0.009414403699338436, + 0.007996119558811188, + 0.00947361346334219, + 0.008766799233853817, + 0.011095952242612839, + 0.014984618872404099, + 0.01448120642453432, + 0.012179415673017502, + 0.010392040014266968, + 0.016934391111135483, + 0.011293034069240093, + 0.015088408254086971, + 0.008000231347978115, + 0.019802197813987732, + 0.012879237532615662, + 0.029327215626835823, + 0.01601802185177803, + 0.018489841371774673, + 0.015579059720039368, + 0.023327570408582687, + 0.016903899610042572, + 0.012445086613297462, + 0.008859531953930855, + 0.0169792789965868, + 0.007171573583036661, + 0.007223361171782017, + 0.011319547891616821, + 0.011432290077209473, + 0.010499294847249985, + 0.01874706894159317, + 0.012355401180684566, + 0.02331974357366562, + 0.013951048254966736, + 0.012212354689836502, + 0.007392550818622112, + 0.014945297501981258, + 0.01144682802259922, + 0.00574159761890769, + 0.003518318058922887, + 0.00934283621609211, + 0.003820090787485242, + 0.0029383397195488214, + 0.0037652861792594194, + 0.0040559121407568455, + 0.004231455270200968, + 0.008867830969393253, + 0.006072626914829016, + 0.008779531344771385, + 0.006915457546710968, + 0.006912927608937025, + 0.004520909395068884, + 0.005976544227451086, + 0.004508121404796839, + 0.0067025963217020035, + 0.006500525865703821, + 0.00969196017831564, + 0.008448377251625061, + 0.00487505504861474, + 0.003802584484219551, + 0.011631276458501816, + 0.004125156439840794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.8320444226264954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005987564101815224 + ], + "opposite_logprob_tokens": [ + -5.130987644195557 + ], + "logprob_reward": -0.005987564101815224, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005987564101815224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3623856008052826, + 0.11279893666505814, + 0.16289471089839935, + 0.06769468635320663, + 0.03185764327645302, + 0.026118215173482895, + 0.03750291466712952, + 0.047724392265081406, + 0.06579164415597916, + 0.08523128926753998 + ], + "credit_map": [ + 0.056625500321388245, + 0.03498435020446777, + 0.054739538580179214, + 0.03484949842095375, + 0.04965171217918396, + 0.04897035285830498, + 0.03685721755027771, + 0.045707426965236664, + 0.01154738012701273, + 0.009709697216749191, + 0.02728853188455105, + 0.012020297348499298, + 0.01085678394883871, + 0.012164624407887459, + 0.013638176023960114, + 0.01557344850152731, + 0.01523922011256218, + 0.01053075771778822, + 0.044630248099565506, + 0.009901400655508041, + 0.017629709094762802, + 0.013304343447089195, + 0.016168920323252678, + 0.03549012541770935, + 0.008849546313285828, + 0.006261932663619518, + 0.012737714685499668, + 0.005839514546096325, + 0.008725079707801342, + 0.00864922720938921, + 0.00922627467662096, + 0.007405392825603485, + 0.003579834708943963, + 0.0034817245323210955, + 0.004455916117876768, + 0.003928687423467636, + 0.004459870047867298, + 0.004642978776246309, + 0.0036675904411822557, + 0.0036410423927009106, + 0.00394087890163064, + 0.003643742995336652, + 0.003103553084656596, + 0.0032981110271066427, + 0.0024972602259367704, + 0.0032432503066956997, + 0.0032186678145080805, + 0.0031727508176118135, + 0.005926915444433689, + 0.004172979388386011, + 0.0068897344172000885, + 0.004776752088218927, + 0.003771924413740635, + 0.0031603421084582806, + 0.004295422229915857, + 0.004508845508098602, + 0.0102675911039114, + 0.006240855436772108, + 0.009201321750879288, + 0.006340017542243004, + 0.0033082470763474703, + 0.0033776736818253994, + 0.0030063868034631014, + 0.005982299335300922, + 0.014977953396737576, + 0.007634260226041079, + 0.01720028929412365, + 0.007145819254219532, + 0.0036066737957298756, + 0.003945622593164444, + 0.0052563841454684734, + 0.006024638656526804, + 0.01692991890013218, + 0.00817810744047165, + 0.024151531979441643, + 0.009707500226795673, + 0.004598855972290039, + 0.004548187367618084, + 0.011881120502948761, + 0.005236074328422546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005987564101815224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354322209954262 + ], + "opposite_logprob_tokens": [ + -3.6513543128967285 + ], + "logprob_reward": -0.026354322209954262, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354322209954262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14767950773239136, + 0.1584816873073578, + 0.11068271100521088, + 0.056941188871860504, + 0.05117716267704964, + 0.023307835683226585, + 0.05189024657011032, + 0.04530615359544754, + 0.10329626500606537, + 0.25123727321624756 + ], + "credit_map": [ + 0.01987982913851738, + 0.013724376447498798, + 0.030284985899925232, + 0.014994459226727486, + 0.011731227859854698, + 0.018542874604463577, + 0.016997549682855606, + 0.021524205803871155, + 0.015274072997272015, + 0.014292090199887753, + 0.032675597816705704, + 0.022364040836691856, + 0.0099839698523283, + 0.01672074943780899, + 0.02188652753829956, + 0.02528463862836361, + 0.010504438541829586, + 0.01070530153810978, + 0.01969490759074688, + 0.011334537528455257, + 0.016124943271279335, + 0.016908925026655197, + 0.01121987123042345, + 0.014189781621098518, + 0.007237507961690426, + 0.005840413272380829, + 0.006818772293627262, + 0.006232334766536951, + 0.010363392531871796, + 0.0053782700560987, + 0.010625074617564678, + 0.004445425234735012, + 0.005937146954238415, + 0.007775557693094015, + 0.009151987731456757, + 0.006395108066499233, + 0.0047539700753986835, + 0.006621495354920626, + 0.0046378858387470245, + 0.005904013756662607, + 0.004150937311351299, + 0.0026537503581494093, + 0.002678650664165616, + 0.0032488706056028605, + 0.0024290618021041155, + 0.002328117610886693, + 0.0030891571659594774, + 0.002729290397837758, + 0.007234489545226097, + 0.005311134736984968, + 0.008020127192139626, + 0.005106799770146608, + 0.0038653805386275053, + 0.004708461929112673, + 0.010505869053304195, + 0.00713797751814127, + 0.007456217426806688, + 0.007575566880404949, + 0.008836571127176285, + 0.006616293452680111, + 0.003478011582046747, + 0.003204836742952466, + 0.003833962604403496, + 0.004304694943130016, + 0.01216461043804884, + 0.010337433777749538, + 0.013481376692652702, + 0.012342439964413643, + 0.0124591663479805, + 0.007427960634231567, + 0.017448220402002335, + 0.017635054886341095, + 0.05782162770628929, + 0.03835183382034302, + 0.06464710831642151, + 0.02925708331167698, + 0.014556621201336384, + 0.013833972625434399, + 0.019349558278918266, + 0.013419492170214653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.026354322209954262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707598507404327 + ], + "opposite_logprob_tokens": [ + -2.1270759105682373 + ], + "logprob_reward": -0.12707598507404327, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12707598507404327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10828344523906708, + 0.3767600953578949, + 0.2577206790447235, + 0.029490208253264427, + 0.02381850965321064, + 0.01304642017930746, + 0.024989565834403038, + 0.06433112174272537, + 0.04067282751202583, + 0.06088712811470032 + ], + "credit_map": [ + 0.01722525805234909, + 0.009446831420063972, + 0.016878871247172356, + 0.009993425570428371, + 0.01800156570971012, + 0.014322280883789062, + 0.009235120378434658, + 0.013180098496377468, + 0.03465127944946289, + 0.02612537518143654, + 0.05631732568144798, + 0.07073090970516205, + 0.03677868843078613, + 0.032372064888477325, + 0.07903864979743958, + 0.0407458059489727, + 0.023469896987080574, + 0.00985398143529892, + 0.09313225746154785, + 0.0117613160982728, + 0.013454580679535866, + 0.022373750805854797, + 0.01326847542077303, + 0.07040642946958542, + 0.004455632995814085, + 0.003240646328777075, + 0.005293374415487051, + 0.003234175033867359, + 0.003911048173904419, + 0.002971045905724168, + 0.003358438378199935, + 0.00302584795281291, + 0.0042122527956962585, + 0.0025200846139341593, + 0.004461086820811033, + 0.0022933720611035824, + 0.002170971594750881, + 0.0027254626620560884, + 0.002478779759258032, + 0.002956500742584467, + 0.001992303878068924, + 0.0017518275417387486, + 0.0020405096001923084, + 0.0014382833614945412, + 0.0011857905192300677, + 0.0010965605033561587, + 0.0018419248517602682, + 0.0016992196906358004, + 0.0035816894378513098, + 0.0023570179473608732, + 0.004437166731804609, + 0.0028892436530441046, + 0.002419230295345187, + 0.00204316433519125, + 0.004339529667049646, + 0.0029225219041109085, + 0.0074250600300729275, + 0.011508580297231674, + 0.007215940859168768, + 0.012708277441561222, + 0.004480529110878706, + 0.006385012995451689, + 0.005678113084286451, + 0.008929606527090073, + 0.005129961296916008, + 0.004275613930076361, + 0.004712734837085009, + 0.0037246751599013805, + 0.0065276953391730785, + 0.004313715733587742, + 0.006060780957341194, + 0.005927647929638624, + 0.008343128487467766, + 0.00675341347232461, + 0.012591826729476452, + 0.010681857354938984, + 0.00442644814029336, + 0.003209885209798813, + 0.010518373921513557, + 0.0043621975928545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.12707598507404327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058335766196251 + ], + "opposite_logprob_tokens": [ + -3.895583391189575 + ], + "logprob_reward": -0.02058335766196251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058335766196251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21662862598896027, + 0.42456138134002686, + 0.09633324295282364, + 0.03831475228071213, + 0.037053804844617844, + 0.02321772649884224, + 0.042488835752010345, + 0.06139833480119705, + 0.044855840504169464, + 0.015147465281188488 + ], + "credit_map": [ + 0.01999390497803688, + 0.021221354603767395, + 0.04000723361968994, + 0.02305617742240429, + 0.02345699444413185, + 0.035842712968587875, + 0.021487601101398468, + 0.031562644988298416, + 0.0484464094042778, + 0.030321864411234856, + 0.060984235256910324, + 0.07603654265403748, + 0.045737527310848236, + 0.03435932844877243, + 0.08532936871051788, + 0.0433460995554924, + 0.009341498836874962, + 0.007748800329864025, + 0.020210541784763336, + 0.008299596607685089, + 0.01135601382702589, + 0.014552430249750614, + 0.009064552374184132, + 0.01575980894267559, + 0.004840807057917118, + 0.00439018290489912, + 0.0050680008716881275, + 0.003195475321263075, + 0.006615275517106056, + 0.0047970665618777275, + 0.005825356114655733, + 0.0035825895611196756, + 0.004196896217763424, + 0.0038040350191295147, + 0.0067881266586482525, + 0.004100276157259941, + 0.003434236627072096, + 0.00581754744052887, + 0.0040895831771194935, + 0.004823100287467241, + 0.003197978250682354, + 0.002014999045059085, + 0.004285349044948816, + 0.002509176265448332, + 0.0021310842130333185, + 0.0020064767450094223, + 0.004723425023257732, + 0.002349236747249961, + 0.005451026372611523, + 0.004627293907105923, + 0.0047168065793812275, + 0.0051073734648525715, + 0.004269234836101532, + 0.0044020796194672585, + 0.008609099313616753, + 0.00530591793358326, + 0.015414980240166187, + 0.004963947460055351, + 0.017981786280870438, + 0.0063963523134589195, + 0.0035017766058444977, + 0.0032713112886995077, + 0.005077556241303682, + 0.0047906264662742615, + 0.011624033562839031, + 0.0030250349082052708, + 0.013203972950577736, + 0.00496752979233861, + 0.002740928204730153, + 0.002241789596155286, + 0.003966434393078089, + 0.0030861173290759325, + 0.00323755806311965, + 0.0019270676421001554, + 0.0032460500951856375, + 0.0017379616620019078, + 0.0010979874059557915, + 0.001038972637616098, + 0.001403158763423562, + 0.0014587098266929388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02058335766196251 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "majestic low-angle shot begins, the camera slowly and steadily tilting upward from the heart of Venezuela's dense, emerald jungle, following the breathtaking descent of Angel Falls. A powerful, relentless stream of white, foaming water cascades down from the flat-topped summit of a massive, ancient cliff face. The water cascades relentlessly down the sheer, weathered rock face, a tapestry of dark gray and ochre covered with damp green moss. As the water plummets through the air, it atomizes into a vast, billowing cloud of fine mist, catching the bright, direct sunlight and creating delicate, shimmering rainbows near the base. The massive waterfall is surrounded by lush, vibrant tropical vegetation on the lower portion of the cliff, forming a dense canopy below, providing a stark, beautiful contrast between life and stone. Above, the scene is crowned by a vast, clear, intense azure sky, completely devoid of cloud, amplifying the spectacular spectacle. The entire scene is filmed on an epic scale, encapsulating the raw, untouched and awe-inspiring beauty of this world-famous landmark, creating an atmosphere that is both serene and incredibly powerful.", + "seed": 14011, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_15/index_3082_seed_14011_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": -0.13990863990038632, + "vlm_logprob": -0.13990863990038632, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.13990863990038632, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5810335278511047 + ], + "opposite_logprob_tokens": [ + -0.8310335278511047 + ], + "logprob_reward": -0.5810335278511047, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5810335278511047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1335289180278778, + 0.06743770092725754, + 0.12312283366918564, + 0.10544287413358688, + 0.040360819548368454, + 0.03642459213733673, + 0.19290369749069214, + 0.14606769382953644, + 0.11577846854925156, + 0.03893241658806801 + ], + "credit_map": [ + 0.022083092480897903, + 0.020057713612914085, + 0.01909501664340496, + 0.014222923666238785, + 0.012319009751081467, + 0.01596062257885933, + 0.01584390364587307, + 0.01394663192331791, + 0.009211757220327854, + 0.008858706802129745, + 0.007515733130276203, + 0.007415858097374439, + 0.004384882282465696, + 0.010923828929662704, + 0.010255160741508007, + 0.008871774189174175, + 0.013224045746028423, + 0.00875681359320879, + 0.012006375007331371, + 0.018594348803162575, + 0.012877336703240871, + 0.016339436173439026, + 0.027292998507618904, + 0.014031480066478252, + 0.010994188487529755, + 0.006356971804052591, + 0.012951957061886787, + 0.0181572325527668, + 0.007171688601374626, + 0.013242239132523537, + 0.024536971002817154, + 0.012031620368361473, + 0.0035347244702279568, + 0.004382838029414415, + 0.005452739540487528, + 0.004212943837046623, + 0.0037866965867578983, + 0.008777358569204807, + 0.004272004589438438, + 0.0059415134601294994, + 0.0034673199988901615, + 0.0037720997352153063, + 0.003992302808910608, + 0.004853476770222187, + 0.005113441497087479, + 0.005502758547663689, + 0.005217254627496004, + 0.004505939781665802, + 0.014713824726641178, + 0.01834685355424881, + 0.01919454149901867, + 0.03240804746747017, + 0.021459044888615608, + 0.021159591153264046, + 0.044354405254125595, + 0.021267404779791832, + 0.04503228887915611, + 0.04465394839644432, + 0.015221749432384968, + 0.008066774345934391, + 0.00913915317505598, + 0.007733626756817102, + 0.00874696858227253, + 0.007473192643374205, + 0.021214665845036507, + 0.024715030565857887, + 0.01485387422144413, + 0.009585636667907238, + 0.008066599257290363, + 0.013746080920100212, + 0.013500554487109184, + 0.010096028447151184, + 0.004355342593044043, + 0.004438153933733702, + 0.004364566877484322, + 0.005458243191242218, + 0.00467942887917161, + 0.004323586355894804, + 0.007126523181796074, + 0.004186570644378662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.5810335278511047 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012578204274177551 + ], + "opposite_logprob_tokens": [ + -4.387578010559082 + ], + "logprob_reward": -0.012578204274177551, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012578204274177551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11886651068925858, + 0.1672523468732834, + 0.08666951954364777, + 0.1007007583975792, + 0.026038141921162605, + 0.07722301781177521, + 0.025811558589339256, + 0.1592039316892624, + 0.10477708280086517, + 0.13345718383789062 + ], + "credit_map": [ + 0.01495302189141512, + 0.012780449353158474, + 0.012818937189877033, + 0.01587812416255474, + 0.00817402359098196, + 0.024265490472316742, + 0.01430345419794321, + 0.01569300703704357, + 0.01775788515806198, + 0.01524688582867384, + 0.018449770286679268, + 0.020711101591587067, + 0.011265158653259277, + 0.03013242408633232, + 0.030936656519770622, + 0.02275245077908039, + 0.009740564040839672, + 0.007456048391759396, + 0.008057921193540096, + 0.014262516051530838, + 0.005557055119425058, + 0.013515369966626167, + 0.017864136025309563, + 0.010215893387794495, + 0.010333145037293434, + 0.00722141657024622, + 0.011594153940677643, + 0.018813857808709145, + 0.009186569601297379, + 0.011372123844921589, + 0.020725728943943977, + 0.011453752405941486, + 0.0029604558367282152, + 0.0026678673457354307, + 0.0034843224566429853, + 0.004172553773969412, + 0.0016769487410783768, + 0.002689710119739175, + 0.0054313866421580315, + 0.0029548965394496918, + 0.007613779045641422, + 0.006857855245471001, + 0.006826036609709263, + 0.010366491042077541, + 0.008292573504149914, + 0.012150773778557777, + 0.015125768259167671, + 0.009989729151129723, + 0.004178933799266815, + 0.003223076928406954, + 0.0025706009473651648, + 0.003577280091121793, + 0.0019382876344025135, + 0.003434445708990097, + 0.004268798511475325, + 0.002620133338496089, + 0.04675101488828659, + 0.04862214997410774, + 0.012368809431791306, + 0.013598739169538021, + 0.008728258311748505, + 0.008172866888344288, + 0.012082478031516075, + 0.008879599161446095, + 0.027357520535588264, + 0.02972521260380745, + 0.012929901480674744, + 0.006821938790380955, + 0.006604448426514864, + 0.008301785215735435, + 0.006208440288901329, + 0.0068278200924396515, + 0.014147806912660599, + 0.015146847814321518, + 0.022046033293008804, + 0.016273338347673416, + 0.009859099984169006, + 0.018573109060525894, + 0.022265207022428513, + 0.015145732089877129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.012578204274177551 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203409664332867 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203409664332867, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203409664332867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1683105081319809, + 0.06101740151643753, + 0.26339617371559143, + 0.1036900207400322, + 0.10666057467460632, + 0.03150692954659462, + 0.09890973567962646, + 0.09791592508554459, + 0.031407274305820465, + 0.0371854342520237 + ], + "credit_map": [ + 0.02446833625435829, + 0.01711386628448963, + 0.03778166323900223, + 0.0183532927185297, + 0.01206300500780344, + 0.022788822650909424, + 0.019254131242632866, + 0.016487373039126396, + 0.007766196504235268, + 0.0065893675200641155, + 0.009227585047483444, + 0.006551723461598158, + 0.00664029736071825, + 0.00812778901308775, + 0.008812081068754196, + 0.007302357815206051, + 0.024646546691656113, + 0.021498525515198708, + 0.024446653202176094, + 0.03196370601654053, + 0.029674947261810303, + 0.04981492832303047, + 0.04803083464503288, + 0.03332002833485603, + 0.008623161353170872, + 0.0077618760988116264, + 0.010805275291204453, + 0.01215424295514822, + 0.00977550819516182, + 0.022371230646967888, + 0.018436867743730545, + 0.013761848211288452, + 0.008346172980964184, + 0.010519974865019321, + 0.0136085981503129, + 0.007957053370773792, + 0.016042089089751244, + 0.024262798950076103, + 0.01110228430479765, + 0.014821586199104786, + 0.002425144659355283, + 0.0024020103737711906, + 0.003722286783158779, + 0.004577830899506807, + 0.0023941867984831333, + 0.0063988761976361275, + 0.005752602592110634, + 0.003833985421806574, + 0.007380262948572636, + 0.00903891958296299, + 0.009291334077715874, + 0.016347736120224, + 0.01035972498357296, + 0.013214867562055588, + 0.021458178758621216, + 0.01181870698928833, + 0.030748654156923294, + 0.031383488327264786, + 0.0067632729187607765, + 0.0052740005776286125, + 0.00601074518635869, + 0.005436474457383156, + 0.006559023167937994, + 0.005740253254771233, + 0.002732692752033472, + 0.0034784742165356874, + 0.0060898903757333755, + 0.003190227784216404, + 0.004057844635099173, + 0.004685681778937578, + 0.003653795225545764, + 0.0035186666063964367, + 0.0049325223080813885, + 0.005782198626548052, + 0.002406098647043109, + 0.0039184256456792355, + 0.0035461827646940947, + 0.005838973447680473, + 0.005905716679990292, + 0.0048553128726780415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.014203409664332867 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866454377770424 + ], + "opposite_logprob_tokens": [ + -3.0486645698547363 + ], + "logprob_reward": -0.04866454377770424, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866454377770424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1004938930273056, + 0.16421687602996826, + 0.35529783368110657, + 0.11076230555772781, + 0.02483287826180458, + 0.03480766341090202, + 0.09238819032907486, + 0.052404262125492096, + 0.022655213251709938, + 0.04214093089103699 + ], + "credit_map": [ + 0.013798908330500126, + 0.013644850812852383, + 0.0186654981225729, + 0.013851716183125973, + 0.00715486379340291, + 0.009487168863415718, + 0.013750667683780193, + 0.010140229016542435, + 0.022621916607022285, + 0.022946946322917938, + 0.020608525723218918, + 0.01957438699901104, + 0.015245657414197922, + 0.022183874621987343, + 0.023188801482319832, + 0.017846770584583282, + 0.040442053228616714, + 0.04408338665962219, + 0.03925612196326256, + 0.04672800749540329, + 0.02106483094394207, + 0.05810600891709328, + 0.06294510513544083, + 0.042672306299209595, + 0.012343751266598701, + 0.01216951198875904, + 0.015362863428890705, + 0.01748541183769703, + 0.009150082245469093, + 0.014570731669664383, + 0.018184607848525047, + 0.011495337821543217, + 0.003093412611633539, + 0.0025553822051733732, + 0.004076663870364428, + 0.0032216403633356094, + 0.0017850506119430065, + 0.0029783251229673624, + 0.0032035047188401222, + 0.003918900154531002, + 0.00295366533100605, + 0.0032288753427565098, + 0.0038078518118709326, + 0.0035296278074383736, + 0.0045629628002643585, + 0.006106266751885414, + 0.005710253491997719, + 0.004908157512545586, + 0.00716545432806015, + 0.00898460391908884, + 0.009747013449668884, + 0.015628430992364883, + 0.009412461891770363, + 0.012876533903181553, + 0.017807424068450928, + 0.010766269639134407, + 0.013505303300917149, + 0.01457828376442194, + 0.004660709761083126, + 0.00453655980527401, + 0.0030450038611888885, + 0.003452765289694071, + 0.004093111492693424, + 0.004532523453235626, + 0.002208487596362829, + 0.0022265110164880753, + 0.003422429086640477, + 0.0033316530752927065, + 0.001863491372205317, + 0.0034612733870744705, + 0.003629719838500023, + 0.002511648926883936, + 0.007065183017402887, + 0.008098727092146873, + 0.005798727739602327, + 0.003673386760056019, + 0.005262976512312889, + 0.004107391461730003, + 0.004784571938216686, + 0.0033499691635370255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.04866454377770424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a low-angle shot and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the waterfall water white and foaming as it cascades down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cliff face dark gray and ochre with damp green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky above vast, clear, and intense azure without any clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are delicate, shimmering rainbows visible in the mist near the base of the falls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063513934612274 + ], + "opposite_logprob_tokens": [ + -3.1680634021759033 + ], + "logprob_reward": -0.043063513934612274, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063513934612274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25644007325172424, + 0.07139340788125992, + 0.2977714538574219, + 0.05927589535713196, + 0.050437141209840775, + 0.059383928775787354, + 0.05388253182172775, + 0.04907175153493881, + 0.05798957869410515, + 0.044354211539030075 + ], + "credit_map": [ + 0.03792927414178848, + 0.03402305766940117, + 0.046454884111881256, + 0.0313250757753849, + 0.016146404668688774, + 0.032108549028635025, + 0.03290240839123726, + 0.025550441816449165, + 0.009670689702033997, + 0.009350518696010113, + 0.010625751689076424, + 0.00866298470646143, + 0.007110829930752516, + 0.008196542970836163, + 0.009246684610843658, + 0.00852940883487463, + 0.03410136327147484, + 0.026638446375727654, + 0.028664981946349144, + 0.0483400784432888, + 0.017513656988739967, + 0.04384696111083031, + 0.06464917212724686, + 0.03401680663228035, + 0.005933670327067375, + 0.004738492425531149, + 0.007847904227674007, + 0.00699904840439558, + 0.005201689433306456, + 0.010017536580562592, + 0.010814935900270939, + 0.007722625508904457, + 0.004658077843487263, + 0.004729609936475754, + 0.007219667546451092, + 0.004300908651202917, + 0.008098679594695568, + 0.0083086509257555, + 0.00664322217926383, + 0.006478330586105585, + 0.00623677670955658, + 0.005709410645067692, + 0.006457589101046324, + 0.010944944806396961, + 0.003544087056070566, + 0.007083548232913017, + 0.01263244729489088, + 0.0067751300521194935, + 0.008670237846672535, + 0.007415665779262781, + 0.005047231446951628, + 0.007598825264722109, + 0.004578019492328167, + 0.006589527241885662, + 0.008258099667727947, + 0.00572493439540267, + 0.01278605218976736, + 0.012305662035942078, + 0.005696968175470829, + 0.0035904180258512497, + 0.0029468750581145287, + 0.003923715092241764, + 0.004205085802823305, + 0.003616977483034134, + 0.008786304853856564, + 0.009758508764207363, + 0.006618687883019447, + 0.005509004462510347, + 0.005854318849742413, + 0.006175053771585226, + 0.008719686418771744, + 0.0065680211409926414, + 0.0041590239852666855, + 0.004848260432481766, + 0.006967059802263975, + 0.005784718319773674, + 0.005195914302021265, + 0.0041200341656804085, + 0.008633370511233807, + 0.00464583607390523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.043063513934612274 + } + ], + "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/0000022/group_2/index_tensor([3771, 3771])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_tensor([3771, 3771])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d0614a18479028a7ba9b09fc93b7e203bbbeec72 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_tensor([3771, 3771])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium shot captures a dormant robot, slumped lifelessly across a heavy diagnostic chair. Its humanoid form, composed of a combination of polished white polymer panels and exposed gunmetal-gray mechanical joints, is completely devoid of tension. Its head tilts forward, its chin resting almost on its chest. Unlit optical sensors give its face a blank, serene expression. Its arms hang limply, fingers slightly curled. From a circular, translucent panel in the center of its chest, a soft, ethereal blue light pulsates with a slow, steady rhythm, like the beat of a sleeping heart. This rhythmic glow, the sole source of significant illumination, casts dynamic, soft-edged highlights across the robot's metal surface and immerses the surrounding cluttered studio in deep, shifting shadows. The overall atmosphere is one of profound stillness and contemplative quiet, a moment of suspended animation within a high-tech sanctuary.", + "seed": 13325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13325_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": -0.021231857873499394, + "vlm_logprob": -0.021231857873499394, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.021231857873499394, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03300538286566734 + ], + "opposite_logprob_tokens": [ + -3.533005475997925 + ], + "logprob_reward": -0.03300538286566734, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03300538286566734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29190561175346375, + 0.07890983670949936, + 0.057619526982307434, + 0.06263939291238785, + 0.1308121681213379, + 0.1024925708770752, + 0.08216627687215805, + 0.049243975430727005, + 0.043495263904333115, + 0.10071537643671036 + ], + "credit_map": [ + 0.049336571246385574, + 0.0277857705950737, + 0.05674704536795616, + 0.025216372683644295, + 0.02334032766520977, + 0.04868941009044647, + 0.02656281739473343, + 0.034227292984724045, + 0.014975915662944317, + 0.008350072428584099, + 0.014876204542815685, + 0.0072332751005887985, + 0.005295862909406424, + 0.009720702655613422, + 0.008652238175272942, + 0.009805561043322086, + 0.008865952491760254, + 0.0058676702901721, + 0.012343540787696838, + 0.006478520110249519, + 0.004240020178258419, + 0.005739794112741947, + 0.007366711273789406, + 0.006717319134622812, + 0.012365709990262985, + 0.005249641370028257, + 0.01514545176178217, + 0.006640616338700056, + 0.004050721414387226, + 0.008636711165308952, + 0.00593674136325717, + 0.004613807424902916, + 0.02658367156982422, + 0.015036518685519695, + 0.030435968190431595, + 0.014401428401470184, + 0.010623667389154434, + 0.01775514893233776, + 0.007410149555653334, + 0.008565625175833702, + 0.018155645579099655, + 0.015549384988844395, + 0.014939436689019203, + 0.010041077621281147, + 0.010611388832330704, + 0.012178538367152214, + 0.01052737794816494, + 0.010489719919860363, + 0.014421162195503712, + 0.012673107907176018, + 0.012338812462985516, + 0.008717451244592667, + 0.009415804408490658, + 0.009029662236571312, + 0.007963540963828564, + 0.007606733124703169, + 0.009106919169425964, + 0.004905004519969225, + 0.012997445650398731, + 0.0024643626529723406, + 0.004178010392934084, + 0.007732033729553223, + 0.003452169941738248, + 0.004408029839396477, + 0.0072836000472307205, + 0.003903799457475543, + 0.01230823714286089, + 0.002398270182311535, + 0.0025227877777069807, + 0.008344392292201519, + 0.0033638174645602703, + 0.0033703583758324385, + 0.015573158860206604, + 0.014395658858120441, + 0.01767808385193348, + 0.008670689538121223, + 0.012586154974997044, + 0.01322224736213684, + 0.009158539585769176, + 0.009430847130715847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.03300538286566734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023312414065003395 + ], + "opposite_logprob_tokens": [ + -3.7733123302459717 + ], + "logprob_reward": -0.023312414065003395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023312414065003395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21025224030017853, + 0.19713464379310608, + 0.05681806057691574, + 0.10500389337539673, + 0.030590256676077843, + 0.1339203119277954, + 0.0698612853884697, + 0.04496559500694275, + 0.08110026270151138, + 0.0703534409403801 + ], + "credit_map": [ + 0.047104887664318085, + 0.021203091368079185, + 0.04363968223333359, + 0.016560090705752373, + 0.02179848775267601, + 0.03735831379890442, + 0.010123836807906628, + 0.012463868595659733, + 0.0460730716586113, + 0.02127886191010475, + 0.03382537141442299, + 0.022991454228758812, + 0.01942388154566288, + 0.026276035234332085, + 0.015897346660494804, + 0.011368632316589355, + 0.0074035124853253365, + 0.005682009272277355, + 0.008370610885322094, + 0.0048678722232580185, + 0.005724603775888681, + 0.00602931622415781, + 0.010620110668241978, + 0.00812002457678318, + 0.015124497003853321, + 0.016536299139261246, + 0.014785672537982464, + 0.011117055080831051, + 0.011981175281107426, + 0.012202858924865723, + 0.011738292872905731, + 0.011518040671944618, + 0.004013093188405037, + 0.002164145465940237, + 0.005135803483426571, + 0.003119109198451042, + 0.00307605997659266, + 0.003069785190746188, + 0.006305295508354902, + 0.0037069644313305616, + 0.025435740128159523, + 0.019965821877121925, + 0.019782986491918564, + 0.014199189841747284, + 0.013610131107270718, + 0.01561394426971674, + 0.013010450638830662, + 0.012302041985094547, + 0.014009146951138973, + 0.0069214883260428905, + 0.015382313169538975, + 0.007466278970241547, + 0.005520249251276255, + 0.010467760264873505, + 0.005513410083949566, + 0.004580637440085411, + 0.008405396714806557, + 0.004518741276115179, + 0.00807227659970522, + 0.004050536546856165, + 0.004757953807711601, + 0.006154659204185009, + 0.004564468748867512, + 0.004441564437001944, + 0.012385894544422626, + 0.009970331564545631, + 0.014092911966145039, + 0.009921003133058548, + 0.009911812841892242, + 0.011373994871973991, + 0.007441444788128138, + 0.006002866663038731, + 0.0097120376303792, + 0.01099967397749424, + 0.011292885057628155, + 0.007296291179955006, + 0.009374944493174553, + 0.010003278031945229, + 0.00565637880936265, + 0.006017953157424927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.023312414065003395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082286946475506 + ], + "opposite_logprob_tokens": [ + -4.511082172393799 + ], + "logprob_reward": -0.011082286946475506, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082286946475506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18665076792240143, + 0.14619514346122742, + 0.07075510174036026, + 0.040100693702697754, + 0.13220369815826416, + 0.11889791488647461, + 0.0585949569940567, + 0.12071705609560013, + 0.08554063737392426, + 0.04034409299492836 + ], + "credit_map": [ + 0.03383173048496246, + 0.0187132116407156, + 0.039925478398799896, + 0.019547121599316597, + 0.014900918118655682, + 0.031119761988520622, + 0.011190381832420826, + 0.01742214895784855, + 0.020710133016109467, + 0.019337987527251244, + 0.02146615833044052, + 0.022353392094373703, + 0.015436431393027306, + 0.017470525577664375, + 0.015446173958480358, + 0.013974330388009548, + 0.009716501459479332, + 0.006946150213479996, + 0.012515885755419731, + 0.008876580744981766, + 0.007688858546316624, + 0.008716591633856297, + 0.008482378907501698, + 0.007812144700437784, + 0.006535614840686321, + 0.004533014725893736, + 0.0057924664579331875, + 0.003733729012310505, + 0.005429155193269253, + 0.005754062905907631, + 0.003912251442670822, + 0.0044103930704295635, + 0.02504434809088707, + 0.019335202872753143, + 0.018181288614869118, + 0.014601301401853561, + 0.015124203637242317, + 0.014311092905700207, + 0.013560337014496326, + 0.012045912444591522, + 0.022829342633485794, + 0.017160311341285706, + 0.017531193792819977, + 0.013725044205784798, + 0.011934204958379269, + 0.013145125471055508, + 0.01145082712173462, + 0.011121849529445171, + 0.011741179041564465, + 0.005040773656219244, + 0.013331891968846321, + 0.004330209922045469, + 0.003933526575565338, + 0.010372824966907501, + 0.0054321326315402985, + 0.00441241217777133, + 0.019425859674811363, + 0.016980571672320366, + 0.020988425239920616, + 0.012436038814485073, + 0.010279637761414051, + 0.01700248196721077, + 0.012165941298007965, + 0.011438085697591305, + 0.011930218897759914, + 0.010357830673456192, + 0.016638170927762985, + 0.010849103331565857, + 0.00961447786539793, + 0.013102385215461254, + 0.0070816148072481155, + 0.0059668272733688354, + 0.0057017314247787, + 0.0068725221790373325, + 0.006305342074483633, + 0.003989541437476873, + 0.005242951214313507, + 0.00537137733772397, + 0.0037557827308773994, + 0.0031048390083014965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011082286946475506 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057693526148796 + ], + "opposite_logprob_tokens": [ + -3.8955769538879395 + ], + "logprob_reward": -0.02057693526148796, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057693526148796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13225685060024261, + 0.0484590083360672, + 0.09258496016263962, + 0.36761850118637085, + 0.022701483219861984, + 0.0954078808426857, + 0.037816621363162994, + 0.1523084044456482, + 0.01852172426879406, + 0.032324548810720444 + ], + "credit_map": [ + 0.01856015995144844, + 0.016376031562685966, + 0.03400849550962448, + 0.00850732158869505, + 0.008070968091487885, + 0.022395992651581764, + 0.010886747390031815, + 0.013451140373945236, + 0.007269683293998241, + 0.006868205964565277, + 0.011104795150458813, + 0.004671731498092413, + 0.003364184405654669, + 0.006726565305143595, + 0.004095208831131458, + 0.00435863621532917, + 0.013839729130268097, + 0.015857651829719543, + 0.012211988680064678, + 0.010611905716359615, + 0.010862977243959904, + 0.01152317225933075, + 0.007990393787622452, + 0.0096871517598629, + 0.06196139007806778, + 0.05905123054981232, + 0.050297629088163376, + 0.04539689049124718, + 0.04006245359778404, + 0.04205009713768959, + 0.03337834030389786, + 0.0354204885661602, + 0.004766808822751045, + 0.0019394690170884132, + 0.0051436349749565125, + 0.0016521214274689555, + 0.0018103154143318534, + 0.003423948073759675, + 0.001868530292995274, + 0.0020966592710465193, + 0.01704220287501812, + 0.014854232780635357, + 0.013596443459391594, + 0.01048586517572403, + 0.010670798830688, + 0.010996920987963676, + 0.008841781876981258, + 0.008919649757444859, + 0.005850539542734623, + 0.005711413454264402, + 0.0055564758367836475, + 0.005338594317436218, + 0.003859652439132333, + 0.004583186004310846, + 0.0040328470058739185, + 0.002883918583393097, + 0.026527833193540573, + 0.0211021825671196, + 0.024283790960907936, + 0.01603691466152668, + 0.01664179563522339, + 0.01948712393641472, + 0.01478895265609026, + 0.013439844362437725, + 0.0026687695644795895, + 0.0022488944232463837, + 0.0033130019437521696, + 0.0019645749125629663, + 0.0020869835279881954, + 0.002661324106156826, + 0.0021019966807216406, + 0.0014761817874386907, + 0.004096117336302996, + 0.004647025838494301, + 0.007086113095283508, + 0.0029708838555961847, + 0.0036753311287611723, + 0.005590394604951143, + 0.002195642562583089, + 0.00206304551102221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.02057693526148796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2344939261674881, + 0.06237128749489784, + 0.06066395342350006, + 0.08108298480510712, + 0.06629126518964767, + 0.09680401533842087, + 0.07642517983913422, + 0.0573689267039299, + 0.14072780311107635, + 0.1237705871462822 + ], + "credit_map": [ + 0.04483412206172943, + 0.02549130655825138, + 0.05088786035776138, + 0.016442520543932915, + 0.02025233954191208, + 0.03348890691995621, + 0.019574753940105438, + 0.023522114381194115, + 0.009169299155473709, + 0.004859492648392916, + 0.014755618758499622, + 0.0048286197707057, + 0.004066814202815294, + 0.009596138261258602, + 0.006281661801040173, + 0.008813641965389252, + 0.010988643392920494, + 0.005899270996451378, + 0.01312022004276514, + 0.006061290856450796, + 0.006544571369886398, + 0.007996627129614353, + 0.005042317323386669, + 0.005011012777686119, + 0.01596567966043949, + 0.006591287441551685, + 0.020312849432229996, + 0.008796267211437225, + 0.006910883821547031, + 0.011674599722027779, + 0.0062386090867221355, + 0.004592811688780785, + 0.010923725552856922, + 0.006133300717920065, + 0.016748806461691856, + 0.006860034540295601, + 0.0053795380517840385, + 0.010325581766664982, + 0.0043700821697711945, + 0.00555019173771143, + 0.01609731651842594, + 0.015331367962062359, + 0.014191127382218838, + 0.010465112514793873, + 0.009400810115039349, + 0.012194688431918621, + 0.009677143767476082, + 0.009446450509130955, + 0.017622079700231552, + 0.008307396434247494, + 0.016758300364017487, + 0.006035253405570984, + 0.004977746866643429, + 0.011524992994964123, + 0.004488019272685051, + 0.006711391266435385, + 0.011031046509742737, + 0.005262281280010939, + 0.01361914537847042, + 0.005358412396162748, + 0.003365308977663517, + 0.00890841893851757, + 0.004909655544906855, + 0.004914658144116402, + 0.023628011345863342, + 0.01908675767481327, + 0.02381720580160618, + 0.015551132149994373, + 0.016627084463834763, + 0.018254848197102547, + 0.0120433010160923, + 0.011719463393092155, + 0.018205618485808372, + 0.018997805193066597, + 0.01910724677145481, + 0.01192776020616293, + 0.01611929200589657, + 0.018525496125221252, + 0.010197458788752556, + 0.010689912363886833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.018182270228862762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a dormant robot, slumped lifelessly across a heavy diagnostic chair. Its humanoid form, composed of a combination of polished white polymer panels and exposed gunmetal-gray mechanical joints, is completely devoid of tension. Its head tilts forward, its chin resting almost on its chest. Unlit optical sensors give its face a blank, serene expression. Its arms hang limply, fingers slightly curled. From a circular, translucent panel in the center of its chest, a soft, ethereal blue light pulsates with a slow, steady rhythm, like the beat of a sleeping heart. This rhythmic glow, the sole source of significant illumination, casts dynamic, soft-edged highlights across the robot's metal surface and immerses the surrounding cluttered studio in deep, shifting shadows. The overall atmosphere is one of profound stillness and contemplative quiet, a moment of suspended animation within a high-tech sanctuary.", + "seed": 13326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13326_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": -0.02371291946619749, + "vlm_logprob": -0.02371291946619749, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.02371291946619749, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026931216940283775 + ], + "opposite_logprob_tokens": [ + -3.7769312858581543 + ], + "logprob_reward": -0.026931216940283775, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026931216940283775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15282605588436127, + 0.19233857095241547, + 0.06184574216604233, + 0.027567915618419647, + 0.1369493156671524, + 0.06723722815513611, + 0.24071574211120605, + 0.031122229993343353, + 0.046494536101818085, + 0.042902637273073196 + ], + "credit_map": [ + 0.02268070913851261, + 0.02398175559937954, + 0.017552684992551804, + 0.015855491161346436, + 0.026512861251831055, + 0.0230714101344347, + 0.01383829303085804, + 0.009332871064543724, + 0.02684517577290535, + 0.019235366955399513, + 0.011866580694913864, + 0.015709297731518745, + 0.04980139806866646, + 0.03779822215437889, + 0.021026473492383957, + 0.010056065395474434, + 0.0072165196761488914, + 0.007495309691876173, + 0.00399658503010869, + 0.004772363696247339, + 0.015734659507870674, + 0.012012689374387264, + 0.007152280770242214, + 0.003465342102572322, + 0.004569093696773052, + 0.006045933347195387, + 0.002677389420568943, + 0.0020716823637485504, + 0.003530060639604926, + 0.004654265474528074, + 0.0024618373718112707, + 0.0015576527221128345, + 0.017577312886714935, + 0.014165293425321579, + 0.008426811546087265, + 0.009576676413416862, + 0.037469472736120224, + 0.027314189821481705, + 0.015721051022410393, + 0.006698503624647856, + 0.009963691234588623, + 0.006742970552295446, + 0.005162099841982126, + 0.004033569246530533, + 0.017555566504597664, + 0.012968067079782486, + 0.007271778304129839, + 0.003539490280672908, + 0.030535826459527016, + 0.02297384850680828, + 0.015791038051247597, + 0.015479200519621372, + 0.06791067868471146, + 0.04876551777124405, + 0.02737562358379364, + 0.01188402809202671, + 0.005039176903665066, + 0.0053888531401753426, + 0.003186695044860244, + 0.002915808465331793, + 0.004711028654128313, + 0.004982989747077227, + 0.0025427842047065496, + 0.002354896394535899, + 0.006613926030695438, + 0.004692082293331623, + 0.003363365773111582, + 0.0025261505506932735, + 0.012128619477152824, + 0.009196736849844456, + 0.005441172048449516, + 0.002532488200813532, + 0.006255842745304108, + 0.0058599552139639854, + 0.0037368489429354668, + 0.0031638348009437323, + 0.009864598512649536, + 0.007729456759989262, + 0.004044818691909313, + 0.0022472830023616552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.026931216940283775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209896981716156 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018209896981716156, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209896981716156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1625407338142395, + 0.15331974625587463, + 0.11484503000974655, + 0.11579802632331848, + 0.270549476146698, + 0.015287482179701328, + 0.04802035540342331, + 0.060417432337999344, + 0.018796196207404137, + 0.04042549803853035 + ], + "credit_map": [ + 0.021235330030322075, + 0.01722424291074276, + 0.010333078913390636, + 0.010588787496089935, + 0.04361354187130928, + 0.03269076347351074, + 0.017550978809595108, + 0.00930403545498848, + 0.0229120422154665, + 0.014816872775554657, + 0.009172511287033558, + 0.015792623162269592, + 0.03737615421414375, + 0.02887805365025997, + 0.01643170788884163, + 0.007939797826111317, + 0.01602320373058319, + 0.010813661850988865, + 0.006907846312969923, + 0.00868916790932417, + 0.03133649751543999, + 0.02287452481687069, + 0.012532838620245457, + 0.005667284596711397, + 0.015787962824106216, + 0.010505110956728458, + 0.006665621884167194, + 0.008039556443691254, + 0.03243636339902878, + 0.023462535813450813, + 0.013153919018805027, + 0.005746970418840647, + 0.0354049988090992, + 0.025713251903653145, + 0.016631554812192917, + 0.017947599291801453, + 0.07543617486953735, + 0.054761990904808044, + 0.031121574342250824, + 0.013532341457903385, + 0.0029659608844667673, + 0.0015855497913435102, + 0.0014083769638091326, + 0.0017413771711289883, + 0.0027741866651922464, + 0.0027081191074103117, + 0.0011163323651999235, + 0.0009875806281343102, + 0.006580563727766275, + 0.004483829252421856, + 0.003165994770824909, + 0.0032493467442691326, + 0.013303508050739765, + 0.009591609239578247, + 0.005369912832975388, + 0.0022755989339202642, + 0.009889435954391956, + 0.006303226109594107, + 0.003455233061686158, + 0.005296220071613789, + 0.014598678797483444, + 0.011384417302906513, + 0.006290682591497898, + 0.003199539612978697, + 0.0023404413368552923, + 0.0019025324145331979, + 0.0011233763070777059, + 0.0013836310245096684, + 0.004998339340090752, + 0.0037442552857100964, + 0.0022988112177699804, + 0.0010048120748251677, + 0.0053508952260017395, + 0.004437021445482969, + 0.002984675345942378, + 0.0029050041921436787, + 0.010659274645149708, + 0.0078076864592731, + 0.0042466288432478905, + 0.0020343137439340353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.018209896981716156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792386554181576 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792386554181576, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792386554181576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6025838851928711, + 0.07625836879014969, + 0.02484070509672165, + 0.10080859810113907, + 0.05383515730500221, + 0.07772711664438248, + 0.01988830603659153, + 0.010330541990697384, + 0.011727550998330116, + 0.021999845281243324 + ], + "credit_map": [ + 0.08468420803546906, + 0.05577964335680008, + 0.03360477834939957, + 0.04957297816872597, + 0.16202865540981293, + 0.11880478262901306, + 0.06731788069009781, + 0.030790915712714195, + 0.012332016602158546, + 0.007906720042228699, + 0.005756894126534462, + 0.007193579338490963, + 0.016977842897176743, + 0.013687772676348686, + 0.007598624099045992, + 0.004804912954568863, + 0.003763513406738639, + 0.0025198098737746477, + 0.0020964634604752064, + 0.0018905473407357931, + 0.005929113831371069, + 0.004637205507606268, + 0.0024728220887482166, + 0.0015312267933040857, + 0.013829450123012066, + 0.009267774410545826, + 0.006195459980517626, + 0.00633239047601819, + 0.028339767828583717, + 0.02042543888092041, + 0.011389227584004402, + 0.005029081832617521, + 0.0069218226708471775, + 0.004994357004761696, + 0.003601326607167721, + 0.003292759647592902, + 0.01506548747420311, + 0.010937914252281189, + 0.006081586703658104, + 0.0029399003833532333, + 0.010302050970494747, + 0.008239898830652237, + 0.005177831742912531, + 0.0056471312418580055, + 0.02069280296564102, + 0.015116493217647076, + 0.00853718537837267, + 0.004013713449239731, + 0.0025123264640569687, + 0.002043290762230754, + 0.0013811950339004397, + 0.001402402645908296, + 0.005466118920594454, + 0.003953447099775076, + 0.002210617996752262, + 0.0009189051925204694, + 0.0014810407301411033, + 0.0012880126014351845, + 0.0012898424174636602, + 0.001181157073006034, + 0.0016850337851792574, + 0.0017554651712998748, + 0.0009771628538146615, + 0.0006728267762809992, + 0.0018763748230412602, + 0.0014640173176303506, + 0.001162945874966681, + 0.0011156847467646003, + 0.002114340662956238, + 0.002120562829077244, + 0.00117771839722991, + 0.0006959042511880398, + 0.003058155532926321, + 0.002209177240729332, + 0.0017914464697241783, + 0.001441683154553175, + 0.005807243753224611, + 0.004257543478161097, + 0.002390335313975811, + 0.0010442598722875118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.009792386554181576 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305369406938553 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.04305369406938553, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305369406938553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1215626522898674, + 0.1848844438791275, + 0.024150662124156952, + 0.2839324176311493, + 0.10827475041151047, + 0.10579347610473633, + 0.033673182129859924, + 0.048985935747623444, + 0.06502574682235718, + 0.023716744035482407 + ], + "credit_map": [ + 0.01839972287416458, + 0.014592260122299194, + 0.012058920226991177, + 0.010648224502801895, + 0.024730663746595383, + 0.02156241610646248, + 0.011724406853318214, + 0.00784603413194418, + 0.027823353186249733, + 0.018344080075621605, + 0.013657032512128353, + 0.01371702179312706, + 0.0472881942987442, + 0.035148777067661285, + 0.019381342455744743, + 0.009524638764560223, + 0.0033893417567014694, + 0.0038169375620782375, + 0.0033152797259390354, + 0.0020368078257888556, + 0.0037758888211101294, + 0.0038834805600345135, + 0.0022086985409259796, + 0.0017242280300706625, + 0.039499394595623016, + 0.025435730814933777, + 0.016192803159356117, + 0.01985953189432621, + 0.07890301942825317, + 0.057243868708610535, + 0.0326576754450798, + 0.014140386134386063, + 0.01485441718250513, + 0.010591618716716766, + 0.006591076496988535, + 0.00903122778981924, + 0.028153810650110245, + 0.02106485888361931, + 0.012301838025450706, + 0.005685910116881132, + 0.018094036728143692, + 0.011854523792862892, + 0.009600629098713398, + 0.008287881501019001, + 0.0227518230676651, + 0.018128719180822372, + 0.010117457248270512, + 0.006958409678190947, + 0.005602357909083366, + 0.0051617794670164585, + 0.00372719787992537, + 0.003428339259698987, + 0.005053911358118057, + 0.0059402501210570335, + 0.0028342376463115215, + 0.0019251102348789573, + 0.0073648071847856045, + 0.0053430041298270226, + 0.0035620476119220257, + 0.0033227482344955206, + 0.012334958650171757, + 0.009233266115188599, + 0.0051966579630970955, + 0.0026284479536116123, + 0.009362513199448586, + 0.006044859066605568, + 0.003893152577802539, + 0.004186906851828098, + 0.018044231459498405, + 0.013009396381676197, + 0.007259381003677845, + 0.003225306747481227, + 0.003120172768831253, + 0.003363208379596472, + 0.002320815110579133, + 0.0019861720502376556, + 0.005195411387830973, + 0.004122974816709757, + 0.0024205157533288, + 0.001187472022138536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.04305369406938553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577402785420418 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.020577402785420418, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577402785420418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3441671133041382, + 0.03478388115763664, + 0.027891116216778755, + 0.08337744325399399, + 0.057450853288173676, + 0.07314728945493698, + 0.22773593664169312, + 0.034585773944854736, + 0.08128660172224045, + 0.035573989152908325 + ], + "credit_map": [ + 0.04760080203413963, + 0.03235236182808876, + 0.018071915954351425, + 0.030820952728390694, + 0.09099813550710678, + 0.06839016079902649, + 0.03846443444490433, + 0.017468368634581566, + 0.006055303383618593, + 0.004647735506296158, + 0.0036682630889117718, + 0.002419116673991084, + 0.007014008704572916, + 0.005726445000618696, + 0.0029375769663602114, + 0.0023154353257268667, + 0.003896172158420086, + 0.004752835258841515, + 0.0028673685155808926, + 0.0020569206681102514, + 0.0054430183954536915, + 0.004501349292695522, + 0.002503767143934965, + 0.0018696885090321302, + 0.011600710451602936, + 0.008532632142305374, + 0.005149432923644781, + 0.005656559951603413, + 0.022679053246974945, + 0.01635310985147953, + 0.009094121865928173, + 0.00431183073669672, + 0.007811611983925104, + 0.006653957534581423, + 0.004070226568728685, + 0.0043137772008776665, + 0.014594612643122673, + 0.010763131082057953, + 0.0060743228532373905, + 0.0031692234333604574, + 0.010645695962011814, + 0.0068765124306082726, + 0.004494314547628164, + 0.005590236280113459, + 0.0194002166390419, + 0.014428078196942806, + 0.008070074021816254, + 0.003642172086983919, + 0.029607180505990982, + 0.02097979560494423, + 0.014251155778765678, + 0.014361363835632801, + 0.0645434632897377, + 0.04648241773247719, + 0.02626228705048561, + 0.011248278431594372, + 0.005158652551472187, + 0.003349118400365114, + 0.0021852224599570036, + 0.002359652193263173, + 0.009209162555634975, + 0.006793767213821411, + 0.003858892945572734, + 0.0016713130753487349, + 0.010098030790686607, + 0.008122284896671772, + 0.005293878726661205, + 0.005207133013755083, + 0.022597886621952057, + 0.016475224867463112, + 0.00952177308499813, + 0.003970395773649216, + 0.006184675730764866, + 0.003539889119565487, + 0.002218155423179269, + 0.002132844179868698, + 0.00930125080049038, + 0.006901996210217476, + 0.003525776555761695, + 0.0017694002017378807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.020577402785420418 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a dormant robot, slumped lifelessly across a heavy diagnostic chair. Its humanoid form, composed of a combination of polished white polymer panels and exposed gunmetal-gray mechanical joints, is completely devoid of tension. Its head tilts forward, its chin resting almost on its chest. Unlit optical sensors give its face a blank, serene expression. Its arms hang limply, fingers slightly curled. From a circular, translucent panel in the center of its chest, a soft, ethereal blue light pulsates with a slow, steady rhythm, like the beat of a sleeping heart. This rhythmic glow, the sole source of significant illumination, casts dynamic, soft-edged highlights across the robot's metal surface and immerses the surrounding cluttered studio in deep, shifting shadows. The overall atmosphere is one of profound stillness and contemplative quiet, a moment of suspended animation within a high-tech sanctuary.", + "seed": 13327, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13327_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": -0.032570981048047544, + "vlm_logprob": -0.032570981048047544, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.032570981048047544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045261092483997345 + ], + "opposite_logprob_tokens": [ + -3.1702611446380615 + ], + "logprob_reward": -0.045261092483997345, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045261092483997345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1199822872877121, + 0.07837996631860733, + 0.055158115923404694, + 0.18242110311985016, + 0.037523481994867325, + 0.10828162729740143, + 0.0945362001657486, + 0.04363124445080757, + 0.22256231307983398, + 0.0575236976146698 + ], + "credit_map": [ + 0.014959045685827732, + 0.022509727627038956, + 0.01520050223916769, + 0.013736337423324585, + 0.014052034355700016, + 0.015232117846608162, + 0.010383746586740017, + 0.013908757828176022, + 0.007087250705808401, + 0.008132443763315678, + 0.007088040001690388, + 0.01297921221703291, + 0.007075377739965916, + 0.010881337337195873, + 0.010950576514005661, + 0.014185715466737747, + 0.00551791163161397, + 0.007505370303988457, + 0.007795719895511866, + 0.0062981597147881985, + 0.007416097912937403, + 0.006832737009972334, + 0.009230028837919235, + 0.004562081303447485, + 0.037358637899160385, + 0.01830936223268509, + 0.01034528948366642, + 0.010999169200658798, + 0.014426175504922867, + 0.01034547295421362, + 0.06722085177898407, + 0.013416117988526821, + 0.005303986836224794, + 0.004291484598070383, + 0.005112047307193279, + 0.003189608920365572, + 0.004418096039444208, + 0.004702989012002945, + 0.004752844572067261, + 0.005752417724579573, + 0.01516985148191452, + 0.01491478644311428, + 0.013939929194748402, + 0.009799836203455925, + 0.01206981111317873, + 0.015856537967920303, + 0.016777802258729935, + 0.009753060527145863, + 0.010220249183475971, + 0.010374282486736774, + 0.011700917966663837, + 0.007458420004695654, + 0.017111042514443398, + 0.008843718096613884, + 0.015814872458577156, + 0.01301268395036459, + 0.0059298621490597725, + 0.009599567390978336, + 0.004767168313264847, + 0.004014587961137295, + 0.004417272750288248, + 0.0045196907594799995, + 0.006992742419242859, + 0.003390346886590123, + 0.03421033173799515, + 0.010570207610726357, + 0.00834107119590044, + 0.010403897613286972, + 0.017056811600923538, + 0.08146645128726959, + 0.04623531177639961, + 0.014278201386332512, + 0.005083774216473103, + 0.011745484545826912, + 0.007583184167742729, + 0.007332191802561283, + 0.0072658308781683445, + 0.008639903739094734, + 0.004411079455167055, + 0.005462240893393755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.045261092483997345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007648942526429892 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007648942526429892, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007648942526429892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15682096779346466, + 0.09559264034032822, + 0.07190117239952087, + 0.2901816964149475, + 0.04283635690808296, + 0.09994757920503616, + 0.03244631737470627, + 0.047010067850351334, + 0.1274634152650833, + 0.035799797624349594 + ], + "credit_map": [ + 0.01704254560172558, + 0.019235115498304367, + 0.02245861291885376, + 0.025423241779208183, + 0.01761145517230034, + 0.022653067484498024, + 0.014293412677943707, + 0.018103521317243576, + 0.011254889890551567, + 0.014852815307676792, + 0.011790798045694828, + 0.01395939290523529, + 0.011058514937758446, + 0.014742773957550526, + 0.008757384493947029, + 0.00917606521397829, + 0.010646087117493153, + 0.005949388258159161, + 0.004806194920092821, + 0.006621551234275103, + 0.0047582597471773624, + 0.020194141194224358, + 0.00982346385717392, + 0.009102084673941135, + 0.05325717478990555, + 0.02868443727493286, + 0.030050557106733322, + 0.01800440438091755, + 0.03385421261191368, + 0.025256287306547165, + 0.08395165205001831, + 0.017122946679592133, + 0.004815076943486929, + 0.005370569881051779, + 0.00474528456106782, + 0.005759405437856913, + 0.0059959120117127895, + 0.0041746231727302074, + 0.007551305461674929, + 0.004424181766808033, + 0.016973387449979782, + 0.009970849379897118, + 0.011768249794840813, + 0.007031592540442944, + 0.008807760663330555, + 0.015383861027657986, + 0.02114141173660755, + 0.00887046754360199, + 0.004569506272673607, + 0.004016940016299486, + 0.0039360858500003815, + 0.002158966613933444, + 0.003956721164286137, + 0.003966034855693579, + 0.0075793578289449215, + 0.0022627071011811495, + 0.007342854514718056, + 0.0045286207459867, + 0.005193124059587717, + 0.004002688452601433, + 0.005120479501783848, + 0.004174634348601103, + 0.012988997623324394, + 0.0036586669739335775, + 0.020487645640969276, + 0.005184540990740061, + 0.0038457277696579695, + 0.005812080577015877, + 0.009245350025594234, + 0.047535091638565063, + 0.030447669327259064, + 0.00490531325340271, + 0.004889505449682474, + 0.006085500121116638, + 0.004252834245562553, + 0.0034402215387672186, + 0.003901739837601781, + 0.0046241749078035355, + 0.0033231116831302643, + 0.0052827103063464165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.007648942526429892 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641592036932707 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.004641592036932707, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641592036932707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15941014885902405, + 0.13467907905578613, + 0.1049891859292984, + 0.1753852218389511, + 0.04473939910531044, + 0.08287625014781952, + 0.03366779536008835, + 0.024002065882086754, + 0.2123449593782425, + 0.027905883267521858 + ], + "credit_map": [ + 0.02278442494571209, + 0.021377606317400932, + 0.013871890492737293, + 0.019752394407987595, + 0.014580019749701023, + 0.037840552628040314, + 0.013648966327309608, + 0.015554291196167469, + 0.01701287552714348, + 0.019472481682896614, + 0.014849658124148846, + 0.01557356771081686, + 0.012717673555016518, + 0.030366776511073112, + 0.011894548311829567, + 0.012791484594345093, + 0.008364198729395866, + 0.014080547727644444, + 0.007576396688818932, + 0.013131298124790192, + 0.008851420134305954, + 0.020677534863352776, + 0.025236044079065323, + 0.00707173952832818, + 0.0201041791588068, + 0.01959041878581047, + 0.04341963678598404, + 0.010786065831780434, + 0.025986334308981895, + 0.01930805668234825, + 0.021684013307094574, + 0.014506511390209198, + 0.004436649847775698, + 0.010180098004639149, + 0.004650020971894264, + 0.007040549535304308, + 0.004773746244609356, + 0.004362606443464756, + 0.004238159395754337, + 0.005057567730545998, + 0.010836758650839329, + 0.007920346222817898, + 0.015421861782670021, + 0.005426841322332621, + 0.012785321101546288, + 0.011261580511927605, + 0.011523966677486897, + 0.007699574343860149, + 0.004163105506449938, + 0.005935767199844122, + 0.004111798480153084, + 0.0029996964149177074, + 0.00469701923429966, + 0.0035125426948070526, + 0.005556412972509861, + 0.002691454952582717, + 0.0023032319732010365, + 0.00407123239710927, + 0.0023342419881373644, + 0.0023172246292233467, + 0.004162176977843046, + 0.003068581223487854, + 0.0020049852319061756, + 0.0037403921596705914, + 0.032467056065797806, + 0.00870399922132492, + 0.006527900230139494, + 0.009496674872934818, + 0.015537208877503872, + 0.07676612585783005, + 0.05520942062139511, + 0.007636583410203457, + 0.0028349782805889845, + 0.0051091923378407955, + 0.003272321308031678, + 0.002596435137093067, + 0.003663404146209359, + 0.003693997859954834, + 0.0027965218760073185, + 0.003939034417271614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.004641592036932707 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896079868078232 + ], + "opposite_logprob_tokens": [ + -2.57896089553833 + ], + "logprob_reward": -0.07896079868078232, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896079868078232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1025252640247345, + 0.07458946853876114, + 0.05792860686779022, + 0.14143246412277222, + 0.030637148767709732, + 0.07884527742862701, + 0.04676336422562599, + 0.0706326887011528, + 0.380219042301178, + 0.01642671786248684 + ], + "credit_map": [ + 0.01132187433540821, + 0.01745240017771721, + 0.01649930700659752, + 0.00860863272100687, + 0.014661520719528198, + 0.012926416471600533, + 0.010880136862397194, + 0.010174962691962719, + 0.0093134930357337, + 0.009976182132959366, + 0.014694464392960072, + 0.006787172053009272, + 0.009938633069396019, + 0.009879346005618572, + 0.005867037922143936, + 0.008133135735988617, + 0.00740957073867321, + 0.006949251517653465, + 0.008531543426215649, + 0.006571362726390362, + 0.005121347960084677, + 0.011600014753639698, + 0.004594622179865837, + 0.007150887046009302, + 0.027607666328549385, + 0.01344617921859026, + 0.012400815263390541, + 0.00883843656629324, + 0.012467151507735252, + 0.011311585083603859, + 0.046117085963487625, + 0.00924352090805769, + 0.002984959864988923, + 0.0032005785033106804, + 0.004771180916577578, + 0.003283634316176176, + 0.004978496581315994, + 0.00445580156520009, + 0.0030727186240255833, + 0.003889775136485696, + 0.008834399282932281, + 0.012553779408335686, + 0.008846301585435867, + 0.006926276255398989, + 0.006942712236195803, + 0.011199146509170532, + 0.008709714747965336, + 0.014832936227321625, + 0.007273635361343622, + 0.004868746269494295, + 0.004679245408624411, + 0.003604284953325987, + 0.005435815546661615, + 0.004977511242032051, + 0.012588072568178177, + 0.0033360517118126154, + 0.011605717241764069, + 0.010061708278954029, + 0.005525743588805199, + 0.004581170156598091, + 0.006181743927299976, + 0.0060186139307916164, + 0.021262457594275475, + 0.005395525600761175, + 0.0601760558784008, + 0.01218641921877861, + 0.010239525698125362, + 0.016998745501041412, + 0.027923839166760445, + 0.14192938804626465, + 0.09739115089178085, + 0.01337388250976801, + 0.001740761217661202, + 0.002646507928147912, + 0.002301587024703622, + 0.0019331722287461162, + 0.0024118702858686447, + 0.0021122379694133997, + 0.0017105706501752138, + 0.00157000997569412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.07896079868078232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634247951209545 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.02634247951209545, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634247951209545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11140582710504532, + 0.04655689001083374, + 0.20133505761623383, + 0.12931138277053833, + 0.021816113963723183, + 0.2700701057910919, + 0.040469735860824585, + 0.031537044793367386, + 0.06940032541751862, + 0.07809752970933914 + ], + "credit_map": [ + 0.015403048135340214, + 0.01476426050066948, + 0.013714801520109177, + 0.00969612319022417, + 0.01434303168207407, + 0.02327062003314495, + 0.01032331120222807, + 0.009890629909932613, + 0.004168996587395668, + 0.009241879917681217, + 0.003948310390114784, + 0.005107020493596792, + 0.004731706343591213, + 0.0053409733809530735, + 0.007994391955435276, + 0.006023612339049578, + 0.017353691160678864, + 0.03136303275823593, + 0.022749049589037895, + 0.03067588247358799, + 0.02073916792869568, + 0.024101562798023224, + 0.038877472281455994, + 0.015475197695195675, + 0.023780465126037598, + 0.012047848664224148, + 0.012408128008246422, + 0.00825592316687107, + 0.008502594195306301, + 0.00962868146598339, + 0.04484090209007263, + 0.009846831671893597, + 0.003180513856932521, + 0.00257485662586987, + 0.0026521256659179926, + 0.002420591190457344, + 0.002891859970986843, + 0.00278035132214427, + 0.002753053791821003, + 0.0025627610739320517, + 0.041115082800388336, + 0.023912400007247925, + 0.04802309349179268, + 0.021337931975722313, + 0.04226130619645119, + 0.032681237906217575, + 0.03988414630293846, + 0.020854927599430084, + 0.004510671366006136, + 0.006221781950443983, + 0.004804170690476894, + 0.003938623238354921, + 0.0043626753613352776, + 0.005712195765227079, + 0.004253680352121592, + 0.006665938068181276, + 0.004119349643588066, + 0.003828551387414336, + 0.003242800012230873, + 0.0020182335283607244, + 0.0038857064209878445, + 0.004530830308794975, + 0.007025032304227352, + 0.0028865409549325705, + 0.0062333629466593266, + 0.010052026249468327, + 0.006085763219743967, + 0.0054724738001823425, + 0.007530955132097006, + 0.012327172793447971, + 0.016151247546076775, + 0.0055473241955041885, + 0.007241969928145409, + 0.015393341891467571, + 0.010814891196787357, + 0.008879972621798515, + 0.010543315671384335, + 0.011132711544632912, + 0.006896553561091423, + 0.007194772828370333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.02634247951209545 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a dormant robot, slumped lifelessly across a heavy diagnostic chair. Its humanoid form, composed of a combination of polished white polymer panels and exposed gunmetal-gray mechanical joints, is completely devoid of tension. Its head tilts forward, its chin resting almost on its chest. Unlit optical sensors give its face a blank, serene expression. Its arms hang limply, fingers slightly curled. From a circular, translucent panel in the center of its chest, a soft, ethereal blue light pulsates with a slow, steady rhythm, like the beat of a sleeping heart. This rhythmic glow, the sole source of significant illumination, casts dynamic, soft-edged highlights across the robot's metal surface and immerses the surrounding cluttered studio in deep, shifting shadows. The overall atmosphere is one of profound stillness and contemplative quiet, a moment of suspended animation within a high-tech sanctuary.", + "seed": 13328, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_2/index_3771_seed_13328_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": -0.04718628264963627, + "vlm_logprob": -0.04718628264963627, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04718628264963627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024711081758141518 + ], + "opposite_logprob_tokens": [ + -3.8997111320495605 + ], + "logprob_reward": -0.024711081758141518, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024711081758141518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30263593792915344, + 0.0702701210975647, + 0.11330495029687881, + 0.21010065078735352, + 0.025477245450019836, + 0.1214364543557167, + 0.05226026475429535, + 0.03550877422094345, + 0.04755333438515663, + 0.02145226299762726 + ], + "credit_map": [ + 0.03216428682208061, + 0.05008235573768616, + 0.043195270001888275, + 0.035997506231069565, + 0.02825181744992733, + 0.041052334010601044, + 0.02621942013502121, + 0.045672956854104996, + 0.009022416546940804, + 0.01121523603796959, + 0.009410775266587734, + 0.008109157904982567, + 0.00777058769017458, + 0.011000615544617176, + 0.00735938735306263, + 0.006381944287568331, + 0.014864305034279823, + 0.018461940810084343, + 0.018407752737402916, + 0.01406139601022005, + 0.009269870817661285, + 0.0152262132614851, + 0.007642734330147505, + 0.015370743349194527, + 0.02645166404545307, + 0.03182452172040939, + 0.033633939921855927, + 0.02732376754283905, + 0.016799265518784523, + 0.029414571821689606, + 0.012838784605264664, + 0.03181413933634758, + 0.0029103599954396486, + 0.003083297051489353, + 0.0035798726603388786, + 0.0031549229752272367, + 0.0030657073948532343, + 0.0037940440233796835, + 0.0024768514558672905, + 0.0034121875651180744, + 0.01482737623155117, + 0.02079990692436695, + 0.018187789246439934, + 0.014640264213085175, + 0.010512081906199455, + 0.01668623834848404, + 0.007904977537691593, + 0.017877807840704918, + 0.005055589601397514, + 0.0065653626807034016, + 0.007342155557125807, + 0.006125680170953274, + 0.00593602005392313, + 0.008857961744070053, + 0.0042197913862764835, + 0.008157702162861824, + 0.004115310497581959, + 0.005288079846650362, + 0.005541094578802586, + 0.0037070405669510365, + 0.0037776047829538584, + 0.005578693468123674, + 0.002835927763953805, + 0.0046650199219584465, + 0.004007398150861263, + 0.006311593111604452, + 0.006642036139965057, + 0.005396039225161076, + 0.007036645896732807, + 0.009180782362818718, + 0.003598051844164729, + 0.005380786024034023, + 0.003089733887463808, + 0.003132301615551114, + 0.0031485818326473236, + 0.00233776425011456, + 0.002814199309796095, + 0.0031368250492960215, + 0.0019344384782016277, + 0.0018584186909720302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.024711081758141518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014216454699635506 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216454699635506, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216454699635506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29377129673957825, + 0.1436358243227005, + 0.1695767492055893, + 0.04756172001361847, + 0.07160042971372604, + 0.059050291776657104, + 0.03486897051334381, + 0.07885251194238663, + 0.040915004909038544, + 0.06016715615987778 + ], + "credit_map": [ + 0.040844932198524475, + 0.044738054275512695, + 0.04045884683728218, + 0.032294031232595444, + 0.03812933340668678, + 0.04492495581507683, + 0.02586604841053486, + 0.02651510015130043, + 0.01664234697818756, + 0.020412921905517578, + 0.022395925596356392, + 0.016503605991601944, + 0.014917484484612942, + 0.02397863008081913, + 0.015293776988983154, + 0.01349111832678318, + 0.016663936898112297, + 0.03195057809352875, + 0.025068603456020355, + 0.01984323002398014, + 0.017055658623576164, + 0.021087994799017906, + 0.013857284560799599, + 0.024049459025263786, + 0.005785929039120674, + 0.006116875447332859, + 0.008087055757641792, + 0.004453266505151987, + 0.006670726463198662, + 0.007906649261713028, + 0.00490784365683794, + 0.0036333741154521704, + 0.008025946095585823, + 0.009727248921990395, + 0.010751818306744099, + 0.008180319331586361, + 0.008293941617012024, + 0.011655754409730434, + 0.006223615724593401, + 0.008741791360080242, + 0.007541204337030649, + 0.010345189832150936, + 0.007686201948672533, + 0.0062113418243825436, + 0.008165526203811169, + 0.007276697084307671, + 0.004777855239808559, + 0.007046274375170469, + 0.005729141179472208, + 0.004205622244626284, + 0.006093012634664774, + 0.0032691187225282192, + 0.004512820392847061, + 0.005423036869615316, + 0.002733455738052726, + 0.0029027650598436594, + 0.009540343657135963, + 0.012948261573910713, + 0.011155031621456146, + 0.008323589339852333, + 0.007889332249760628, + 0.01067402958869934, + 0.007927956059575081, + 0.010393969714641571, + 0.00513581233099103, + 0.004902157466858625, + 0.005872807931154966, + 0.004735839087516069, + 0.004574690945446491, + 0.006576027255505323, + 0.003934216219931841, + 0.005183455068618059, + 0.008344165049493313, + 0.009050876833498478, + 0.008681279607117176, + 0.005960994865745306, + 0.007499282248318195, + 0.009352135471999645, + 0.006310043390840292, + 0.004968378227204084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.014216454699635506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607787050306797 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.01607787050306797, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607787050306797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2603135406970978, + 0.055984560400247574, + 0.09096349775791168, + 0.12563490867614746, + 0.14295393228530884, + 0.07324521243572235, + 0.06728590279817581, + 0.07417763024568558, + 0.06282778829336166, + 0.04661300405859947 + ], + "credit_map": [ + 0.032425712794065475, + 0.042774513363838196, + 0.042400769889354706, + 0.026710066944360733, + 0.0298724714666605, + 0.03540081903338432, + 0.01932186633348465, + 0.031407326459884644, + 0.006125656887888908, + 0.006640577223151922, + 0.00791411753743887, + 0.007129849400371313, + 0.007418446708470583, + 0.009208723902702332, + 0.005771464668214321, + 0.0057757250033319, + 0.010301092639565468, + 0.015135799534618855, + 0.013540598563849926, + 0.011355690658092499, + 0.009269404225051403, + 0.012917278334498405, + 0.005932648666203022, + 0.012510997243225574, + 0.018381765112280846, + 0.0179964117705822, + 0.02014858089387417, + 0.014911283738911152, + 0.01126683596521616, + 0.0172291100025177, + 0.008777308277785778, + 0.016923606395721436, + 0.016434894874691963, + 0.02524586208164692, + 0.022607270628213882, + 0.018210913985967636, + 0.010337588377296925, + 0.019963156431913376, + 0.009654615074396133, + 0.020499637350440025, + 0.009269996546208858, + 0.012466288171708584, + 0.01088685728609562, + 0.008858680725097656, + 0.0061215790919959545, + 0.009618311189115047, + 0.005934903863817453, + 0.0100886020809412, + 0.008211781270802021, + 0.011640509590506554, + 0.010288613848388195, + 0.008480285294353962, + 0.005787432659417391, + 0.00944958534091711, + 0.005163057707250118, + 0.008264640346169472, + 0.008857093751430511, + 0.011954803951084614, + 0.011349836364388466, + 0.009359936229884624, + 0.0063030654564499855, + 0.011682190001010895, + 0.005408068653196096, + 0.009262645617127419, + 0.0073701078072190285, + 0.010045977309346199, + 0.009490189142525196, + 0.007408388424664736, + 0.005644382443279028, + 0.008286066353321075, + 0.00507690105587244, + 0.00950577761977911, + 0.006449478678405285, + 0.007344551384449005, + 0.006504785269498825, + 0.004067918285727501, + 0.005476400721818209, + 0.0072143166325986385, + 0.0037696650251746178, + 0.005785895045846701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.01607787050306797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034474968910217 + ], + "opposite_logprob_tokens": [ + -1.9103447198867798 + ], + "logprob_reward": -0.16034474968910217, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034474968910217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1964820921421051, + 0.09931761026382446, + 0.049823369830846786, + 0.11826933920383453, + 0.0775633230805397, + 0.11753349006175995, + 0.08032897859811783, + 0.12178755551576614, + 0.09454934298992157, + 0.044344935566186905 + ], + "credit_map": [ + 0.025792060419917107, + 0.026929697021842003, + 0.0287962444126606, + 0.02102937176823616, + 0.028332695364952087, + 0.03219744563102722, + 0.014570890925824642, + 0.018833676353096962, + 0.013150465674698353, + 0.010334526188671589, + 0.016373667865991592, + 0.014773334376513958, + 0.010175018571317196, + 0.01726238615810871, + 0.007750381715595722, + 0.009497838094830513, + 0.007254302501678467, + 0.005018051713705063, + 0.009525824338197708, + 0.00596789363771677, + 0.005480295978486538, + 0.008749824948608875, + 0.004249463323503733, + 0.003577711759135127, + 0.016288746148347855, + 0.018054721876978874, + 0.018108181655406952, + 0.01477913185954094, + 0.00982983410358429, + 0.015450511127710342, + 0.007921183481812477, + 0.017837032675743103, + 0.008395938202738762, + 0.01172858290374279, + 0.011921362951397896, + 0.009076807647943497, + 0.006734755355864763, + 0.010987009853124619, + 0.006128213834017515, + 0.01259065605700016, + 0.015540353953838348, + 0.01789274998009205, + 0.018574155867099762, + 0.014657835476100445, + 0.00954294390976429, + 0.01690901257097721, + 0.007821418344974518, + 0.01659502647817135, + 0.01071067713201046, + 0.013961724936962128, + 0.01260544266551733, + 0.009960769675672054, + 0.006574680563062429, + 0.009548009373247623, + 0.006874448154121637, + 0.01009322702884674, + 0.015538221225142479, + 0.020831091329455376, + 0.0192047581076622, + 0.01548317726701498, + 0.010026725940406322, + 0.016450025141239166, + 0.008113457821309566, + 0.01614009216427803, + 0.012266648933291435, + 0.015483061783015728, + 0.015353092923760414, + 0.011315714567899704, + 0.007782662753015757, + 0.011791097931563854, + 0.006371605675667524, + 0.014185458421707153, + 0.005531536415219307, + 0.007977470755577087, + 0.0061362506821751595, + 0.005303733982145786, + 0.004529011435806751, + 0.006429630797356367, + 0.0042833443731069565, + 0.00415395712479949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.16034474968910217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a humanoid robot slumped lifelessly across a heavy diagnostic chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the robot's body composed of polished white polymer panels and exposed gunmetal-gray mechanical joints?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a soft, ethereal blue light pulsate rhythmically from a circular translucent panel on the robot's chest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the surrounding environment a cluttered studio illuminated by deep, shifting shadows cast by the blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot captured from a static, low-angle, medium perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2228018343448639, + 0.09579308331012726, + 0.08096515387296677, + 0.12633374333381653, + 0.10127910226583481, + 0.11427100747823715, + 0.035929448902606964, + 0.0572158619761467, + 0.10561539977788925, + 0.059795353561639786 + ], + "credit_map": [ + 0.03538969159126282, + 0.03280951827764511, + 0.034035395830869675, + 0.025794919580221176, + 0.023460794240236282, + 0.029314370825886726, + 0.018735993653535843, + 0.02326115593314171, + 0.013580739498138428, + 0.010361481457948685, + 0.017600486055016518, + 0.010088360868394375, + 0.010261801071465015, + 0.01843719743192196, + 0.008997909724712372, + 0.006465115584433079, + 0.012192622758448124, + 0.010818623006343842, + 0.014497668482363224, + 0.008084258064627647, + 0.008784344419836998, + 0.01053804624825716, + 0.008875995874404907, + 0.007173597346991301, + 0.012070502154529095, + 0.015386639162898064, + 0.018535621464252472, + 0.019200146198272705, + 0.011444319039583206, + 0.019375838339328766, + 0.010588440112769604, + 0.019732246175408363, + 0.01033664122223854, + 0.012648394331336021, + 0.01622827537357807, + 0.009717414155602455, + 0.015747543424367905, + 0.015073036774992943, + 0.011374526657164097, + 0.010153274983167648, + 0.014745019376277924, + 0.01859964244067669, + 0.01836812123656273, + 0.013104911893606186, + 0.009126179851591587, + 0.01707661710679531, + 0.007597927935421467, + 0.015652596950531006, + 0.0049803247675299644, + 0.0053735035471618176, + 0.005369252059608698, + 0.0038045994006097317, + 0.004598870407789946, + 0.004771116655319929, + 0.0033421963453292847, + 0.003689587116241455, + 0.007559630088508129, + 0.009213810786604881, + 0.009699656628072262, + 0.006164387334138155, + 0.004788634367287159, + 0.008997424505650997, + 0.004034537356346846, + 0.006757788825780153, + 0.013419696129858494, + 0.01686045154929161, + 0.01674460805952549, + 0.011691314168274403, + 0.009944083169102669, + 0.016082724556326866, + 0.007150437217205763, + 0.013722102157771587, + 0.007493186742067337, + 0.009703965857625008, + 0.008656383492052555, + 0.00742552662268281, + 0.006305927876383066, + 0.009109726175665855, + 0.004735956899821758, + 0.0063646831549704075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.020581256598234177 + } + ], + "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/0000022/group_3/index_tensor([3184, 3184])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_tensor([3184, 3184])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..86491080805f2e3b9cc33296cc6f4bc501ab37cd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_tensor([3184, 3184])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "grand drone camera, looking slightly upward, smoothly and sideways tracks an open-top anti-gravity chariot, steadily and calmly traversing the cascading night sky of a futuristic metropolis. Atop the chariot, Zhuge Liang stands poised, clad in a futuristic Hanfu garment crafted from deep blue nanofabric. The classic silhouette of the wide, wide-sleeved robe remains, but the robe is embellished with a pattern reminiscent of shimmering circuitry, and the wide cuffs and collar are inlaid with soft LED light strips. His signature long hair is held back by a simple metal crown. His face is calm and sagacious, his gaze deeply surveying the endless cyberpunk cityscape below. In his hand, he gently grasps a translucent holographic feather fan composed of blue data streams. The chariot he controls is constructed from a polished, obsidian-like alloy, its lines sleek and aggressive. Etched on its sides are glowing, classical auspicious cloud motifs, and beneath it, several anti-gravity thrusters emit a steady cyan glow. Below, towering, interlocking skyscrapers pierced the polluted clouds. Giant holographic billboards flashed animated dragons and shimmering Chinese characters. Countless small aircraft, like fireflies, darted through the airways between the buildings. The relentless drizzle cast a dazzling halo of neon light across the city. Against a cool backdrop, the entire scene was illuminated by neon hues of pink, cyan, and gold, creating a solemn, tranquil, and futuristic atmosphere.", + "seed": 8660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8660_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": -0.051113540027290585, + "vlm_logprob": -0.051113540027290585, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.051113540027290585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09948841482400894 + ], + "opposite_logprob_tokens": [ + -2.8494884967803955 + ], + "logprob_reward": -0.09948841482400894, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09948841482400894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16162630915641785, + 0.1815628707408905, + 0.044238537549972534, + 0.04364974424242973, + 0.07974974811077118, + 0.04229355603456497, + 0.05641573667526245, + 0.10593875497579575, + 0.23309430480003357, + 0.05143051594495773 + ], + "credit_map": [ + 0.017089568078517914, + 0.013887214474380016, + 0.01332763209939003, + 0.011781860142946243, + 0.014546660706400871, + 0.022452348843216896, + 0.028561580926179886, + 0.039979416877031326, + 0.024544086307287216, + 0.01361544243991375, + 0.01579061895608902, + 0.013785864226520061, + 0.015474081970751286, + 0.022108744829893112, + 0.02422022633254528, + 0.05202377587556839, + 0.004160664044320583, + 0.005460401531308889, + 0.005767786409705877, + 0.004668325651437044, + 0.006289033219218254, + 0.005546411033719778, + 0.006586798466742039, + 0.0057591102086007595, + 0.005021828692406416, + 0.007613638881593943, + 0.005063590127974749, + 0.003334163688123226, + 0.00656913174316287, + 0.005950632970780134, + 0.005667336285114288, + 0.004429415799677372, + 0.006256463471800089, + 0.0077451784163713455, + 0.005532246083021164, + 0.004615365993231535, + 0.011738419532775879, + 0.010443351231515408, + 0.00928120594471693, + 0.024137506261467934, + 0.005270859692245722, + 0.007923929020762444, + 0.004876324441283941, + 0.003953431732952595, + 0.005407540127635002, + 0.004685593768954277, + 0.003795844269916415, + 0.006380025297403336, + 0.006270731333643198, + 0.010407090187072754, + 0.005734055768698454, + 0.007900742813944817, + 0.006442995276302099, + 0.007389180362224579, + 0.006001144181936979, + 0.006269788835197687, + 0.01183931715786457, + 0.011665171012282372, + 0.006108480505645275, + 0.006086952053010464, + 0.028365986421704292, + 0.02586926706135273, + 0.007505352608859539, + 0.00849821325391531, + 0.02318606525659561, + 0.04587550833821297, + 0.026041319593787193, + 0.02566801942884922, + 0.02512461319565773, + 0.03171992301940918, + 0.021099817007780075, + 0.0343790240585804, + 0.007987713441252708, + 0.007255693431943655, + 0.008405093103647232, + 0.004434620495885611, + 0.007506735157221556, + 0.006825852673500776, + 0.004636220168322325, + 0.004378580953925848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.09948841482400894 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372933343052864 + ], + "opposite_logprob_tokens": [ + -3.408729314804077 + ], + "logprob_reward": -0.03372933343052864, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372933343052864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884458065032959, + 0.11501219123601913, + 0.035434793680906296, + 0.03599158674478531, + 0.12371420860290527, + 0.06295198202133179, + 0.07541119307279587, + 0.12221908569335938, + 0.13284820318222046, + 0.10797102004289627 + ], + "credit_map": [ + 0.02795196697115898, + 0.023539148271083832, + 0.021198730915784836, + 0.01990175060927868, + 0.01750798523426056, + 0.0217930618673563, + 0.032804541289806366, + 0.023748593404889107, + 0.008318192325532436, + 0.01136694010347128, + 0.008847910910844803, + 0.008374684490263462, + 0.022942356765270233, + 0.015510743483901024, + 0.0228232741355896, + 0.01682807318866253, + 0.004125382751226425, + 0.004336513578891754, + 0.003850274719297886, + 0.003971240483224392, + 0.004062771797180176, + 0.0042399209924042225, + 0.0053666927851736546, + 0.0054819961078464985, + 0.003839698387309909, + 0.005535217933356762, + 0.004206698387861252, + 0.0029238779097795486, + 0.004552352707833052, + 0.004381922073662281, + 0.004805407486855984, + 0.005746407434344292, + 0.008873592130839825, + 0.013043941929936409, + 0.006387942470610142, + 0.005246439483016729, + 0.018043003976345062, + 0.016230879351496696, + 0.016279369592666626, + 0.039609022438526154, + 0.006375649478286505, + 0.01246484275907278, + 0.004856386687606573, + 0.0045790462754666805, + 0.010247396305203438, + 0.009769000113010406, + 0.005031534470617771, + 0.009628125466406345, + 0.009655827656388283, + 0.01650983840227127, + 0.006844285409897566, + 0.007148600649088621, + 0.009112700819969177, + 0.010022149421274662, + 0.007005578838288784, + 0.009112206287682056, + 0.010755793191492558, + 0.011125260032713413, + 0.010184112004935741, + 0.0090806158259511, + 0.03116694651544094, + 0.03030518814921379, + 0.008827243000268936, + 0.010773921385407448, + 0.01079218927770853, + 0.012824379839003086, + 0.018993867561221123, + 0.019003281369805336, + 0.015197916887700558, + 0.0170164592564106, + 0.020728426054120064, + 0.018291668966412544, + 0.019147146493196487, + 0.014247982762753963, + 0.0171840637922287, + 0.011279174126684666, + 0.012891457416117191, + 0.012511835433542728, + 0.008139616809785366, + 0.012569720856845379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03372933343052864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034342855215073 + ], + "opposite_logprob_tokens": [ + -2.3503434658050537 + ], + "logprob_reward": -0.10034342855215073, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034342855215073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18598341941833496, + 0.07659147679805756, + 0.03354036435484886, + 0.021459871903061867, + 0.048034366220235825, + 0.032134268432855606, + 0.061265163123607635, + 0.2738412022590637, + 0.19942410290241241, + 0.06772586703300476 + ], + "credit_map": [ + 0.024883920326828957, + 0.02170451171696186, + 0.019757390022277832, + 0.017629573121666908, + 0.019515538588166237, + 0.020288784056901932, + 0.03642343357205391, + 0.02578023262321949, + 0.009100987575948238, + 0.010287418030202389, + 0.0071218982338905334, + 0.005140346474945545, + 0.010399982333183289, + 0.00956115871667862, + 0.012773425318300724, + 0.012206249870359898, + 0.0034011700190603733, + 0.004592382349073887, + 0.0032523456029593945, + 0.0032239577267318964, + 0.004267797339707613, + 0.003996072802692652, + 0.00488576665520668, + 0.005920868366956711, + 0.0023851110599935055, + 0.003578533884137869, + 0.0021824550349265337, + 0.001404517563059926, + 0.0029157663229852915, + 0.0027283395174890757, + 0.002780183218419552, + 0.0034849594812840223, + 0.004288661293685436, + 0.006788230501115322, + 0.0036139939911663532, + 0.002741388510912657, + 0.009033502079546452, + 0.008360383100807667, + 0.00465491134673357, + 0.008553287014365196, + 0.0037166804540902376, + 0.00451697176322341, + 0.0035534570924937725, + 0.002467858837917447, + 0.005070718005299568, + 0.005139155313372612, + 0.0023561962880194187, + 0.005313224624842405, + 0.007306782063096762, + 0.013201561756432056, + 0.008378516882658005, + 0.005980685353279114, + 0.007281026337295771, + 0.007634470704942942, + 0.005647097714245319, + 0.005835010204464197, + 0.027127010747790337, + 0.03337651863694191, + 0.028611309826374054, + 0.030746372416615486, + 0.06145284324884415, + 0.055184684693813324, + 0.020225875079631805, + 0.017116524279117584, + 0.020121203735470772, + 0.034690987318754196, + 0.020433485507965088, + 0.016739318147301674, + 0.026232833042740822, + 0.0278464462608099, + 0.01820102520287037, + 0.035158783197402954, + 0.009928010404109955, + 0.007392426021397114, + 0.007353314198553562, + 0.005039751995354891, + 0.014297853223979473, + 0.012601238675415516, + 0.005499799735844135, + 0.005613463465124369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.10034342855215073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950473714619875 + ], + "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": [ + 0.23845909535884857, + 0.1694459617137909, + 0.04974159598350525, + 0.03195155784487724, + 0.04551423341035843, + 0.0549272783100605, + 0.07224955409765244, + 0.1548977792263031, + 0.11579127609729767, + 0.06702172756195068 + ], + "credit_map": [ + 0.02878453955054283, + 0.02672739513218403, + 0.030567146837711334, + 0.023325573652982712, + 0.02129446715116501, + 0.026608744636178017, + 0.043783485889434814, + 0.03736774995923042, + 0.012521923519670963, + 0.01592056266963482, + 0.014695857651531696, + 0.012237847782671452, + 0.032743435353040695, + 0.022188615053892136, + 0.029240358620882034, + 0.029897360131144524, + 0.004717162810266018, + 0.006211405619978905, + 0.006724770180881023, + 0.004576892126351595, + 0.006402386352419853, + 0.006618902087211609, + 0.0076460582204163074, + 0.006844021379947662, + 0.0031063060741871595, + 0.004111784044653177, + 0.002134717535227537, + 0.0019815650302916765, + 0.004915421828627586, + 0.005040895193815231, + 0.003928239457309246, + 0.006732627749443054, + 0.0035991952754557133, + 0.0060573299415409565, + 0.0037155188620090485, + 0.002679844619706273, + 0.007498056627810001, + 0.0064580137841403484, + 0.005128777120262384, + 0.010377496480941772, + 0.006248540710657835, + 0.007742704823613167, + 0.008150177076458931, + 0.007340478710830212, + 0.008238296955823898, + 0.006945418659597635, + 0.004530916456133127, + 0.005730742122977972, + 0.00847933255136013, + 0.010510509833693504, + 0.009571521542966366, + 0.008799765259027481, + 0.008254100568592548, + 0.0081565473228693, + 0.008541337214410305, + 0.009936438873410225, + 0.017166873440146446, + 0.021498549729585648, + 0.018367914482951164, + 0.01691405475139618, + 0.030716076493263245, + 0.02681932970881462, + 0.013772455975413322, + 0.00964253582060337, + 0.01133529283106327, + 0.01585332676768303, + 0.012106011621654034, + 0.010352478362619877, + 0.016322918236255646, + 0.018613139167428017, + 0.01488501112908125, + 0.016323093324899673, + 0.008730264380574226, + 0.008310038596391678, + 0.01022503525018692, + 0.005171601194888353, + 0.01175162848085165, + 0.010529076680541039, + 0.005968169774860144, + 0.006335909012705088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.005950473714619875 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605604961514473 + ], + "opposite_logprob_tokens": [ + -4.141056060791016 + ], + "logprob_reward": -0.01605604961514473, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605604961514473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4748213291168213, + 0.024631967768073082, + 0.057219356298446655, + 0.01955578662455082, + 0.07147286087274551, + 0.04525677114725113, + 0.13550078868865967, + 0.08684021234512329, + 0.03281988576054573, + 0.0518811009824276 + ], + "credit_map": [ + 0.02999585121870041, + 0.026727262884378433, + 0.030011916533112526, + 0.032534074038267136, + 0.038416195660829544, + 0.06855162233114243, + 0.11238225549459457, + 0.1362021267414093, + 0.001809737179428339, + 0.002706679282709956, + 0.002297723200172186, + 0.002073196228593588, + 0.0035857404582202435, + 0.00314884795807302, + 0.004278187640011311, + 0.004731854423880577, + 0.003744251560419798, + 0.004629322327673435, + 0.004382097162306309, + 0.0035730956587940454, + 0.005278048571199179, + 0.007265837397426367, + 0.011387796141207218, + 0.016958903521299362, + 0.001712262979708612, + 0.0027799273375421762, + 0.0016184597043320537, + 0.0013718304689973593, + 0.003133001271635294, + 0.0025619990192353725, + 0.0032073576003313065, + 0.0031709461472928524, + 0.0038876745384186506, + 0.006920394953340292, + 0.005604679696261883, + 0.0041294400580227375, + 0.0178383756428957, + 0.01757812686264515, + 0.005723222624510527, + 0.009790947660803795, + 0.003472652519121766, + 0.005022660829126835, + 0.003927518147975206, + 0.002662384184077382, + 0.00950054731220007, + 0.009763232432305813, + 0.0027722788508981466, + 0.008135494776070118, + 0.012800122611224651, + 0.014483477920293808, + 0.008074603043496609, + 0.008848695084452629, + 0.022164463996887207, + 0.03645516186952591, + 0.010596591979265213, + 0.022077662870287895, + 0.008485255762934685, + 0.007039680145680904, + 0.005991281941533089, + 0.005402103066444397, + 0.01904275268316269, + 0.0204415712505579, + 0.008072057738900185, + 0.012365496717393398, + 0.003532619448378682, + 0.005853837821632624, + 0.003073833417147398, + 0.003404793096706271, + 0.004491119645535946, + 0.005805716384202242, + 0.003540933830663562, + 0.003117028623819351, + 0.007111493963748217, + 0.005836767144501209, + 0.006277274340391159, + 0.0033007613383233547, + 0.006430190987884998, + 0.010035721585154533, + 0.004401789512485266, + 0.008487094193696976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.01605604961514473 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand drone camera, looking slightly upward, smoothly and sideways tracks an open-top anti-gravity chariot, steadily and calmly traversing the cascading night sky of a futuristic metropolis. Atop the chariot, Zhuge Liang stands poised, clad in a futuristic Hanfu garment crafted from deep blue nanofabric. The classic silhouette of the wide, wide-sleeved robe remains, but the robe is embellished with a pattern reminiscent of shimmering circuitry, and the wide cuffs and collar are inlaid with soft LED light strips. His signature long hair is held back by a simple metal crown. His face is calm and sagacious, his gaze deeply surveying the endless cyberpunk cityscape below. In his hand, he gently grasps a translucent holographic feather fan composed of blue data streams. The chariot he controls is constructed from a polished, obsidian-like alloy, its lines sleek and aggressive. Etched on its sides are glowing, classical auspicious cloud motifs, and beneath it, several anti-gravity thrusters emit a steady cyan glow. Below, towering, interlocking skyscrapers pierced the polluted clouds. Giant holographic billboards flashed animated dragons and shimmering Chinese characters. Countless small aircraft, like fireflies, darted through the airways between the buildings. The relentless drizzle cast a dazzling halo of neon light across the city. Against a cool backdrop, the entire scene was illuminated by neon hues of pink, cyan, and gold, creating a solemn, tranquil, and futuristic atmosphere.", + "seed": 8661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8661_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": -0.07609179988503456, + "vlm_logprob": -0.07609179988503456, + "vlm_yes_no_margin": 2.949999976158142, + "vlm_token_credit": -0.07609179988503456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1646781712770462 + ], + "opposite_logprob_tokens": [ + -2.0396780967712402 + ], + "logprob_reward": -0.1646781712770462, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1646781712770462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015078447759151459, + 0.9018034934997559, + 0.006915382109582424, + 0.010713149793446064, + 0.00607343390583992, + 0.009145955555140972, + 0.007604176178574562, + 0.014712626114487648, + 0.005766818765550852, + 0.02218654565513134 + ], + "credit_map": [ + 0.0022585082333534956, + 0.0020074015483260155, + 0.0016314882086589932, + 0.0009935302659869194, + 0.0020971253979951143, + 0.0031865620985627174, + 0.0012065034825354815, + 0.0016973273595795035, + 0.07458307594060898, + 0.3440353274345398, + 0.025302303954958916, + 0.02066955529153347, + 0.03352360427379608, + 0.3556692600250244, + 0.026835724711418152, + 0.021184567362070084, + 0.001102697104215622, + 0.000866183137986809, + 0.0008041891851462424, + 0.0007206547888927162, + 0.000963321013841778, + 0.0010836133733391762, + 0.0005573611706495285, + 0.0008173623355105519, + 0.001969914184883237, + 0.0027088106144219637, + 0.0009956571739166975, + 0.0009080285090021789, + 0.0011356929317116737, + 0.0014389118878170848, + 0.0006555648287758231, + 0.0009005676256492734, + 0.0014145650202408433, + 0.0009728812729008496, + 0.0007740925066173077, + 0.0006286195130087435, + 0.0006139779579825699, + 0.0008342002402059734, + 0.0004061312647536397, + 0.00042896586819551885, + 0.0008729809778742492, + 0.0014933875063434243, + 0.0012813502689823508, + 0.001179798855446279, + 0.001337925554253161, + 0.0012168087996542454, + 0.0007899589836597443, + 0.0009737437940202653, + 0.001794512034393847, + 0.0008519950788468122, + 0.0005495712393894792, + 0.0005565903265960515, + 0.001182620064355433, + 0.0018281190423294902, + 0.0005166626651771367, + 0.0003241046506445855, + 0.0036060637794435024, + 0.0017419342184439301, + 0.0008975520613603294, + 0.0006178321782499552, + 0.0009495242848061025, + 0.005246824584901333, + 0.0008465105784125626, + 0.0008063839632086456, + 0.0010044729569926858, + 0.0007959498907439411, + 0.0003782297426369041, + 0.00042189733358100057, + 0.0008422086830250919, + 0.0015534545527771115, + 0.0003955774591304362, + 0.0003750272444449365, + 0.008722388185560703, + 0.00117768335621804, + 0.0007561292732134461, + 0.0006006552139297128, + 0.000685650622472167, + 0.009120570495724678, + 0.0005638026050291955, + 0.0005596630508080125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.1646781712770462 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812185674905777 + ], + "opposite_logprob_tokens": [ + -3.2881219387054443 + ], + "logprob_reward": -0.03812185674905777, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812185674905777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01342453621327877, + 0.8574634194374084, + 0.008507495746016502, + 0.012328912504017353, + 0.007390724029392004, + 0.014369962736964226, + 0.024353478103876114, + 0.007743166293948889, + 0.006305484101176262, + 0.048112768679857254 + ], + "credit_map": [ + 0.0021853437647223473, + 0.001534589915536344, + 0.0009410541388206184, + 0.0009382702992297709, + 0.0024349058512598276, + 0.0026000060606747866, + 0.0013854376738891006, + 0.001404928625561297, + 0.07115501910448074, + 0.32684192061424255, + 0.02386161871254444, + 0.019868645817041397, + 0.03192933648824692, + 0.3379833996295929, + 0.025958262383937836, + 0.019865253940224648, + 0.0012967509683221579, + 0.0011440558591857553, + 0.0010286426404491067, + 0.0008519936818629503, + 0.0010425483342260122, + 0.0015733266482129693, + 0.0006708443397656083, + 0.0008993331575766206, + 0.0024670190177857876, + 0.00260541751049459, + 0.0014956871746107936, + 0.0010729990899562836, + 0.0012678629718720913, + 0.0014054853236302733, + 0.0008963220752775669, + 0.0011181191075593233, + 0.0020616664551198483, + 0.0010463843354955316, + 0.0008782505756244063, + 0.00047198947868309915, + 0.0008161068544723094, + 0.0011557383695617318, + 0.0004979124641977251, + 0.0004626756126526743, + 0.0015095560811460018, + 0.0025862138718366623, + 0.002415329683572054, + 0.0010694715892896056, + 0.0018707483541220427, + 0.0026636850088834763, + 0.0013756284024566412, + 0.0008793296874500811, + 0.007351716514676809, + 0.002169576706364751, + 0.0018195263110101223, + 0.0012871082872152328, + 0.0021204263903200626, + 0.007192442659288645, + 0.0015805577859282494, + 0.0008321225177496672, + 0.001283689052797854, + 0.001092690508812666, + 0.0011640734737738967, + 0.0005199629813432693, + 0.0008331964490935206, + 0.0016165135893970728, + 0.0006079854792915285, + 0.0006250545266084373, + 0.0013282415457069874, + 0.0007481571519747376, + 0.0005805797409266233, + 0.0005130710196681321, + 0.0007470257696695626, + 0.0013493971200659871, + 0.0004386116052046418, + 0.0006003999733366072, + 0.019282273948192596, + 0.002507775789126754, + 0.0012256692862138152, + 0.0008649816154502332, + 0.0014436012133955956, + 0.02051187865436077, + 0.001074657659046352, + 0.0012019308051094413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.03812185674905777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497550964355469 + ], + "opposite_logprob_tokens": [ + -2.9299755096435547 + ], + "logprob_reward": -0.05497550964355469, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497550964355469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02593451738357544, + 0.7705588340759277, + 0.020220840349793434, + 0.015770521014928818, + 0.005471198353916407, + 0.06710182130336761, + 0.02754969149827957, + 0.005086070857942104, + 0.005388012621551752, + 0.05691856890916824 + ], + "credit_map": [ + 0.004387735854834318, + 0.002773200161755085, + 0.002334238262847066, + 0.0019407484214752913, + 0.003736240789294243, + 0.005367293022572994, + 0.0026106054428964853, + 0.0027844561263918877, + 0.06506959348917007, + 0.29340261220932007, + 0.02149307169020176, + 0.01811797358095646, + 0.02900274284183979, + 0.30271604657173157, + 0.023061048239469528, + 0.0176957119256258, + 0.0034774134401232004, + 0.0027221818454563618, + 0.0016854555578902364, + 0.0015402394346892834, + 0.004666029941290617, + 0.0028051212430000305, + 0.001717469422146678, + 0.0016069303965196013, + 0.0030515280086547136, + 0.004011660814285278, + 0.0016189457383006811, + 0.0011085446458309889, + 0.0014178585261106491, + 0.0020956809166818857, + 0.0009352329070679843, + 0.0015310683520510793, + 0.0010856993030756712, + 0.0008288469398394227, + 0.0007216167286969721, + 0.0006735653150826693, + 0.0006949083181098104, + 0.0005388687714003026, + 0.0004709626082330942, + 0.00045673007844015956, + 0.006368307862430811, + 0.015623198822140694, + 0.004639620892703533, + 0.006237500347197056, + 0.00919946189969778, + 0.01581350527703762, + 0.004824075847864151, + 0.004396151285618544, + 0.00907337199896574, + 0.0027061826549470425, + 0.0013990342849865556, + 0.001303918077610433, + 0.0017725921934470534, + 0.009226166643202305, + 0.0010680453851819038, + 0.0010003804927691817, + 0.0009054275578819215, + 0.0011017194483429193, + 0.0004003664362244308, + 0.0004047863185405731, + 0.0006808458711020648, + 0.0007889916887506843, + 0.00042571997619234025, + 0.00037821344449184835, + 0.000991575070656836, + 0.0005043898709118366, + 0.0004447412211447954, + 0.0004321739834267646, + 0.000694233167450875, + 0.0014655161648988724, + 0.0003324753197375685, + 0.0005229084054008126, + 0.022025438025593758, + 0.0033433467615395784, + 0.0016790679655969143, + 0.0013790149241685867, + 0.0023779314942657948, + 0.023208385333418846, + 0.0014673474943265319, + 0.00143803795799613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 9 + ], + "selected_reward": -0.05497550964355469 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011510997079312801, + 0.8775138258934021, + 0.008952281437814236, + 0.011325567029416561, + 0.007842333056032658, + 0.02092166244983673, + 0.0053334953263401985, + 0.0035542326513677835, + 0.005808423738926649, + 0.04723726958036423 + ], + "credit_map": [ + 0.0018794452771544456, + 0.0012059989385306835, + 0.0010744696483016014, + 0.000812250713352114, + 0.0017488630255684257, + 0.0023029844742268324, + 0.00115178432315588, + 0.001335199922323227, + 0.07231657207012177, + 0.3344721794128418, + 0.02463654801249504, + 0.020252926275134087, + 0.03263184428215027, + 0.3464295566082001, + 0.02635124884545803, + 0.020422831177711487, + 0.0011708780657500029, + 0.0012463248567655683, + 0.0012940075248479843, + 0.0007598470547236502, + 0.0016637054504826665, + 0.00125609221868217, + 0.0007053716690279543, + 0.0008560538408346474, + 0.002126022009178996, + 0.002859797328710556, + 0.0010759147116914392, + 0.0009355195797979832, + 0.0010129879228770733, + 0.001623575808480382, + 0.0006878824206069112, + 0.0010038657346740365, + 0.001938869128935039, + 0.0011545199668034911, + 0.001103690592572093, + 0.0005759461782872677, + 0.0006381525890901685, + 0.001456226222217083, + 0.0004094933101441711, + 0.0005654345732182264, + 0.0023188453633338213, + 0.004660249222069979, + 0.0016080060740932822, + 0.0021304714027792215, + 0.00248371041379869, + 0.0049094329588115215, + 0.001544103492051363, + 0.001266842009499669, + 0.0014481988037005067, + 0.00048432499170303345, + 0.00042566092452034354, + 0.00046795938396826386, + 0.0005275727016851306, + 0.0013791925739496946, + 0.0003161630011163652, + 0.0002844221016857773, + 0.0004935555043630302, + 0.0006046927883289754, + 0.00041009054984897375, + 0.00027024484006688, + 0.0003271375026088208, + 0.0008216634741984308, + 0.00030223841895349324, + 0.0003246090782340616, + 0.0013517169281840324, + 0.0007449396653100848, + 0.000428084924351424, + 0.00035752117400988936, + 0.0005611138767562807, + 0.001594863715581596, + 0.0003777637321036309, + 0.000392419082345441, + 0.01903809793293476, + 0.0025328027550131083, + 0.0012281107483431697, + 0.0007670329650864005, + 0.0014708690578117967, + 0.02013830468058586, + 0.0009278351790271699, + 0.0011342110810801387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290370672941208 + ], + "opposite_logprob_tokens": [ + -2.237903594970703 + ], + "logprob_reward": -0.11290370672941208, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290370672941208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027611369267106056, + 0.6415908336639404, + 0.09122385084629059, + 0.09573780000209808, + 0.02782827615737915, + 0.04551241919398308, + 0.00635919813066721, + 0.012559524737298489, + 0.016912762075662613, + 0.03466394543647766 + ], + "credit_map": [ + 0.005320908967405558, + 0.00317393452860415, + 0.0019234739011153579, + 0.0015543103218078613, + 0.004207824356853962, + 0.006697474513202906, + 0.0022997537162154913, + 0.0024336909409612417, + 0.05824689939618111, + 0.2430265247821808, + 0.017619801685214043, + 0.015875408425927162, + 0.024287419393658638, + 0.2478611320257187, + 0.019952688366174698, + 0.014721010811626911, + 0.015909330919384956, + 0.010718543082475662, + 0.009223840199410915, + 0.011758966371417046, + 0.010686587542295456, + 0.014452372677624226, + 0.00798938050866127, + 0.0104848463088274, + 0.01671019196510315, + 0.02442902699112892, + 0.008971204981207848, + 0.007089617196470499, + 0.011639561504125595, + 0.013468701392412186, + 0.005162371788173914, + 0.008267142809927464, + 0.0066129411570727825, + 0.004758702591061592, + 0.003487702924758196, + 0.0017968633910641074, + 0.002040222752839327, + 0.005762220360338688, + 0.001555361901409924, + 0.0018142631743103266, + 0.003979029133915901, + 0.010396596975624561, + 0.0035228305496275425, + 0.0038617949467152357, + 0.0067783561535179615, + 0.010967628099024296, + 0.0032307456713169813, + 0.0027754439506679773, + 0.0011282300110906363, + 0.0007487295661121607, + 0.0004907098482362926, + 0.0005914916982874274, + 0.001305046142078936, + 0.0012219187337905169, + 0.0005191164673306048, + 0.0003539559547789395, + 0.0019113448215648532, + 0.002646929817274213, + 0.0009314605849795043, + 0.0009608917171135545, + 0.0014671576209366322, + 0.002720740158110857, + 0.0009587736567482352, + 0.0009622277575545013, + 0.003090853802859783, + 0.002561175962910056, + 0.001337813213467598, + 0.0010460219345986843, + 0.0025902017951011658, + 0.003873497946187854, + 0.0012218098854646087, + 0.001191390329040587, + 0.013929205946624279, + 0.0018318055663257837, + 0.000887851812876761, + 0.0006285670679062605, + 0.0011146805481985211, + 0.014715301804244518, + 0.0007036547176539898, + 0.0008528818725608289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.11290370672941208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand drone camera, looking slightly upward, smoothly and sideways tracks an open-top anti-gravity chariot, steadily and calmly traversing the cascading night sky of a futuristic metropolis. Atop the chariot, Zhuge Liang stands poised, clad in a futuristic Hanfu garment crafted from deep blue nanofabric. The classic silhouette of the wide, wide-sleeved robe remains, but the robe is embellished with a pattern reminiscent of shimmering circuitry, and the wide cuffs and collar are inlaid with soft LED light strips. His signature long hair is held back by a simple metal crown. His face is calm and sagacious, his gaze deeply surveying the endless cyberpunk cityscape below. In his hand, he gently grasps a translucent holographic feather fan composed of blue data streams. The chariot he controls is constructed from a polished, obsidian-like alloy, its lines sleek and aggressive. Etched on its sides are glowing, classical auspicious cloud motifs, and beneath it, several anti-gravity thrusters emit a steady cyan glow. Below, towering, interlocking skyscrapers pierced the polluted clouds. Giant holographic billboards flashed animated dragons and shimmering Chinese characters. Countless small aircraft, like fireflies, darted through the airways between the buildings. The relentless drizzle cast a dazzling halo of neon light across the city. Against a cool backdrop, the entire scene was illuminated by neon hues of pink, cyan, and gold, creating a solemn, tranquil, and futuristic atmosphere.", + "seed": 8662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8662_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": -0.06366615109145642, + "vlm_logprob": -0.06366615109145642, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.06366615109145642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15196171402931213 + ], + "opposite_logprob_tokens": [ + -2.1519618034362793 + ], + "logprob_reward": -0.15196171402931213, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.15196171402931213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11877516657114029, + 0.1550098955631256, + 0.06676055490970612, + 0.04278360679745674, + 0.17296619713306427, + 0.047207314521074295, + 0.04757750779390335, + 0.23553192615509033, + 0.08124891668558121, + 0.032138943672180176 + ], + "credit_map": [ + 0.01259130984544754, + 0.019834039732813835, + 0.010363016277551651, + 0.013546351343393326, + 0.0128525635227561, + 0.021943213418126106, + 0.013064667582511902, + 0.014580008573830128, + 0.019557906314730644, + 0.030440712347626686, + 0.015275978483259678, + 0.01799517124891281, + 0.0204310305416584, + 0.01693359762430191, + 0.019268188625574112, + 0.015107305720448494, + 0.010033108294010162, + 0.00891854427754879, + 0.00584002397954464, + 0.006441538687795401, + 0.01018531247973442, + 0.008386759087443352, + 0.00697465380653739, + 0.009980614297091961, + 0.006045214366167784, + 0.006140724290162325, + 0.0028859893791377544, + 0.005996112711727619, + 0.005780166946351528, + 0.0062600900419056416, + 0.005399428773671389, + 0.004275880288332701, + 0.01744280569255352, + 0.01927114650607109, + 0.020309392362833023, + 0.019478239119052887, + 0.024290993809700012, + 0.020729221403598785, + 0.03579862788319588, + 0.015645772218704224, + 0.005336307920515537, + 0.004190038423985243, + 0.005617313552647829, + 0.0069033256731927395, + 0.006299481261521578, + 0.0052067735232412815, + 0.008685575798153877, + 0.004968499764800072, + 0.006227121688425541, + 0.003883558791130781, + 0.00804801844060421, + 0.00522961188107729, + 0.0037111900746822357, + 0.006604130379855633, + 0.00876401737332344, + 0.005109858699142933, + 0.07958778738975525, + 0.012476260773837566, + 0.009945963509380817, + 0.009214961901307106, + 0.012348031625151634, + 0.09368504583835602, + 0.010766533203423023, + 0.007507328875362873, + 0.014746520668268204, + 0.009603557176887989, + 0.009372127242386341, + 0.007022412959486246, + 0.009421197697520256, + 0.01487155444920063, + 0.009091658517718315, + 0.007119887508451939, + 0.007605441380292177, + 0.0038796067237854004, + 0.0029489989392459393, + 0.002717251656576991, + 0.0034081374760717154, + 0.0066353194415569305, + 0.0027813883498311043, + 0.0021627983078360558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.15196171402931213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048675332218408585 + ], + "opposite_logprob_tokens": [ + -3.048675298690796 + ], + "logprob_reward": -0.048675332218408585, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048675332218408585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23736682534217834, + 0.05306265503168106, + 0.058417126536369324, + 0.05963010713458061, + 0.0682031661272049, + 0.10999859869480133, + 0.059025075286626816, + 0.14362432062625885, + 0.1364247351884842, + 0.074247345328331 + ], + "credit_map": [ + 0.019524630159139633, + 0.02579035982489586, + 0.023953506723046303, + 0.03269905596971512, + 0.026309439912438393, + 0.038971196860075, + 0.028027232736349106, + 0.04209142550826073, + 0.006703094579279423, + 0.006731273606419563, + 0.00554525526240468, + 0.006585701368749142, + 0.00727942306548357, + 0.006221204996109009, + 0.006996641401201487, + 0.0070000700652599335, + 0.00917072780430317, + 0.006195317953824997, + 0.005290069617331028, + 0.0070079416036605835, + 0.007361740339547396, + 0.00853324681520462, + 0.006031763274222612, + 0.008826322853565216, + 0.00928165391087532, + 0.007468894589692354, + 0.006849372759461403, + 0.006182321347296238, + 0.007751275785267353, + 0.007932658307254314, + 0.007817523553967476, + 0.006346414797008038, + 0.00700464379042387, + 0.009083526208996773, + 0.009239769540727139, + 0.007932393811643124, + 0.008580885827541351, + 0.009642073884606361, + 0.009408004581928253, + 0.007311872206628323, + 0.015774628147482872, + 0.006908657029271126, + 0.016790932044386864, + 0.013653740286827087, + 0.010651914402842522, + 0.014438265934586525, + 0.017547115683555603, + 0.01423335075378418, + 0.009432134218513966, + 0.007720232009887695, + 0.007305137813091278, + 0.006483732722699642, + 0.007461764384061098, + 0.008037255145609379, + 0.007586232852190733, + 0.004998595919460058, + 0.043077584356069565, + 0.00961267575621605, + 0.007699329871684313, + 0.006970192771404982, + 0.008011659607291222, + 0.05040125548839569, + 0.010420114733278751, + 0.007431526202708483, + 0.028689103201031685, + 0.013850408606231213, + 0.01736229844391346, + 0.010731110349297523, + 0.011909070424735546, + 0.028804680332541466, + 0.015236707404255867, + 0.009841359220445156, + 0.024228280410170555, + 0.007524962071329355, + 0.004041646607220173, + 0.0025374290999025106, + 0.005890413653105497, + 0.02470940165221691, + 0.002751742023974657, + 0.002563475165516138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.048675332218408585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307219386100769 + ], + "opposite_logprob_tokens": [ + -3.16807222366333 + ], + "logprob_reward": -0.04307219386100769, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307219386100769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09931197017431259, + 0.05436411872506142, + 0.0380726121366024, + 0.019950073212385178, + 0.05754741653800011, + 0.042366672307252884, + 0.04730052128434181, + 0.065550297498703, + 0.031264375895261765, + 0.5442719459533691 + ], + "credit_map": [ + 0.012030168436467648, + 0.011692728847265244, + 0.008653519675135612, + 0.009349418804049492, + 0.012554213404655457, + 0.023778611794114113, + 0.01025362964719534, + 0.010999681428074837, + 0.00486041558906436, + 0.007141624111682177, + 0.004907859023660421, + 0.01249407883733511, + 0.004522255621850491, + 0.008402418345212936, + 0.006023345049470663, + 0.006012121215462685, + 0.005026856437325478, + 0.005616402253508568, + 0.003998101688921452, + 0.004730509594082832, + 0.00473281042650342, + 0.0049997796304523945, + 0.00443512387573719, + 0.004533029161393642, + 0.002148919040337205, + 0.0027072993107140064, + 0.002062321873381734, + 0.0021003212314099073, + 0.0027885884046554565, + 0.0023869809228926897, + 0.0030234423466026783, + 0.002732199151068926, + 0.005331871565431356, + 0.007018318399786949, + 0.00874922052025795, + 0.006284681148827076, + 0.008185010403394699, + 0.007201893255114555, + 0.00862659327685833, + 0.006149827968329191, + 0.007396178785711527, + 0.005446390248835087, + 0.004508634563535452, + 0.004368678200989962, + 0.005185446236282587, + 0.005701766349375248, + 0.005585365928709507, + 0.00417421106249094, + 0.006888176780194044, + 0.004814485087990761, + 0.008634676225483418, + 0.005200129933655262, + 0.004283858463168144, + 0.005487163085490465, + 0.006542180199176073, + 0.00544985244050622, + 0.013744230382144451, + 0.008369021117687225, + 0.0067286016419529915, + 0.004255852196365595, + 0.006419131066650152, + 0.016224926337599754, + 0.005328366532921791, + 0.004480164963752031, + 0.0054223802872002125, + 0.003320840420201421, + 0.0032208289485424757, + 0.002816429128870368, + 0.0031063260976225138, + 0.006729604676365852, + 0.003739457344636321, + 0.0029085089918226004, + 0.2170531302690506, + 0.021124020218849182, + 0.02288087084889412, + 0.0064404928125441074, + 0.02742708846926689, + 0.22322842478752136, + 0.014420014806091785, + 0.011697943322360516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.04307219386100769 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109740972518921, + 0.052557218819856644, + 0.057892683893442154, + 0.015884192660450935, + 0.03285204991698265, + 0.015591558068990707, + 0.03451629728078842, + 0.025951923802495003, + 0.022679636254906654, + 0.5311003923416138 + ], + "credit_map": [ + 0.020345471799373627, + 0.022221654653549194, + 0.025425095111131668, + 0.019947493448853493, + 0.027346763759851456, + 0.032019972801208496, + 0.022809285670518875, + 0.040858346968889236, + 0.006741955876350403, + 0.010540508665144444, + 0.005029435735195875, + 0.005681437440216541, + 0.006392280571162701, + 0.005612423177808523, + 0.006336056161671877, + 0.006223117001354694, + 0.007591683883219957, + 0.008370901457965374, + 0.006012280005961657, + 0.006899259518831968, + 0.0065460484474897385, + 0.00750740198418498, + 0.006913613993674517, + 0.008051492273807526, + 0.0023305683862417936, + 0.0016452852869406343, + 0.0013808195944875479, + 0.001538761192932725, + 0.00214772904291749, + 0.0022055511362850666, + 0.002495699329301715, + 0.0021397790405899286, + 0.003532730508595705, + 0.0038761242758482695, + 0.0031560829374939203, + 0.003899344941601157, + 0.004459401126950979, + 0.0035000585485249758, + 0.007072820328176022, + 0.003355485387146473, + 0.00229679886251688, + 0.0014351282734423876, + 0.0021314697805792093, + 0.0019655253272503614, + 0.0014614167157560587, + 0.0021142021287232637, + 0.0025195206981152296, + 0.0016674963990226388, + 0.004236782900989056, + 0.0024920175783336163, + 0.006541764363646507, + 0.003466062480583787, + 0.00325103965587914, + 0.00392901012673974, + 0.004910737741738558, + 0.005688882898539305, + 0.005923400167375803, + 0.0027012380305677652, + 0.0024132532998919487, + 0.0017379665514454246, + 0.002284381538629532, + 0.006900151260197163, + 0.002195499837398529, + 0.001796032185666263, + 0.0038438269402831793, + 0.002169220242649317, + 0.0026003418024629354, + 0.0021453462541103363, + 0.0023541338741779327, + 0.004130338318645954, + 0.0035199776757508516, + 0.0019164516124874353, + 0.21078786253929138, + 0.02041168510913849, + 0.02340477891266346, + 0.006373973563313484, + 0.026801548898220062, + 0.2172451913356781, + 0.0142517676576972, + 0.011823579668998718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.004634710028767586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06998680531978607 + ], + "opposite_logprob_tokens": [ + -2.6949868202209473 + ], + "logprob_reward": -0.06998680531978607, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06998680531978607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06812811642885208, + 0.2618029713630676, + 0.10157511383295059, + 0.021444767713546753, + 0.08242351561784744, + 0.022999534383416176, + 0.020096484571695328, + 0.009318330325186253, + 0.023694150149822235, + 0.38851702213287354 + ], + "credit_map": [ + 0.007813322357833385, + 0.011971477419137955, + 0.006251719314604998, + 0.006090626120567322, + 0.009574967436492443, + 0.009839491918683052, + 0.007938080467283726, + 0.008648435585200787, + 0.03177766501903534, + 0.04679951071739197, + 0.02608528546988964, + 0.03427514806389809, + 0.03500955551862717, + 0.027720147743821144, + 0.03350211679935455, + 0.026633549481630325, + 0.013293796218931675, + 0.012280126102268696, + 0.009965190663933754, + 0.010998157784342766, + 0.014141281135380268, + 0.016436899080872536, + 0.012001005001366138, + 0.012458649463951588, + 0.0029810036066919565, + 0.0032880958169698715, + 0.0017656356794759631, + 0.00274475896731019, + 0.0026775819715112448, + 0.0031380841974169016, + 0.0022688002791255713, + 0.002580807078629732, + 0.008775088004767895, + 0.009015468880534172, + 0.00826994702219963, + 0.009815671481192112, + 0.01084932591766119, + 0.009342647157609463, + 0.01823102869093418, + 0.008124335668981075, + 0.0028777355328202248, + 0.0016294235829263926, + 0.0038363621570169926, + 0.002987385494634509, + 0.002153779147192836, + 0.002723789308220148, + 0.003586000297218561, + 0.003205059329047799, + 0.0022795074619352818, + 0.0017612824449315667, + 0.0030353134498000145, + 0.0019641369581222534, + 0.002379540354013443, + 0.0021621861960738897, + 0.0035688732750713825, + 0.002945643151178956, + 0.0015442537842318416, + 0.001244085724465549, + 0.0013593198964372277, + 0.0008950367919169366, + 0.0013956994516775012, + 0.001208966481499374, + 0.0008455631323158741, + 0.0008254048298113048, + 0.003711930476129055, + 0.002880144165828824, + 0.003370184451341629, + 0.0020245269406586885, + 0.002959021832793951, + 0.003871851833537221, + 0.0027361400425434113, + 0.00214035133831203, + 0.15576505661010742, + 0.015092169865965843, + 0.01666337437927723, + 0.0039453208446502686, + 0.01982574351131916, + 0.15992280840873718, + 0.009760413318872452, + 0.007542137987911701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.06998680531978607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "grand drone camera, looking slightly upward, smoothly and sideways tracks an open-top anti-gravity chariot, steadily and calmly traversing the cascading night sky of a futuristic metropolis. Atop the chariot, Zhuge Liang stands poised, clad in a futuristic Hanfu garment crafted from deep blue nanofabric. The classic silhouette of the wide, wide-sleeved robe remains, but the robe is embellished with a pattern reminiscent of shimmering circuitry, and the wide cuffs and collar are inlaid with soft LED light strips. His signature long hair is held back by a simple metal crown. His face is calm and sagacious, his gaze deeply surveying the endless cyberpunk cityscape below. In his hand, he gently grasps a translucent holographic feather fan composed of blue data streams. The chariot he controls is constructed from a polished, obsidian-like alloy, its lines sleek and aggressive. Etched on its sides are glowing, classical auspicious cloud motifs, and beneath it, several anti-gravity thrusters emit a steady cyan glow. Below, towering, interlocking skyscrapers pierced the polluted clouds. Giant holographic billboards flashed animated dragons and shimmering Chinese characters. Countless small aircraft, like fireflies, darted through the airways between the buildings. The relentless drizzle cast a dazzling halo of neon light across the city. Against a cool backdrop, the entire scene was illuminated by neon hues of pink, cyan, and gold, creating a solemn, tranquil, and futuristic atmosphere.", + "seed": 8663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_3/index_3184_seed_8663_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": -0.06607902329415083, + "vlm_logprob": -0.06607902329415083, + "vlm_yes_no_margin": 3.124999976158142, + "vlm_token_credit": -0.06607902329415083, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15549004077911377 + ], + "opposite_logprob_tokens": [ + -2.155489921569824 + ], + "logprob_reward": -0.15549004077911377, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.15549004077911377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3333669602870941, + 0.09494928270578384, + 0.05749852582812309, + 0.06857208907604218, + 0.07711581885814667, + 0.05883289501070976, + 0.0735066682100296, + 0.07820914685726166, + 0.10850996524095535, + 0.049438659101724625 + ], + "credit_map": [ + 0.07251698523759842, + 0.027718720957636833, + 0.04783867299556732, + 0.042071688920259476, + 0.03641544282436371, + 0.03457086160778999, + 0.0329471230506897, + 0.03928748145699501, + 0.015298348851501942, + 0.00863607693463564, + 0.023962877690792084, + 0.008359454572200775, + 0.006021957378834486, + 0.011524323374032974, + 0.007801254745572805, + 0.013344995677471161, + 0.011716642417013645, + 0.004902368877083063, + 0.007330409716814756, + 0.004712166264653206, + 0.005549236666411161, + 0.011797748506069183, + 0.0035383892245590687, + 0.007951569743454456, + 0.0083100451156497, + 0.007029772270470858, + 0.012841817922890186, + 0.007320734206587076, + 0.007099992595613003, + 0.006243743002414703, + 0.0073024616576731205, + 0.012423529289662838, + 0.008529981598258018, + 0.009299823082983494, + 0.008062331937253475, + 0.008528689853847027, + 0.016973940655589104, + 0.008696362376213074, + 0.007633855566382408, + 0.009390839375555515, + 0.011775102466344833, + 0.005168273579329252, + 0.010772204026579857, + 0.004495872650295496, + 0.009264475665986538, + 0.007121762726455927, + 0.0049013402312994, + 0.005333870183676481, + 0.010874050669372082, + 0.008439688943326473, + 0.007886581122875214, + 0.006758340634405613, + 0.011948658153414726, + 0.009631558321416378, + 0.01099226251244545, + 0.006975528318434954, + 0.011697371490299702, + 0.00625919084995985, + 0.01577550359070301, + 0.00983250979334116, + 0.007750878110527992, + 0.010229911655187607, + 0.006617730483412743, + 0.010046051815152168, + 0.014272413216531277, + 0.013250835239887238, + 0.015889102593064308, + 0.016843898221850395, + 0.011433301493525505, + 0.011525876820087433, + 0.013276951387524605, + 0.012017584405839443, + 0.008796566165983677, + 0.0057962932623922825, + 0.006801725830882788, + 0.005690563470125198, + 0.005233499687165022, + 0.0075804986990988255, + 0.004810517653822899, + 0.004729000851511955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.15549004077911377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029824934899806976 + ], + "opposite_logprob_tokens": [ + -3.52982497215271 + ], + "logprob_reward": -0.029824934899806976, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029824934899806976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14380067586898804, + 0.2045053094625473, + 0.10946892946958542, + 0.11895406991243362, + 0.07793720066547394, + 0.0594484806060791, + 0.05912524461746216, + 0.05066050589084625, + 0.12349865585565567, + 0.052600931376218796 + ], + "credit_map": [ + 0.01970501057803631, + 0.014934955164790154, + 0.018021706491708755, + 0.01910966821014881, + 0.010941511020064354, + 0.019428743049502373, + 0.016147837042808533, + 0.0255112387239933, + 0.026946108788251877, + 0.018469305709004402, + 0.04935420677065849, + 0.019914036616683006, + 0.014232330024242401, + 0.021610094234347343, + 0.021871238946914673, + 0.03210798650979996, + 0.016021957620978355, + 0.020064380019903183, + 0.010810124687850475, + 0.022242259234189987, + 0.008595393970608711, + 0.012239550240337849, + 0.006304453127086163, + 0.013190816156566143, + 0.01724918559193611, + 0.016141651198267937, + 0.01955327019095421, + 0.015344033017754555, + 0.01724020577967167, + 0.010065549053251743, + 0.01070959772914648, + 0.012650573626160622, + 0.0077148121781647205, + 0.008615844883024693, + 0.008932698518037796, + 0.011260268278419971, + 0.01622837968170643, + 0.008206125348806381, + 0.007161843124777079, + 0.009817228652536869, + 0.011669053696095943, + 0.004863935988396406, + 0.013776245526969433, + 0.004600172862410545, + 0.008433881215751171, + 0.0065334253013134, + 0.004927702248096466, + 0.004644065164029598, + 0.008670334704220295, + 0.006766895763576031, + 0.010979215614497662, + 0.005727178417146206, + 0.00584098557010293, + 0.008239450864493847, + 0.006678832229226828, + 0.006222349591553211, + 0.009358745999634266, + 0.004343102220445871, + 0.008262543007731438, + 0.004934702068567276, + 0.006189548876136541, + 0.00606708275154233, + 0.004974845331162214, + 0.0065299347043037415, + 0.019222252070903778, + 0.014964170753955841, + 0.01563131809234619, + 0.019199268892407417, + 0.01099289208650589, + 0.014001964591443539, + 0.017121093347668648, + 0.01236569695174694, + 0.009270376525819302, + 0.0054140822030603886, + 0.009842639788985252, + 0.006813072599470615, + 0.0033892940264195204, + 0.008162099868059158, + 0.00523387361317873, + 0.004475493915379047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.029824934899806976 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033156722784042 + ], + "opposite_logprob_tokens": [ + -2.3503315448760986 + ], + "logprob_reward": -0.10033156722784042, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033156722784042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13807712495326996, + 0.0958200991153717, + 0.11203857511281967, + 0.09671179950237274, + 0.0726887434720993, + 0.061636485159397125, + 0.12460637837648392, + 0.13929569721221924, + 0.10902750492095947, + 0.050097640603780746 + ], + "credit_map": [ + 0.029700331389904022, + 0.011472917161881924, + 0.023739060387015343, + 0.012917637825012207, + 0.00839934777468443, + 0.023480335250496864, + 0.01088466215878725, + 0.01748281717300415, + 0.011428817175328732, + 0.009568258188664913, + 0.019595175981521606, + 0.012272194027900696, + 0.008587823249399662, + 0.010068900883197784, + 0.009590690955519676, + 0.014708233997225761, + 0.020120054483413696, + 0.015880271792411804, + 0.013433012180030346, + 0.01888558454811573, + 0.006988681387156248, + 0.01461031287908554, + 0.007530427537858486, + 0.014590220525860786, + 0.012233556248247623, + 0.019323261454701424, + 0.012655275873839855, + 0.01850188337266445, + 0.008995953015983105, + 0.007788147311657667, + 0.010794107802212238, + 0.006419605575501919, + 0.008691812865436077, + 0.008896143175661564, + 0.009492224082350731, + 0.007479693740606308, + 0.01587603986263275, + 0.007668484468013048, + 0.007065253797918558, + 0.007519085425883532, + 0.013806191273033619, + 0.006060001906007528, + 0.013661288656294346, + 0.004931255709379911, + 0.006850360427051783, + 0.006160789169371128, + 0.004912781063467264, + 0.0052538104355335236, + 0.018997501581907272, + 0.01704948954284191, + 0.017235131934285164, + 0.012369412928819656, + 0.01568213663995266, + 0.015181303024291992, + 0.014059425331652164, + 0.01403197180479765, + 0.02252589538693428, + 0.01229703240096569, + 0.02709571272134781, + 0.011178892105817795, + 0.017303643748164177, + 0.019148722290992737, + 0.011369237676262856, + 0.0183765459805727, + 0.014641378074884415, + 0.014307712204754353, + 0.01485848892480135, + 0.018255524337291718, + 0.011278203688561916, + 0.011890921741724014, + 0.013486528769135475, + 0.010308739729225636, + 0.005889350548386574, + 0.004861216992139816, + 0.007939619943499565, + 0.005869600921869278, + 0.00848972238600254, + 0.00668928399682045, + 0.006582044996321201, + 0.0037767975591123104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.10033156722784042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107191201299429 + ], + "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": [ + 0.22924357652664185, + 0.16292712092399597, + 0.16595447063446045, + 0.09003150463104248, + 0.06833348423242569, + 0.07433135062456131, + 0.050363924354314804, + 0.04764247685670853, + 0.06902405619621277, + 0.042148031294345856 + ], + "credit_map": [ + 0.04902974143624306, + 0.01855405792593956, + 0.04068216681480408, + 0.02577662654221058, + 0.022548705339431763, + 0.02590956725180149, + 0.018532220274209976, + 0.028210483491420746, + 0.032373979687690735, + 0.0151775311678648, + 0.029424069449305534, + 0.02080066315829754, + 0.01506772730499506, + 0.019728796556591988, + 0.015187721699476242, + 0.01516662072390318, + 0.03434446454048157, + 0.022245293483138084, + 0.020641526207327843, + 0.02600323036313057, + 0.009654802270233631, + 0.028510114178061485, + 0.008398648351430893, + 0.016156386584043503, + 0.013776682317256927, + 0.010863912291824818, + 0.015468288213014603, + 0.012137600220739841, + 0.012723708525300026, + 0.007995862513780594, + 0.009012418799102306, + 0.00805302057415247, + 0.008597212843596935, + 0.009019440039992332, + 0.008695087395608425, + 0.008449357002973557, + 0.014339489862322807, + 0.007136660162359476, + 0.005521020852029324, + 0.006575208157300949, + 0.016300329938530922, + 0.006828912068158388, + 0.0173349492251873, + 0.0060587553307414055, + 0.008261101320385933, + 0.00735826138406992, + 0.006124677136540413, + 0.006064359564334154, + 0.008351300843060017, + 0.00681381905451417, + 0.006452702917158604, + 0.0044444953091442585, + 0.006560641340911388, + 0.006043668370693922, + 0.006068256683647633, + 0.005629031453281641, + 0.00889085978269577, + 0.005378678906708956, + 0.00877860002219677, + 0.006522544659674168, + 0.004074903205037117, + 0.0059912195429205894, + 0.0036701620556414127, + 0.004335504956543446, + 0.008523805066943169, + 0.009407088160514832, + 0.010379891842603683, + 0.012324158102273941, + 0.00810915231704712, + 0.007585442159324884, + 0.007086590398102999, + 0.00560792488977313, + 0.005756974220275879, + 0.006274626590311527, + 0.004143751226365566, + 0.006977608427405357, + 0.004550994839519262, + 0.0067092557437717915, + 0.00467107305303216, + 0.003063743468374014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.01107191201299429 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view from a high drone angle looking slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera smoothly and sideways tracking an open-top anti-gravity chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character Zhuge Liang visible wearing a deep blue nanofabric Hanfu with circuitry patterns and LED light strips?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot made of polished obsidian-like alloy with glowing cyan anti-gravity thrusters?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a futuristic metropolis with towering skyscrapers, holographic billboards, and neon lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367666155099869 + ], + "opposite_logprob_tokens": [ + -3.4086766242980957 + ], + "logprob_reward": -0.03367666155099869, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367666155099869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3086922764778137, + 0.03757943585515022, + 0.1523975133895874, + 0.034884318709373474, + 0.08065594732761383, + 0.048741620033979416, + 0.03291923180222511, + 0.18838469684123993, + 0.02549944631755352, + 0.09024551510810852 + ], + "credit_map": [ + 0.058059293776750565, + 0.027103299275040627, + 0.04466623440384865, + 0.03417881578207016, + 0.02879307046532631, + 0.04806405305862427, + 0.030170727521181107, + 0.037656791508197784, + 0.006124687381088734, + 0.003197313053533435, + 0.008604194968938828, + 0.003995510749518871, + 0.002551726531237364, + 0.0039050686173141003, + 0.003259466728195548, + 0.005941466428339481, + 0.021946167573332787, + 0.023027285933494568, + 0.01880946382880211, + 0.028204306960105896, + 0.011579900979995728, + 0.022505104541778564, + 0.008843690156936646, + 0.017481595277786255, + 0.004979029297828674, + 0.0030945432372391224, + 0.007501919753849506, + 0.0037677146028727293, + 0.003418933367356658, + 0.0030437931418418884, + 0.0030342540703713894, + 0.006044128909707069, + 0.009536482393741608, + 0.007897519506514072, + 0.007818866521120071, + 0.010338380001485348, + 0.01570849120616913, + 0.008816007524728775, + 0.006926637142896652, + 0.013613569550216198, + 0.008060360327363014, + 0.004959293641149998, + 0.00630164472386241, + 0.005024385638535023, + 0.00937560759484768, + 0.006021030247211456, + 0.0049023996107280254, + 0.004096898250281811, + 0.007784058339893818, + 0.004025042988359928, + 0.006328326649963856, + 0.0023658722639083862, + 0.002980364253744483, + 0.0047599999234080315, + 0.002125354018062353, + 0.0025502166245132685, + 0.02834608592092991, + 0.015545422211289406, + 0.03780175372958183, + 0.018884792923927307, + 0.021308686584234238, + 0.025322915986180305, + 0.015322582796216011, + 0.02585245668888092, + 0.003906548023223877, + 0.002710810862481594, + 0.0043266527354717255, + 0.004394089803099632, + 0.0023364597000181675, + 0.002645287197083235, + 0.0025377424899488688, + 0.0026418541092425585, + 0.014544107019901276, + 0.012261600233614445, + 0.015581903047859669, + 0.009796585887670517, + 0.007219432387501001, + 0.016237234696745872, + 0.006891274359077215, + 0.0077133760787546635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.03367666155099869 + } + ], + "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/0000022/group_4/index_tensor([8601, 8601])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_tensor([8601, 8601])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..93fe02861a04825a31ee0853fdce2e7c36d2cfd0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_tensor([8601, 8601])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle close-up captures a profoundly serene scene, rendered in a whimsical Ghibli style with muted colors. The focal point of the view is an ancient-looking boulder, its weathered, rounded surface almost entirely covered in a thick, velvety carpet of deep emerald moss, still shimmering with tiny, jewel-like droplets left behind by a recent rain shower. This solitary stone sits within a garden of meticulously raked white gravel. The smooth, pearly stones are arranged in gentle concentric waves, their damp surfaces now softly reflective. Above, the sky clears, fluffy, hand-painted white clouds revealing a soft azure. Arcing through this opening is a bright, vibrant rainbow, its distinctive bands of saturated color glowing warmly. The rising sun filters through the moist air, causing the moss to glow and softly casting the rainbow's vibrant colors onto the damp gravel below, creating a serene and magical atmosphere imbued with a tranquil sense of wonder.", + "seed": 5704, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_8601_seed_5704_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": -0.1285468996502459, + "vlm_logprob": -0.1285468996502459, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.1285468996502459, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5778816342353821 + ], + "opposite_logprob_tokens": [ + -0.8278816342353821 + ], + "logprob_reward": -0.5778816342353821, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5778816342353821, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23203310370445251, + 0.04039692133665085, + 0.08675765991210938, + 0.10964877158403397, + 0.09986111521720886, + 0.045822624117136, + 0.027558185160160065, + 0.09069855511188507, + 0.11090495437383652, + 0.15631811320781708 + ], + "credit_map": [ + 0.03491261601448059, + 0.039959486573934555, + 0.02366400510072708, + 0.024588175117969513, + 0.03281503543257713, + 0.034099824726581573, + 0.025852488353848457, + 0.016141489148139954, + 0.0053866347298026085, + 0.006154011003673077, + 0.004769166931509972, + 0.005063402000814676, + 0.005096790846437216, + 0.005255034193396568, + 0.0053955004550516605, + 0.0032763818744570017, + 0.0080689936876297, + 0.013288344256579876, + 0.006688263267278671, + 0.013622977770864964, + 0.007338763680309057, + 0.015974203124642372, + 0.014784757047891617, + 0.006991367321461439, + 0.010948927141726017, + 0.03391211852431297, + 0.00629225792363286, + 0.010239562951028347, + 0.009839425794780254, + 0.024433637037873268, + 0.008683614432811737, + 0.005299241282045841, + 0.009058896452188492, + 0.019328899681568146, + 0.0042119985446333885, + 0.018529487773776054, + 0.009723053313791752, + 0.01444999035447836, + 0.020303554832935333, + 0.004255235195159912, + 0.0035169196780771017, + 0.0036394991911947727, + 0.003513130359351635, + 0.012548359110951424, + 0.0035021137446165085, + 0.00384487840346992, + 0.012236076407134533, + 0.0030216516461223364, + 0.003657096764072776, + 0.0035745904315263033, + 0.0027359556406736374, + 0.004335131961852312, + 0.0028038250748068094, + 0.0033477614633738995, + 0.004491265397518873, + 0.0026125595904886723, + 0.006300746463239193, + 0.016198353841900826, + 0.003319642972201109, + 0.022384576499462128, + 0.007090821396559477, + 0.005536701064556837, + 0.02552921511232853, + 0.004338509868830442, + 0.010268901474773884, + 0.0119868703186512, + 0.004845782648772001, + 0.03195572271943092, + 0.006339711137115955, + 0.008686620742082596, + 0.0320466049015522, + 0.004774737637490034, + 0.008491545915603638, + 0.007569647394120693, + 0.005284330807626247, + 0.050633788108825684, + 0.007536814082413912, + 0.017969707027077675, + 0.05195537209510803, + 0.006876911036670208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.5778816342353821 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004646101035177708 + ], + "opposite_logprob_tokens": [ + -5.379646301269531 + ], + "logprob_reward": -0.004646101035177708, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004646101035177708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1812809705734253, + 0.07133765518665314, + 0.07829432189464569, + 0.18455220758914948, + 0.06634103506803513, + 0.07176890969276428, + 0.04117279127240181, + 0.09288148581981659, + 0.1769929677248001, + 0.03537767753005028 + ], + "credit_map": [ + 0.020876046270132065, + 0.023948173969984055, + 0.016925811767578125, + 0.026612628251314163, + 0.024776414036750793, + 0.023643752560019493, + 0.03176012635231018, + 0.012737996876239777, + 0.008292765356600285, + 0.009132696315646172, + 0.00753307668492198, + 0.01204767357558012, + 0.007822249084711075, + 0.00772958854213357, + 0.01271449401974678, + 0.00606510927900672, + 0.006495954468846321, + 0.012012559920549393, + 0.0042369733564555645, + 0.012631848454475403, + 0.006920743267983198, + 0.01806691102683544, + 0.01335326861590147, + 0.004576053004711866, + 0.015869474038481712, + 0.03567120432853699, + 0.015119743533432484, + 0.020194577053189278, + 0.01350263599306345, + 0.05128259211778641, + 0.020539959892630577, + 0.012372022494673729, + 0.005921503994613886, + 0.019139427691698074, + 0.0034982284996658564, + 0.009017747826874256, + 0.006537246517837048, + 0.008189251646399498, + 0.010851665399968624, + 0.0031859574373811483, + 0.006495644338428974, + 0.014506029896438122, + 0.004828120116144419, + 0.014597602188587189, + 0.006218482740223408, + 0.005831557791680098, + 0.016012992709875107, + 0.0032784724608063698, + 0.0034474635031074286, + 0.00878127757459879, + 0.0030333958566188812, + 0.006968308240175247, + 0.003008435945957899, + 0.005697552114725113, + 0.007158105261623859, + 0.003078250680118799, + 0.0080603938549757, + 0.022784365341067314, + 0.0025025000795722008, + 0.02022813819348812, + 0.007666455581784248, + 0.0065084961242973804, + 0.02193593978881836, + 0.003195187309756875, + 0.015254326164722443, + 0.026778582483530045, + 0.008412041701376438, + 0.03989074379205704, + 0.011432124301791191, + 0.023810308426618576, + 0.04297327250242233, + 0.008441543206572533, + 0.0033088827040046453, + 0.003020043484866619, + 0.0021884755697101355, + 0.007842371240258217, + 0.0024352283217012882, + 0.006750799249857664, + 0.007817302830517292, + 0.0020145715679973364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.004646101035177708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0205698125064373 + ], + "opposite_logprob_tokens": [ + -3.8955698013305664 + ], + "logprob_reward": -0.0205698125064373, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205698125064373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831217259168625, + 0.04184296727180481, + 0.028676453977823257, + 0.39643394947052, + 0.06109362095594406, + 0.05561283230781555, + 0.029740523546934128, + 0.06732375919818878, + 0.08671082556247711, + 0.049443263560533524 + ], + "credit_map": [ + 0.02096378244459629, + 0.025035589933395386, + 0.03081519901752472, + 0.02430691570043564, + 0.020390836521983147, + 0.020910188555717468, + 0.023412048816680908, + 0.017287185415625572, + 0.005742101930081844, + 0.006734116468578577, + 0.004964417777955532, + 0.0052587599493563175, + 0.005515501368790865, + 0.00507492758333683, + 0.005092907231301069, + 0.0034602368250489235, + 0.0030847848393023014, + 0.0036450603511184454, + 0.003106529824435711, + 0.0045033716596663, + 0.004226258490234613, + 0.002916146768257022, + 0.004654503893107176, + 0.0025398009456694126, + 0.050031084567308426, + 0.03634992986917496, + 0.04988991469144821, + 0.07334341108798981, + 0.033740486949682236, + 0.04488811641931534, + 0.06789523363113403, + 0.0402958057820797, + 0.006250164471566677, + 0.01594054326415062, + 0.004357847850769758, + 0.008695363067090511, + 0.006201297976076603, + 0.005794315133243799, + 0.01070363912731409, + 0.0031504530925303698, + 0.004953278694301844, + 0.007698053028434515, + 0.004227861762046814, + 0.012787152081727982, + 0.0040303491987288, + 0.0053281234577298164, + 0.013804829679429531, + 0.0027831955812871456, + 0.00313261104747653, + 0.007027890533208847, + 0.003073242725804448, + 0.004245226737111807, + 0.002846134826540947, + 0.002848586766049266, + 0.004336683079600334, + 0.0022301487624645233, + 0.006746962666511536, + 0.014332503080368042, + 0.0032682770397514105, + 0.012352715246379375, + 0.00616831798106432, + 0.008393298834562302, + 0.012676402926445007, + 0.0033852867782115936, + 0.00733528845012188, + 0.00918668694794178, + 0.003591521643102169, + 0.02220062166452408, + 0.00581956934183836, + 0.010844859294593334, + 0.02312900871038437, + 0.004603279754519463, + 0.0039766826666891575, + 0.002360113663598895, + 0.003705656621605158, + 0.015088658779859543, + 0.002945274580270052, + 0.0037884244229644537, + 0.015080327168107033, + 0.0024981312453746796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0205698125064373 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11760874092578888, + 0.05379706621170044, + 0.12029775232076645, + 0.03248092532157898, + 0.06975428760051727, + 0.12946176528930664, + 0.02185160666704178, + 0.026650995016098022, + 0.16240207850933075, + 0.2656947672367096 + ], + "credit_map": [ + 0.011273750104010105, + 0.010213162750005722, + 0.017556102946400642, + 0.023178594186902046, + 0.011159450747072697, + 0.010110949166119099, + 0.021176286041736603, + 0.012940456159412861, + 0.005230194889008999, + 0.007151203230023384, + 0.0061140586622059345, + 0.01034344919025898, + 0.006083220709115267, + 0.004483679309487343, + 0.009279248304665089, + 0.005112018436193466, + 0.010360938496887684, + 0.016983361914753914, + 0.01167030818760395, + 0.022200606763362885, + 0.00933957751840353, + 0.01683124341070652, + 0.022369004786014557, + 0.010542732663452625, + 0.004522831179201603, + 0.005443748086690903, + 0.002279842272400856, + 0.004483138211071491, + 0.004930294118821621, + 0.002999423071742058, + 0.005330717656761408, + 0.0024909365456551313, + 0.007708654273301363, + 0.014071367681026459, + 0.0038196290843188763, + 0.009200925938785076, + 0.005854683928191662, + 0.01378585584461689, + 0.010493923909962177, + 0.004819253925234079, + 0.010950409807264805, + 0.023371748626232147, + 0.005172437056899071, + 0.031129363924264908, + 0.008956946432590485, + 0.013590048998594284, + 0.031155819073319435, + 0.005134991370141506, + 0.0024309689179062843, + 0.005905311554670334, + 0.0015465349424630404, + 0.0028860520105808973, + 0.0022454322315752506, + 0.0028563346713781357, + 0.0026623369194567204, + 0.001318638096563518, + 0.002144333440810442, + 0.0021670451387763023, + 0.002274690195918083, + 0.005513588432222605, + 0.002751655410975218, + 0.004003351088613272, + 0.00585308438166976, + 0.0019432510016486049, + 0.013036145828664303, + 0.012109381146728992, + 0.007382833398878574, + 0.05016638711094856, + 0.008353748358786106, + 0.01197334099560976, + 0.052341315895318985, + 0.007038949988782406, + 0.015017643570899963, + 0.013898634351789951, + 0.010217607021331787, + 0.0824967622756958, + 0.012734406627714634, + 0.035045042634010315, + 0.08404670655727386, + 0.012237992137670517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.005946918856352568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369003161787987 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.03369003161787987, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369003161787987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11476513743400574, + 0.03324466198682785, + 0.03269628807902336, + 0.25891777873039246, + 0.0641101822257042, + 0.05847024545073509, + 0.024222364649176598, + 0.04463554918766022, + 0.11502567678689957, + 0.2539121210575104 + ], + "credit_map": [ + 0.011622116900980473, + 0.009463724680244923, + 0.010538388974964619, + 0.02316187508404255, + 0.013472861610352993, + 0.010828281752765179, + 0.024919873103499413, + 0.010758014395833015, + 0.004421075340360403, + 0.0046518705785274506, + 0.0028144356328994036, + 0.004965485539287329, + 0.00429587671533227, + 0.0041817715391516685, + 0.005133252590894699, + 0.0027808917220681906, + 0.0024926546029746532, + 0.002821357222273946, + 0.0025753192603588104, + 0.006522926967591047, + 0.005239951424300671, + 0.003357690991833806, + 0.006736861541867256, + 0.0029495249036699533, + 0.03395290672779083, + 0.028353706002235413, + 0.03703945130109787, + 0.048376504331827164, + 0.020963920280337334, + 0.022609077394008636, + 0.03877630457282066, + 0.028845887631177902, + 0.005260718520730734, + 0.017147677019238472, + 0.003754631383344531, + 0.008333944715559483, + 0.00520390085875988, + 0.01135878637433052, + 0.008913316763937473, + 0.00413720915094018, + 0.0032603300642222166, + 0.007542459759861231, + 0.0023540647234767675, + 0.0167701356112957, + 0.0034018107689917088, + 0.004822811111807823, + 0.017321854829788208, + 0.002996778814122081, + 0.001411101664416492, + 0.0035074851475656033, + 0.001290519256144762, + 0.00590859167277813, + 0.0013286458561196923, + 0.0028444549534469843, + 0.005832890048623085, + 0.0020986765157431364, + 0.0035171466879546642, + 0.010526045225560665, + 0.0021774659398943186, + 0.008836211636662483, + 0.004230511840432882, + 0.003373492509126663, + 0.009402402676641941, + 0.0025722712744027376, + 0.009126137010753155, + 0.015208355151116848, + 0.006561052054166794, + 0.027901973575353622, + 0.007680678274482489, + 0.010989365167915821, + 0.03214387595653534, + 0.005414239596575499, + 0.018261821940541267, + 0.014809805899858475, + 0.009254286997020245, + 0.08400900661945343, + 0.011740860529243946, + 0.01948682591319084, + 0.08401008695363998, + 0.012339429929852486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.03369003161787987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a profoundly serene scene, rendered in a whimsical Ghibli style with muted colors. The focal point of the view is an ancient-looking boulder, its weathered, rounded surface almost entirely covered in a thick, velvety carpet of deep emerald moss, still shimmering with tiny, jewel-like droplets left behind by a recent rain shower. This solitary stone sits within a garden of meticulously raked white gravel. The smooth, pearly stones are arranged in gentle concentric waves, their damp surfaces now softly reflective. Above, the sky clears, fluffy, hand-painted white clouds revealing a soft azure. Arcing through this opening is a bright, vibrant rainbow, its distinctive bands of saturated color glowing warmly. The rising sun filters through the moist air, causing the moss to glow and softly casting the rainbow's vibrant colors onto the damp gravel below, creating a serene and magical atmosphere imbued with a tranquil sense of wonder.", + "seed": 5705, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_8601_seed_5705_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": -0.057543596858158706, + "vlm_logprob": -0.057543596858158706, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.057543596858158706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20578217506408691 + ], + "opposite_logprob_tokens": [ + -1.705782175064087 + ], + "logprob_reward": -0.20578217506408691, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20578217506408691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2365775853395462, + 0.040659885853528976, + 0.02582535333931446, + 0.07398734241724014, + 0.14119023084640503, + 0.19323298335075378, + 0.1200156956911087, + 0.04279569163918495, + 0.06611493229866028, + 0.059600260108709335 + ], + "credit_map": [ + 0.040107499808073044, + 0.03884988650679588, + 0.018653035163879395, + 0.019629940390586853, + 0.025074895471334457, + 0.05164110288023949, + 0.016370443627238274, + 0.026250790804624557, + 0.005149151664227247, + 0.006366118788719177, + 0.0048162671737372875, + 0.004135154187679291, + 0.007810697425156832, + 0.005158847197890282, + 0.003972588572651148, + 0.0032510601449757814, + 0.0039128707721829414, + 0.0035539763048291206, + 0.0030132397077977657, + 0.002811441430822015, + 0.0026933581102639437, + 0.00396546209231019, + 0.002278747735545039, + 0.003596259281039238, + 0.01146872341632843, + 0.009300352074205875, + 0.007697166875004768, + 0.0061842529103159904, + 0.007979155518114567, + 0.014751069247722626, + 0.006080025807023048, + 0.010526604950428009, + 0.015711994841694832, + 0.013067602179944515, + 0.01058485358953476, + 0.043489132076501846, + 0.013759068213403225, + 0.007399566005915403, + 0.030157780274748802, + 0.007020239252597094, + 0.03363155201077461, + 0.018766140565276146, + 0.011191553436219692, + 0.05698930472135544, + 0.014329679310321808, + 0.011999948881566525, + 0.03978000581264496, + 0.006544816307723522, + 0.017206629738211632, + 0.016559720039367676, + 0.010818480513989925, + 0.022733889520168304, + 0.01657654158771038, + 0.010583406314253807, + 0.01725188083946705, + 0.008285156451165676, + 0.007524208165705204, + 0.004113039001822472, + 0.0054975044913589954, + 0.005051606334745884, + 0.003828930202871561, + 0.008956152014434338, + 0.004285025410354137, + 0.003539229976013303, + 0.007138566114008427, + 0.007473082281649113, + 0.00437379302456975, + 0.016328923404216766, + 0.006229399237781763, + 0.008928768336772919, + 0.011710293591022491, + 0.003932110499590635, + 0.009867572225630283, + 0.006337776314467192, + 0.008347622118890285, + 0.011349475011229515, + 0.0043676868081092834, + 0.005138843785971403, + 0.00858735665678978, + 0.005603929981589317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.20578217506408691 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794983081519604 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009794983081519604, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794983081519604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2262144833803177, + 0.19932448863983154, + 0.06014988198876381, + 0.0635804608464241, + 0.05006438121199608, + 0.08713051676750183, + 0.03555985540151596, + 0.06472024321556091, + 0.05023210123181343, + 0.16302363574504852 + ], + "credit_map": [ + 0.02590191178023815, + 0.02275898866355419, + 0.0418563112616539, + 0.03577189892530441, + 0.018808091059327126, + 0.025421947240829468, + 0.028745708987116814, + 0.026949595659971237, + 0.023630158975720406, + 0.021402692422270775, + 0.035384055227041245, + 0.030162539333105087, + 0.016734950244426727, + 0.02306916005909443, + 0.02463552914559841, + 0.024305379018187523, + 0.007892305962741375, + 0.0061442116275429726, + 0.010030231438577175, + 0.006503662094473839, + 0.007400298956781626, + 0.007363473996520042, + 0.007341619115322828, + 0.007474072277545929, + 0.007093019783496857, + 0.003869557986035943, + 0.007454134523868561, + 0.016351819038391113, + 0.004164894111454487, + 0.007981055416166782, + 0.011045479215681553, + 0.005620501935482025, + 0.004933962598443031, + 0.005584754049777985, + 0.005278682801872492, + 0.013180908747017384, + 0.005060739815235138, + 0.0037362375296652317, + 0.009559093043208122, + 0.0027300009969621897, + 0.008521591313183308, + 0.00671184528619051, + 0.010516819544136524, + 0.024200551211833954, + 0.006818273104727268, + 0.0077933380380272865, + 0.017161374911665916, + 0.005406717304140329, + 0.004165508318692446, + 0.0041494243778288364, + 0.0055989911779761314, + 0.006015936844050884, + 0.003346402198076248, + 0.0037618696223944426, + 0.0051893554627895355, + 0.003332365769892931, + 0.011265093460679054, + 0.006587019190192223, + 0.007738505955785513, + 0.008699610829353333, + 0.004956166259944439, + 0.012347439303994179, + 0.0069547682069242, + 0.006171636749058962, + 0.005503189284354448, + 0.0051286546513438225, + 0.004586744587868452, + 0.012571227736771107, + 0.0037692575715482235, + 0.006972290575504303, + 0.008916085585951805, + 0.002784650307148695, + 0.022127317264676094, + 0.013582599349319935, + 0.026493921875953674, + 0.032263677567243576, + 0.010057426989078522, + 0.01191078219562769, + 0.02878294140100479, + 0.01780497282743454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.009794983081519604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203024834394455 + ], + "opposite_logprob_tokens": [ + -2.81203031539917 + ], + "logprob_reward": -0.06203024834394455, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203024834394455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13997147977352142, + 0.04674419388175011, + 0.05603503808379173, + 0.057290587574243546, + 0.0734640508890152, + 0.07325632870197296, + 0.11786967515945435, + 0.14821983873844147, + 0.04277642443776131, + 0.24437235295772552 + ], + "credit_map": [ + 0.016403928399086, + 0.020708534866571426, + 0.015778057277202606, + 0.0156132522970438, + 0.01338402833789587, + 0.02144554816186428, + 0.0158709529787302, + 0.020767180249094963, + 0.005403486080467701, + 0.005643339827656746, + 0.007415646221488714, + 0.006643086206167936, + 0.005628334823995829, + 0.005147725809365511, + 0.005749909672886133, + 0.005112667568027973, + 0.007908791303634644, + 0.006441420875489712, + 0.004980702418833971, + 0.0055381180718541145, + 0.008770039305090904, + 0.010307494550943375, + 0.004728853236883879, + 0.007359624374657869, + 0.00684584304690361, + 0.005728262010961771, + 0.004517998546361923, + 0.014431833289563656, + 0.005037022288888693, + 0.004627799615263939, + 0.00943128950893879, + 0.0066705429926514626, + 0.008150841109454632, + 0.008893676102161407, + 0.007227639202028513, + 0.01918216608464718, + 0.006828688085079193, + 0.005876982118934393, + 0.013424056582152843, + 0.0038800102192908525, + 0.010247672908008099, + 0.007765491027384996, + 0.004957705270498991, + 0.01974448561668396, + 0.007653453387320042, + 0.004598897881805897, + 0.014662256464362144, + 0.0036263742949813604, + 0.018236396834254265, + 0.013697977177798748, + 0.00953904539346695, + 0.02407732419669628, + 0.015079284086823463, + 0.011469232849776745, + 0.01811824180185795, + 0.007652175612747669, + 0.026399316266179085, + 0.014887724071741104, + 0.013044786639511585, + 0.024266403168439865, + 0.013346954248845577, + 0.026572998613119125, + 0.01788649894297123, + 0.011815181002020836, + 0.0060730017721652985, + 0.004300259985029697, + 0.005308851599693298, + 0.004268244374543428, + 0.003758231643587351, + 0.010663996450603008, + 0.0034998413175344467, + 0.004903998225927353, + 0.042896684259176254, + 0.032400667667388916, + 0.022951656952500343, + 0.024849358946084976, + 0.035071104764938354, + 0.04759513959288597, + 0.02019849419593811, + 0.018409263342618942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.06203024834394455 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024876620154827833 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024876620154827833, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024876620154827833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15165285766124725, + 0.05539853498339653, + 0.04610413312911987, + 0.028645621612668037, + 0.10922587662935257, + 0.16426438093185425, + 0.09133438020944595, + 0.08909069746732712, + 0.15559718012809753, + 0.10868632048368454 + ], + "credit_map": [ + 0.01992844231426716, + 0.019390316680073738, + 0.019475383684039116, + 0.018556242808699608, + 0.022539300844073296, + 0.02281373180449009, + 0.015541395172476768, + 0.013408055528998375, + 0.006094301585108042, + 0.005612771958112717, + 0.00802891980856657, + 0.008577542379498482, + 0.007306964602321386, + 0.0049473256804049015, + 0.00772046996280551, + 0.007110241334885359, + 0.006459202151745558, + 0.005194412544369698, + 0.005827120505273342, + 0.005431683734059334, + 0.004624308552592993, + 0.008291634730994701, + 0.004965855274349451, + 0.005309917498379946, + 0.004181771073490381, + 0.002516436856240034, + 0.0026138813700526953, + 0.005635196343064308, + 0.00257691228762269, + 0.003148950170725584, + 0.004671853501349688, + 0.0033006174489855766, + 0.01409047283232212, + 0.010001116432249546, + 0.006786391604691744, + 0.036996856331825256, + 0.004792747087776661, + 0.005584878847002983, + 0.025895221158862114, + 0.005078186746686697, + 0.01916486583650112, + 0.027180764824151993, + 0.01372185442596674, + 0.025959277525544167, + 0.030618302524089813, + 0.016577161848545074, + 0.019292686134576797, + 0.011749454773962498, + 0.009344219230115414, + 0.012406079098582268, + 0.006493799388408661, + 0.021071594208478928, + 0.0136850755661726, + 0.006306502968072891, + 0.015724001452326775, + 0.006303108762949705, + 0.01467590406537056, + 0.008309729397296906, + 0.010030381381511688, + 0.018228722736239433, + 0.006166183855384588, + 0.011349781416356564, + 0.013881590217351913, + 0.006448410917073488, + 0.017909586429595947, + 0.015937797725200653, + 0.013011844828724861, + 0.034328069537878036, + 0.011991084553301334, + 0.026526611298322678, + 0.023956116288900375, + 0.01193607971072197, + 0.01806877925992012, + 0.012782762758433819, + 0.01234592217952013, + 0.01379624754190445, + 0.013367557898163795, + 0.015319901518523693, + 0.011768042109906673, + 0.011237100698053837 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.0024876620154827833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007622915785759687 + ], + "opposite_logprob_tokens": [ + -4.882622718811035 + ], + "logprob_reward": -0.007622915785759687, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007622915785759687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28510016202926636, + 0.10123376548290253, + 0.03860758617520332, + 0.07081092894077301, + 0.02506098710000515, + 0.08321447670459747, + 0.05531392991542816, + 0.10237710922956467, + 0.15433625876903534, + 0.08394480496644974 + ], + "credit_map": [ + 0.040319327265024185, + 0.04555223509669304, + 0.0415182001888752, + 0.03814151883125305, + 0.030129536986351013, + 0.03228497877717018, + 0.031416378915309906, + 0.025737982243299484, + 0.016309168189764023, + 0.013401541858911514, + 0.014121092855930328, + 0.012761642225086689, + 0.011936900205910206, + 0.013292661868035793, + 0.010719711892306805, + 0.008691046386957169, + 0.006041897926479578, + 0.0051054279319942, + 0.003898513736203313, + 0.00469714030623436, + 0.004672299139201641, + 0.004199352581053972, + 0.004756831098347902, + 0.005236123222857714, + 0.009149796329438686, + 0.005612371489405632, + 0.006371676456183195, + 0.020049039274454117, + 0.005086553283035755, + 0.005630644038319588, + 0.014020176604390144, + 0.0048906817100942135, + 0.002195848850533366, + 0.0022605492267757654, + 0.0028072320856153965, + 0.005583096761256456, + 0.003180354367941618, + 0.0026448334101587534, + 0.004307679831981659, + 0.0020813937298953533, + 0.010291140526533127, + 0.006650328636169434, + 0.008326329290866852, + 0.024255331605672836, + 0.005248210392892361, + 0.006985211279243231, + 0.017101066187024117, + 0.004356859717518091, + 0.008073963224887848, + 0.007866371423006058, + 0.006078349892050028, + 0.006323195528239012, + 0.009256836958229542, + 0.007942507043480873, + 0.005203367210924625, + 0.004569336771965027, + 0.014285548590123653, + 0.01087520644068718, + 0.013343624770641327, + 0.013268318958580494, + 0.008838014677166939, + 0.02186538279056549, + 0.009842929430305958, + 0.0100580845028162, + 0.01712176948785782, + 0.013317319564521313, + 0.012028860859572887, + 0.044882435351610184, + 0.009396220557391644, + 0.018398767337203026, + 0.03146528825163841, + 0.00772560341283679, + 0.010199460200965405, + 0.00850546732544899, + 0.016071341931819916, + 0.013296901248395443, + 0.006461564917117357, + 0.007703331299126148, + 0.012197841890156269, + 0.009508893825113773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007622915785759687 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a profoundly serene scene, rendered in a whimsical Ghibli style with muted colors. The focal point of the view is an ancient-looking boulder, its weathered, rounded surface almost entirely covered in a thick, velvety carpet of deep emerald moss, still shimmering with tiny, jewel-like droplets left behind by a recent rain shower. This solitary stone sits within a garden of meticulously raked white gravel. The smooth, pearly stones are arranged in gentle concentric waves, their damp surfaces now softly reflective. Above, the sky clears, fluffy, hand-painted white clouds revealing a soft azure. Arcing through this opening is a bright, vibrant rainbow, its distinctive bands of saturated color glowing warmly. The rising sun filters through the moist air, causing the moss to glow and softly casting the rainbow's vibrant colors onto the damp gravel below, creating a serene and magical atmosphere imbued with a tranquil sense of wonder.", + "seed": 5706, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_8601_seed_5706_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": -0.03306913189589977, + "vlm_logprob": -0.03306913189589977, + "vlm_yes_no_margin": 4.775000023841858, + "vlm_token_credit": -0.03306913189589977, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1451658308506012 + ], + "opposite_logprob_tokens": [ + -2.0201659202575684 + ], + "logprob_reward": -0.1451658308506012, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1451658308506012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26306629180908203, + 0.10449456423521042, + 0.036340467631816864, + 0.050884973257780075, + 0.040355268865823746, + 0.09841611236333847, + 0.06603404879570007, + 0.11706674844026566, + 0.13549071550369263, + 0.08785079419612885 + ], + "credit_map": [ + 0.02346641756594181, + 0.056682784110307693, + 0.021583503112196922, + 0.03444218635559082, + 0.025685284286737442, + 0.03248711675405502, + 0.024974366649985313, + 0.04374462738633156, + 0.009477528743445873, + 0.015042838640511036, + 0.011572049930691719, + 0.011251568794250488, + 0.012517738156020641, + 0.014115610159933567, + 0.01314047072082758, + 0.01737675815820694, + 0.003943719435483217, + 0.007598737720400095, + 0.003499547950923443, + 0.004568963777273893, + 0.0027032066136598587, + 0.005000242497771978, + 0.003915490582585335, + 0.005110557656735182, + 0.00546651054173708, + 0.009299840778112411, + 0.004977357108145952, + 0.004881958942860365, + 0.004187485668808222, + 0.007986174896359444, + 0.0030577301513403654, + 0.011027914471924305, + 0.00466477544978261, + 0.004377807956188917, + 0.004197642207145691, + 0.0032452333252876997, + 0.003663192503154278, + 0.006163208745419979, + 0.0045450832694768906, + 0.009498321451246738, + 0.0091443732380867, + 0.01635252870619297, + 0.009266454726457596, + 0.006334853824228048, + 0.017560863867402077, + 0.015913179144263268, + 0.01000006403774023, + 0.01384380180388689, + 0.006169391795992851, + 0.005857027135789394, + 0.004947652108967304, + 0.0030462525319308043, + 0.006120564416050911, + 0.012218194082379341, + 0.005448371171951294, + 0.022226596251130104, + 0.009738394990563393, + 0.01697913557291031, + 0.013282646425068378, + 0.008797062560915947, + 0.01528710313141346, + 0.0186949260532856, + 0.010189635679125786, + 0.024097846820950508, + 0.016380896791815758, + 0.011011906899511814, + 0.012856998480856419, + 0.0086764981970191, + 0.023344196379184723, + 0.023862291127443314, + 0.01307874545454979, + 0.026279186829924583, + 0.008659303188323975, + 0.008865856565535069, + 0.009869632311165333, + 0.007127760909497738, + 0.01105012558400631, + 0.012935265898704529, + 0.011525935493409634, + 0.017816921696066856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.1451658308506012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011863820254802704 + ], + "opposite_logprob_tokens": [ + -6.751186370849609 + ], + "logprob_reward": -0.0011863820254802704, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011863820254802704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3128000795841217, + 0.17532867193222046, + 0.059138212352991104, + 0.034559883177280426, + 0.048916734755039215, + 0.16808564960956573, + 0.05395199730992317, + 0.05724392086267471, + 0.037679802626371384, + 0.0522950142621994 + ], + "credit_map": [ + 0.02718210034072399, + 0.05184872820973396, + 0.03014882281422615, + 0.03307649493217468, + 0.021242881193757057, + 0.042023878544569016, + 0.03252977877855301, + 0.07474742084741592, + 0.013154773972928524, + 0.023614957928657532, + 0.021008504554629326, + 0.015525381080806255, + 0.020489545539021492, + 0.02365475706756115, + 0.018788103014230728, + 0.03909265622496605, + 0.005962500348687172, + 0.010553468950092793, + 0.006971622817218304, + 0.005896515212953091, + 0.005277891643345356, + 0.007397763431072235, + 0.0062269363552331924, + 0.01085151731967926, + 0.003945885691791773, + 0.004939734935760498, + 0.0030583601910620928, + 0.003472567070275545, + 0.0032589163165539503, + 0.0050845337100327015, + 0.0029051469173282385, + 0.007894742302596569, + 0.006882993970066309, + 0.008478603325784206, + 0.0046651712618768215, + 0.005377546884119511, + 0.0039194622077047825, + 0.007119082845747471, + 0.003301858203485608, + 0.009172017686069012, + 0.014062917791306973, + 0.0392695814371109, + 0.01761971227824688, + 0.018992723897099495, + 0.02061871998012066, + 0.02058585174381733, + 0.015947390347719193, + 0.020988766103982925, + 0.004818376619368792, + 0.008669357746839523, + 0.00522221252322197, + 0.003853557398542762, + 0.006744416896253824, + 0.008539482019841671, + 0.004862139467149973, + 0.0112424585968256, + 0.005455113481730223, + 0.008110197260975838, + 0.007041417993605137, + 0.004688207525759935, + 0.005850870627909899, + 0.007594090420752764, + 0.005457061342895031, + 0.013046962209045887, + 0.003524575848132372, + 0.006839343346655369, + 0.004606659524142742, + 0.003287080442532897, + 0.005301992874592543, + 0.00499514676630497, + 0.0030840677209198475, + 0.006040940061211586, + 0.00516407610848546, + 0.007575885392725468, + 0.005218698643147945, + 0.0045515322126448154, + 0.009164126589894295, + 0.007180655840784311, + 0.005565510131418705, + 0.007874533534049988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0011863820254802704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004631625022739172 + ], + "opposite_logprob_tokens": [ + -5.379631519317627 + ], + "logprob_reward": -0.004631625022739172, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004631625022739172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19917874038219452, + 0.18004761636257172, + 0.2211333066225052, + 0.03995895013213158, + 0.10961318761110306, + 0.0578751377761364, + 0.03127572685480118, + 0.07217863202095032, + 0.02943866141140461, + 0.059300076216459274 + ], + "credit_map": [ + 0.02548114024102688, + 0.020145907998085022, + 0.01643216796219349, + 0.013298282399773598, + 0.013619610108435154, + 0.03432389721274376, + 0.013507794588804245, + 0.062369946390390396, + 0.019896728917956352, + 0.010467584244906902, + 0.010977406986057758, + 0.007806754671037197, + 0.010532382875680923, + 0.03901755437254906, + 0.009307410567998886, + 0.07204177975654602, + 0.024962827563285828, + 0.012452110648155212, + 0.01192264724522829, + 0.008600444532930851, + 0.01466467883437872, + 0.04901454970240593, + 0.00955356564372778, + 0.08996248245239258, + 0.0052954270504415035, + 0.0029459623619914055, + 0.0026962996926158667, + 0.0025665676221251488, + 0.002708795014768839, + 0.007749097887426615, + 0.002378757344558835, + 0.013618041761219501, + 0.012526252306997776, + 0.006332827731966972, + 0.0066846138797700405, + 0.005730253644287586, + 0.008874516002833843, + 0.022732499986886978, + 0.004979357123374939, + 0.041752856224775314, + 0.004903342109173536, + 0.006153570022433996, + 0.004276907071471214, + 0.0036823733244091272, + 0.011224773712456226, + 0.01099634449928999, + 0.006072903051972389, + 0.010564925149083138, + 0.003419618122279644, + 0.0030385623686015606, + 0.0031459678430110216, + 0.002017795108258724, + 0.0038024892564862967, + 0.005239255726337433, + 0.0028866964858025312, + 0.007725341245532036, + 0.005705280229449272, + 0.008642353117465973, + 0.007994652725756168, + 0.0045391167514026165, + 0.01019321009516716, + 0.011648335494101048, + 0.006158515810966492, + 0.01729716546833515, + 0.003482967847958207, + 0.002578320214524865, + 0.002448443789035082, + 0.0019635979551821947, + 0.005103083327412605, + 0.00459820544347167, + 0.0034869227092713118, + 0.005777119658887386, + 0.007154098246246576, + 0.004725700709968805, + 0.005659669172018766, + 0.00394509406760335, + 0.010594744235277176, + 0.009848386980593204, + 0.007102371193468571, + 0.0102700125426054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004631625022739172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630723994225264 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630723994225264, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630723994225264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12230581790208817, + 0.03996512293815613, + 0.023330634459853172, + 0.02224200963973999, + 0.01877324841916561, + 0.06756643950939178, + 0.12325672060251236, + 0.0956350713968277, + 0.1281876116991043, + 0.35873734951019287 + ], + "credit_map": [ + 0.012826364487409592, + 0.015094826929271221, + 0.011696981266140938, + 0.017096957191824913, + 0.012915183790028095, + 0.015634378418326378, + 0.015420056879520416, + 0.02162107452750206, + 0.0038887406699359417, + 0.005136100575327873, + 0.0037688876036554575, + 0.004078855738043785, + 0.004929408896714449, + 0.005444745533168316, + 0.004219553433358669, + 0.008498831652104855, + 0.0022681166883558035, + 0.0030630677938461304, + 0.002433568239212036, + 0.0015428694896399975, + 0.0017925803549587727, + 0.0036992691457271576, + 0.001880817930214107, + 0.006650345865637064, + 0.002586187096312642, + 0.0029004705138504505, + 0.0015252670273184776, + 0.0014127141330391169, + 0.002201839815825224, + 0.003665501018986106, + 0.0016015442088246346, + 0.006348486058413982, + 0.002147407503798604, + 0.0032952374313026667, + 0.0016638626111671329, + 0.00235932320356369, + 0.0019311826908960938, + 0.0020602275617420673, + 0.0021928795613348484, + 0.00312312925234437, + 0.0063536944799125195, + 0.00853785127401352, + 0.0049249860458076, + 0.004218669608235359, + 0.013717938214540482, + 0.013511361554265022, + 0.008091463707387447, + 0.008210479281842709, + 0.016008421778678894, + 0.011259589344263077, + 0.012528905645012856, + 0.009171238169074059, + 0.015647882595658302, + 0.018094250932335854, + 0.012755893170833588, + 0.02779053896665573, + 0.009449698962271214, + 0.007237850222736597, + 0.009425665251910686, + 0.0065596094354987144, + 0.008672638796269894, + 0.01740807108581066, + 0.009317666292190552, + 0.02756386622786522, + 0.012132720090448856, + 0.018465830013155937, + 0.009177541360259056, + 0.008736333809792995, + 0.026640567928552628, + 0.01965469866991043, + 0.016323182731866837, + 0.017056729644536972, + 0.043497081845998764, + 0.033764515072107315, + 0.034305594861507416, + 0.023061349987983704, + 0.06341584771871567, + 0.06235485523939133, + 0.0395803302526474, + 0.05875777453184128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.007630723994225264 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731097586452961 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731097586452961, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731097586452961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25004202127456665, + 0.08748205006122589, + 0.04025077074766159, + 0.050906673073768616, + 0.05372720584273338, + 0.11323204636573792, + 0.18141868710517883, + 0.1381729692220688, + 0.03575469180941582, + 0.04901288077235222 + ], + "credit_map": [ + 0.02897653914988041, + 0.021116282790899277, + 0.016133161261677742, + 0.01402029674500227, + 0.017670443281531334, + 0.04465794563293457, + 0.01906914822757244, + 0.08839816600084305, + 0.009213817305862904, + 0.008098291233181953, + 0.00654254388064146, + 0.007865307852625847, + 0.012227172031998634, + 0.01398204080760479, + 0.009921183809638023, + 0.01963169313967228, + 0.003828982822597027, + 0.0046077147126197815, + 0.005123271606862545, + 0.004129085224121809, + 0.0033367376308888197, + 0.0061528985388576984, + 0.0037883773911744356, + 0.009283700957894325, + 0.006224615499377251, + 0.006835253443568945, + 0.005851196590811014, + 0.005252189468592405, + 0.004599215462803841, + 0.006868736352771521, + 0.0045057120732963085, + 0.010769747197628021, + 0.005866525694727898, + 0.009332503192126751, + 0.0054106274619698524, + 0.005338550545275211, + 0.00459287641569972, + 0.00696199806407094, + 0.004607531242072582, + 0.011616588570177555, + 0.010490135289728642, + 0.020629070699214935, + 0.009343203157186508, + 0.011651923879981041, + 0.01422739215195179, + 0.01480347104370594, + 0.012340772897005081, + 0.019746065139770508, + 0.019529905170202255, + 0.013947173953056335, + 0.013775509782135487, + 0.013687021099030972, + 0.018564175814390182, + 0.029977241531014442, + 0.014860148541629314, + 0.05707748979330063, + 0.012142060324549675, + 0.015481988899409771, + 0.010870725847780704, + 0.008430114947259426, + 0.02946796640753746, + 0.02456437610089779, + 0.012412267737090588, + 0.02480345219373703, + 0.0033059341367334127, + 0.005430144723504782, + 0.005320390686392784, + 0.003158832900226116, + 0.00434712041169405, + 0.005065057426691055, + 0.0032642544247210026, + 0.005862955003976822, + 0.004989295266568661, + 0.005278388969600201, + 0.005259244702756405, + 0.0036149495281279087, + 0.005860723555088043, + 0.008398035541176796, + 0.0049881841987371445, + 0.010624054819345474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.006731097586452961 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up captures a profoundly serene scene, rendered in a whimsical Ghibli style with muted colors. The focal point of the view is an ancient-looking boulder, its weathered, rounded surface almost entirely covered in a thick, velvety carpet of deep emerald moss, still shimmering with tiny, jewel-like droplets left behind by a recent rain shower. This solitary stone sits within a garden of meticulously raked white gravel. The smooth, pearly stones are arranged in gentle concentric waves, their damp surfaces now softly reflective. Above, the sky clears, fluffy, hand-painted white clouds revealing a soft azure. Arcing through this opening is a bright, vibrant rainbow, its distinctive bands of saturated color glowing warmly. The rising sun filters through the moist air, causing the moss to glow and softly casting the rainbow's vibrant colors onto the damp gravel below, creating a serene and magical atmosphere imbued with a tranquil sense of wonder.", + "seed": 5707, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_4/index_8601_seed_5707_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": -0.08130882317200303, + "vlm_logprob": -0.08130882317200303, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.08130882317200303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3174034357070923 + ], + "opposite_logprob_tokens": [ + -1.3174034357070923 + ], + "logprob_reward": -0.3174034357070923, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3174034357070923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4105410873889923, + 0.0705358237028122, + 0.04284434765577316, + 0.057058513164520264, + 0.09793908894062042, + 0.09933579713106155, + 0.09574463218450546, + 0.036428797990083694, + 0.060991864651441574, + 0.028580041602253914 + ], + "credit_map": [ + 0.0269709974527359, + 0.04649295657873154, + 0.04217730462551117, + 0.05843905359506607, + 0.04059309512376785, + 0.06710787862539291, + 0.045713622123003006, + 0.08304619044065475, + 0.005504336673766375, + 0.007678034249693155, + 0.005675615277141333, + 0.015363273210823536, + 0.004929643124341965, + 0.012238837778568268, + 0.010083948262035847, + 0.00906213466078043, + 0.00431599747389555, + 0.007264890242367983, + 0.004367385525256395, + 0.004502140451222658, + 0.006087005138397217, + 0.007856277748942375, + 0.0031728169415146112, + 0.005277836229652166, + 0.004069723188877106, + 0.009412405081093311, + 0.005899342708289623, + 0.008590015582740307, + 0.006945818196982145, + 0.009181368164718151, + 0.006071793846786022, + 0.006888048257678747, + 0.009185283444821835, + 0.012373634614050388, + 0.00723002664744854, + 0.01758197881281376, + 0.008717966265976429, + 0.019812870770692825, + 0.01178828813135624, + 0.011249036528170109, + 0.009882492013275623, + 0.014354678802192211, + 0.007609211839735508, + 0.011676616035401821, + 0.011319567449390888, + 0.020330220460891724, + 0.009097781032323837, + 0.015065224841237068, + 0.008255789056420326, + 0.014810596592724323, + 0.007226059213280678, + 0.014162776060402393, + 0.010976778343319893, + 0.01710698753595352, + 0.01077465433627367, + 0.012430990114808083, + 0.002856173552572727, + 0.006931767798960209, + 0.00401655025780201, + 0.005563539452850819, + 0.004863888490945101, + 0.003764956956729293, + 0.00516285328194499, + 0.00326906586997211, + 0.004491180181503296, + 0.01155443862080574, + 0.006050214171409607, + 0.010222653858363628, + 0.008162282407283783, + 0.006431350484490395, + 0.007969620637595654, + 0.006110126618295908, + 0.0022126752883195877, + 0.0059595233760774136, + 0.0032456195913255215, + 0.003691300516948104, + 0.004392239265143871, + 0.0029359173495322466, + 0.0037153272423893213, + 0.002427438274025917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.3174034357070923 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025015748105943203 + ], + "opposite_logprob_tokens": [ + -6.002501487731934 + ], + "logprob_reward": -0.0025015748105943203, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025015748105943203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.360667884349823, + 0.16104820370674133, + 0.10914883017539978, + 0.07829995453357697, + 0.0633745864033699, + 0.03629957139492035, + 0.03367829695343971, + 0.07339856773614883, + 0.03856953606009483, + 0.045514605939388275 + ], + "credit_map": [ + 0.022156739607453346, + 0.037606265395879745, + 0.0381162166595459, + 0.08924362808465958, + 0.03003501705825329, + 0.047550883144140244, + 0.06484781205654144, + 0.031111298128962517, + 0.011143138632178307, + 0.02181040495634079, + 0.018749374896287918, + 0.031232275068759918, + 0.01987661048769951, + 0.015067217871546745, + 0.028549732640385628, + 0.014619441702961922, + 0.008606752380728722, + 0.021580547094345093, + 0.012438138946890831, + 0.014089195989072323, + 0.01691703125834465, + 0.01000300608575344, + 0.014867095276713371, + 0.010647070594131947, + 0.006319983396679163, + 0.014628435485064983, + 0.008687554858624935, + 0.0112247783690691, + 0.011771641671657562, + 0.008547434583306313, + 0.009828099049627781, + 0.007292028982192278, + 0.005382612347602844, + 0.009811598807573318, + 0.005569805856794119, + 0.008758027106523514, + 0.006663888227194548, + 0.010873041115701199, + 0.008304856717586517, + 0.008010756224393845, + 0.0035200517158955336, + 0.005873378366231918, + 0.0034510877449065447, + 0.004900249186903238, + 0.004146169871091843, + 0.005908583290874958, + 0.0031885334756225348, + 0.005311519838869572, + 0.002562835579738021, + 0.005619417876005173, + 0.003930448554456234, + 0.005684516858309507, + 0.0046005286276340485, + 0.0037636361084878445, + 0.004338664468377829, + 0.003178250975906849, + 0.004910468589514494, + 0.01409424189478159, + 0.007673361338675022, + 0.012445399537682533, + 0.010179839096963406, + 0.005106754135340452, + 0.012397384271025658, + 0.006591119337826967, + 0.0031140518840402365, + 0.005824702326208353, + 0.004975089803338051, + 0.00470136571675539, + 0.004887620452791452, + 0.00563921919092536, + 0.004383041523396969, + 0.00504444632679224, + 0.0037504504434764385, + 0.007805327419191599, + 0.00504786754027009, + 0.005974750965833664, + 0.006368087138980627, + 0.006066602189093828, + 0.0060975817032158375, + 0.004403939936310053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0025015748105943203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001592963933945 + ], + "opposite_logprob_tokens": [ + -2.6950159072875977 + ], + "logprob_reward": -0.07001592963933945, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001592963933945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3958702087402344, + 0.22047244012355804, + 0.09798729419708252, + 0.05691651999950409, + 0.058627866208553314, + 0.036714840680360794, + 0.028112970292568207, + 0.030778178945183754, + 0.03291061893105507, + 0.041609037667512894 + ], + "credit_map": [ + 0.0239657461643219, + 0.03533975034952164, + 0.041006434708833694, + 0.0679156556725502, + 0.025877533480525017, + 0.06303642690181732, + 0.04953017458319664, + 0.08919849246740341, + 0.016773903742432594, + 0.02946934662759304, + 0.022494075819849968, + 0.03763905540108681, + 0.021124189719557762, + 0.03323955461382866, + 0.026604432612657547, + 0.03312789276242256, + 0.007469602394849062, + 0.015065547078847885, + 0.009464011527597904, + 0.01256503351032734, + 0.011490510776638985, + 0.014836792834103107, + 0.010911766439676285, + 0.01618403196334839, + 0.004144732374697924, + 0.007393885403871536, + 0.005423790775239468, + 0.0067004431039094925, + 0.005816384684294462, + 0.010839972645044327, + 0.006217188201844692, + 0.010380122810602188, + 0.004257598891854286, + 0.007486263755708933, + 0.003709049429744482, + 0.008280936628580093, + 0.005437142681330442, + 0.011026784777641296, + 0.006056235171854496, + 0.012373855337500572, + 0.0027601346373558044, + 0.0047607263550162315, + 0.0033506755717098713, + 0.006550188641995192, + 0.0034630235750228167, + 0.0048275236040353775, + 0.004489877261221409, + 0.006512689869850874, + 0.002258437918499112, + 0.00403530104085803, + 0.0018784969579428434, + 0.00442931242287159, + 0.0032825691159814596, + 0.004604628775268793, + 0.003435984253883362, + 0.004188238643109798, + 0.002708061831071973, + 0.004870975390076637, + 0.0034525198861956596, + 0.003550999565050006, + 0.004047623835504055, + 0.004987483844161034, + 0.003635331057012081, + 0.0035251809749752283, + 0.0032744058407843113, + 0.004670720547437668, + 0.002424315083771944, + 0.0053976159542799, + 0.003802077379077673, + 0.0051341853104531765, + 0.003906935453414917, + 0.004300363827496767, + 0.003655350301414728, + 0.005291083361953497, + 0.0033483034931123257, + 0.006671424955129623, + 0.0043821781873703, + 0.006959439255297184, + 0.004884346388280392, + 0.006416910793632269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07001592963933945 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092295654118061 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092295654118061, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092295654118061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45462125539779663, + 0.1322687268257141, + 0.06655316054821014, + 0.05387444421648979, + 0.06422454118728638, + 0.051220137625932693, + 0.05450334772467613, + 0.05059894919395447, + 0.028099877759814262, + 0.04403551667928696 + ], + "credit_map": [ + 0.03126884996891022, + 0.043946169316768646, + 0.04363135248422623, + 0.10731799155473709, + 0.0372026301920414, + 0.0556495301425457, + 0.06729764491319656, + 0.06830713152885437, + 0.010731209069490433, + 0.019474884495139122, + 0.015620812773704529, + 0.019019048660993576, + 0.017013750970363617, + 0.019569387659430504, + 0.01828855648636818, + 0.012551099061965942, + 0.006974548567086458, + 0.00851665809750557, + 0.007297474425286055, + 0.008308789692819118, + 0.008213636465370655, + 0.011349315755069256, + 0.008032352663576603, + 0.007860387675464153, + 0.004491589032113552, + 0.00638242345303297, + 0.0046658385545015335, + 0.005473312921822071, + 0.006284557282924652, + 0.012060613371431828, + 0.0040723783895373344, + 0.010443734005093575, + 0.007557873614132404, + 0.00797785259783268, + 0.005724535323679447, + 0.009450690820813179, + 0.005930474027991295, + 0.01190941408276558, + 0.00751015730202198, + 0.008163551799952984, + 0.005907457787543535, + 0.0068766530603170395, + 0.00511613255366683, + 0.005774625577032566, + 0.005977047141641378, + 0.009319735690951347, + 0.0049758125096559525, + 0.00727267749607563, + 0.00455436622723937, + 0.007994553074240685, + 0.004568499978631735, + 0.007890358567237854, + 0.00679380539804697, + 0.010876467451453209, + 0.005437122192233801, + 0.0063881780952215195, + 0.0045925322920084, + 0.006427147425711155, + 0.004347085952758789, + 0.00847777258604765, + 0.0061176796443760395, + 0.0077157458290457726, + 0.007534153759479523, + 0.005386837758123875, + 0.0023326065856963396, + 0.003752975957468152, + 0.002923819934949279, + 0.003134507220238447, + 0.003414271166548133, + 0.004931871313601732, + 0.0030173026025295258, + 0.004592526238411665, + 0.004583726171404123, + 0.005832043942064047, + 0.0038330198731273413, + 0.00605774624273181, + 0.0044588143937289715, + 0.008783869445323944, + 0.005016000475734472, + 0.005470302887260914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004092295654118061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a whimsical Ghibli style with muted colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the focal point an ancient-looking boulder covered in deep emerald moss with tiny droplets?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boulder situated in a garden of meticulously raked white gravel arranged in concentric waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a bright, vibrant rainbow arcing through the sky with saturated color bands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at a static, low-angle close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253088004887104 + ], + "opposite_logprob_tokens": [ + -4.38753080368042 + ], + "logprob_reward": -0.01253088004887104, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253088004887104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48193639516830444, + 0.10279598832130432, + 0.05344365909695625, + 0.042399726808071136, + 0.04570845514535904, + 0.06456387788057327, + 0.06854373216629028, + 0.024704638868570328, + 0.053637195378541946, + 0.06226632744073868 + ], + "credit_map": [ + 0.03533945977687836, + 0.050218161195516586, + 0.07227027416229248, + 0.08108343929052353, + 0.04046359658241272, + 0.07053189724683762, + 0.0738283321261406, + 0.05820121616125107, + 0.008713462390005589, + 0.014327529817819595, + 0.009663891978561878, + 0.014150422066450119, + 0.00973933469504118, + 0.020404411479830742, + 0.012324378825724125, + 0.013472553342580795, + 0.003724884008988738, + 0.009030194953083992, + 0.004630623385310173, + 0.007098689209669828, + 0.005070877261459827, + 0.009486669674515724, + 0.005689145997166634, + 0.00871257483959198, + 0.003925065044313669, + 0.005546422675251961, + 0.0032868150155991316, + 0.006020648870617151, + 0.0035875991452485323, + 0.009349982254207134, + 0.004387951921671629, + 0.006295241415500641, + 0.0036624656058847904, + 0.006931399926543236, + 0.0030746404081583023, + 0.007739474065601826, + 0.0043679336085915565, + 0.008430130779743195, + 0.005860998760908842, + 0.005641408730298281, + 0.005549797788262367, + 0.009540718048810959, + 0.004883522633463144, + 0.008074890822172165, + 0.007030752021819353, + 0.011938506737351418, + 0.007028964813798666, + 0.01051672175526619, + 0.006715982686728239, + 0.009815686382353306, + 0.005925904028117657, + 0.00897984579205513, + 0.008144360035657883, + 0.013385334983468056, + 0.007354116067290306, + 0.008222498930990696, + 0.002177146263420582, + 0.003972702194005251, + 0.00290154991671443, + 0.003946992103010416, + 0.002677388722077012, + 0.0031987344846129417, + 0.002901609055697918, + 0.002928513567894697, + 0.003367892000824213, + 0.009401797316968441, + 0.004244155716150999, + 0.008934993296861649, + 0.006150553468614817, + 0.006190825719386339, + 0.009081167168915272, + 0.0062658097594976425, + 0.0054417988285422325, + 0.009189862757921219, + 0.006195799447596073, + 0.009696164168417454, + 0.006446446292102337, + 0.009634967893362045, + 0.00912516936659813, + 0.006536117754876614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.01253088004887104 + } + ], + "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/0000022/group_5/index_tensor([9209, 9209])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_tensor([9209, 9209])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6abdf3d0d818703195bef0e769a9406c0b79f853 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_tensor([9209, 9209])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a slightly overhead medium shot, then slowly pans forward to focus on a professional chef in a crisp white uniform and tall hat. He concentrates, wielding a gleaming stainless steel ladle, carefully pouring a ladleful of rich, amber-colored broth into a dark stoneware casserole heating on the stovetop. As the hot broth pours in, the meat and lush green vegetables already in the pot gently tumble in the bubbling broth. Puffs of white steam rise from the pot's rim, illuminating it with a warm yellow spotlight overhead, transforming into a visible, lingering cloud of aroma. The camera steadily pans forward, ultimately capturing an extreme close-up of the bubbling, oily broth within the casserole, each bubble visible as it bursts. The chef's profile is clearly outlined. He stares intently at the food in the pot, his eyes full of precise control of the heat. The whole picture, set against warm colors, creates a warm, quiet and craftsman-like cooking atmosphere, as if you can smell the tempting aroma through the screen.", + "seed": 26456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_9209_seed_26456_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": -0.1618542006239295, + "vlm_logprob": -0.1618542006239295, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.1618542006239295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5783093571662903 + ], + "opposite_logprob_tokens": [ + -0.8283093571662903 + ], + "logprob_reward": -0.5783093571662903, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5783093571662903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5785163044929504, + 0.09339481592178345, + 0.025795195251703262, + 0.025844121351838112, + 0.07450666278600693, + 0.058165568858385086, + 0.022185055539011955, + 0.02554200030863285, + 0.05596752464771271, + 0.040082670748233795 + ], + "credit_map": [ + 0.04993791505694389, + 0.0340193547308445, + 0.18398417532444, + 0.03429492935538292, + 0.036408185958862305, + 0.03334783762693405, + 0.17401833832263947, + 0.03250564634799957, + 0.009536733850836754, + 0.011956391856074333, + 0.023115336894989014, + 0.009354791603982449, + 0.006013000849634409, + 0.007422801572829485, + 0.02016228809952736, + 0.005833482835441828, + 0.004173104651272297, + 0.0030841787811368704, + 0.0033068531192839146, + 0.003698557149618864, + 0.0026168718468397856, + 0.002640770049765706, + 0.003564039710909128, + 0.002710824366658926, + 0.0029254970140755177, + 0.0033911787904798985, + 0.00357227330096066, + 0.00515943206846714, + 0.0025017978623509407, + 0.002771778730675578, + 0.0029282234609127045, + 0.002593942452222109, + 0.020154332742094994, + 0.0036935065872967243, + 0.02749248407781124, + 0.004348908085376024, + 0.004258978180587292, + 0.005395706743001938, + 0.0050594075582921505, + 0.00410334300249815, + 0.0070973788388073444, + 0.006134441122412682, + 0.006835370324552059, + 0.008627657778561115, + 0.006025610025972128, + 0.009073933586478233, + 0.006530328653752804, + 0.007840855978429317, + 0.002819129265844822, + 0.0024626669473946095, + 0.0022291301283985376, + 0.0020639952272176743, + 0.003164263442158699, + 0.004517518449574709, + 0.0025055650621652603, + 0.0024227905087172985, + 0.001645063515752554, + 0.0057524386793375015, + 0.0017441710224375129, + 0.0018395169172436, + 0.005833460949361324, + 0.005019323900341988, + 0.0015284533146768808, + 0.0021795749198645353, + 0.0052298279479146, + 0.005254067946225405, + 0.006804234813898802, + 0.0052951970137655735, + 0.00897288043051958, + 0.0145913390442729, + 0.00470310915261507, + 0.005116873886436224, + 0.004580027889460325, + 0.0041322666220366955, + 0.004383545834571123, + 0.0038546998985111713, + 0.008088375441730022, + 0.006750915665179491, + 0.0031558142509311438, + 0.0051370300352573395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.5783093571662903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029838934540748596 + ], + "opposite_logprob_tokens": [ + -3.529839038848877 + ], + "logprob_reward": -0.029838934540748596, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029838934540748596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5571912527084351, + 0.11027015745639801, + 0.040126096457242966, + 0.05865954980254173, + 0.04697272926568985, + 0.05607897788286209, + 0.02088611014187336, + 0.014013118110597134, + 0.026721807196736336, + 0.06908012926578522 + ], + "credit_map": [ + 0.04660719633102417, + 0.033810295164585114, + 0.17630982398986816, + 0.035230472683906555, + 0.034227047115564346, + 0.03146100416779518, + 0.1705136001110077, + 0.029031839221715927, + 0.008963785134255886, + 0.010227312333881855, + 0.03150555491447449, + 0.010941347107291222, + 0.006161956582218409, + 0.008993644267320633, + 0.027141954749822617, + 0.006334608420729637, + 0.006307157687842846, + 0.004018494393676519, + 0.007868179120123386, + 0.004929150454699993, + 0.0037285652942955494, + 0.0032204727176576853, + 0.005788343958556652, + 0.0042657312005758286, + 0.008090333081781864, + 0.01217573881149292, + 0.009658639319241047, + 0.00702022947371006, + 0.00642473204061389, + 0.006244021002203226, + 0.004195979796350002, + 0.004849874880164862, + 0.012314083985984325, + 0.0023934782948344946, + 0.017249923199415207, + 0.0028860203456133604, + 0.00309964781627059, + 0.0030835438519716263, + 0.003332967171445489, + 0.0026130664627999067, + 0.006724648643285036, + 0.006932466756552458, + 0.0071983071975409985, + 0.004059962462633848, + 0.006755426991730928, + 0.012624281458556652, + 0.005633464083075523, + 0.006150422617793083, + 0.0023094448260962963, + 0.003183497581630945, + 0.0015401147538796067, + 0.0019764662720263004, + 0.0030268921982496977, + 0.004421262536197901, + 0.0021516061387956142, + 0.002276825252920389, + 0.0018823002465069294, + 0.0032600664999336004, + 0.0012367714662104845, + 0.0011876614298671484, + 0.0024635144509375095, + 0.0022151072043925524, + 0.0007566213607788086, + 0.0010110753355547786, + 0.002600496867671609, + 0.002446299884468317, + 0.0035791194532066584, + 0.0030353404581546783, + 0.004349332302808762, + 0.005545562598854303, + 0.002713957102969289, + 0.002451699459925294, + 0.008941514417529106, + 0.007000255398452282, + 0.008531135506927967, + 0.004545085597783327, + 0.011977031826972961, + 0.015128841623663902, + 0.005299821030348539, + 0.007656442932784557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.029838934540748596 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993953824043274 + ], + "opposite_logprob_tokens": [ + -1.8049395084381104 + ], + "logprob_reward": -0.17993953824043274, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993953824043274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1309187114238739, + 0.11609399318695068, + 0.023203449323773384, + 0.06978730112314224, + 0.07553569227457047, + 0.045289296656847, + 0.015831423923373222, + 0.03519425913691521, + 0.04825051501393318, + 0.4398953914642334 + ], + "credit_map": [ + 0.01363054197281599, + 0.01217157393693924, + 0.03848884627223015, + 0.009826491586863995, + 0.008324626833200455, + 0.010285363532602787, + 0.028696594759821892, + 0.009494678117334843, + 0.009542692452669144, + 0.011289751157164574, + 0.033833567053079605, + 0.008849190548062325, + 0.005862973164767027, + 0.007387776859104633, + 0.03195196017622948, + 0.007376089226454496, + 0.003952319268137217, + 0.0020380564965307713, + 0.006027343217283487, + 0.0026415789034217596, + 0.0015054374234750867, + 0.0018763960106298327, + 0.0027347097638994455, + 0.0024276089388877153, + 0.005629359278827906, + 0.009855842217803001, + 0.01535905059427023, + 0.012206180952489376, + 0.007702778559178114, + 0.006845787167549133, + 0.0068068429827690125, + 0.005381457507610321, + 0.008669141680002213, + 0.008289619348943233, + 0.0183834508061409, + 0.010594265535473824, + 0.010960308834910393, + 0.007899729534983635, + 0.004823167342692614, + 0.0059160105884075165, + 0.00493593281134963, + 0.00808639544993639, + 0.0047215684317052364, + 0.004958091303706169, + 0.005319319665431976, + 0.007270903792232275, + 0.00610341876745224, + 0.0038936659693717957, + 0.001969950972124934, + 0.00122824648860842, + 0.0023732814006507397, + 0.0017289454117417336, + 0.0016654767096042633, + 0.0029115178622305393, + 0.0019351241644471884, + 0.002018881030380726, + 0.0032347480300813913, + 0.005677126348018646, + 0.0028244126588106155, + 0.0034912354312837124, + 0.005174749996513128, + 0.008636397309601307, + 0.002608021255582571, + 0.0035475699696689844, + 0.00530231324955821, + 0.004339830484241247, + 0.00569050433114171, + 0.006688167341053486, + 0.007753725163638592, + 0.01037134975194931, + 0.004723664373159409, + 0.003380960552021861, + 0.05864556133747101, + 0.03688894584774971, + 0.05501265823841095, + 0.036876823753118515, + 0.056362085044384, + 0.10940520465373993, + 0.03849904239177704, + 0.04820505157113075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.17993953824043274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254300493746996 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.01254300493746996, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254300493746996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31630739569664, + 0.11520091444253922, + 0.032113976776599884, + 0.0842934176325798, + 0.07978921383619308, + 0.15944938361644745, + 0.057238951325416565, + 0.04591112583875656, + 0.03579014539718628, + 0.07390546053647995 + ], + "credit_map": [ + 0.025514597073197365, + 0.022094564512372017, + 0.0951247438788414, + 0.01894724741578102, + 0.020600203424692154, + 0.023043710738420486, + 0.09343215823173523, + 0.01755017228424549, + 0.009702119044959545, + 0.012253182008862495, + 0.03314897045493126, + 0.009748926386237144, + 0.0055783092975616455, + 0.005856897681951523, + 0.031916625797748566, + 0.006995868869125843, + 0.0045790523290634155, + 0.003520890139043331, + 0.0034764546435326338, + 0.004869346506893635, + 0.0029907438438385725, + 0.004732821602374315, + 0.0031235304195433855, + 0.004821138922125101, + 0.011452695354819298, + 0.014754684641957283, + 0.012718096375465393, + 0.012235320173203945, + 0.009567447938024998, + 0.012964466586709023, + 0.005365464370697737, + 0.005235246382653713, + 0.009072049520909786, + 0.007325448095798492, + 0.020824823528528214, + 0.008649645373225212, + 0.01296621561050415, + 0.008505851961672306, + 0.005480815190821886, + 0.006964370142668486, + 0.018280522897839546, + 0.023220712319016457, + 0.01608402468264103, + 0.014138390310108662, + 0.017837954685091972, + 0.032113585621118546, + 0.020494554191827774, + 0.017279645428061485, + 0.00624491274356842, + 0.0096082529053092, + 0.005059857852756977, + 0.006402528844773769, + 0.009583931416273117, + 0.008513635024428368, + 0.004804456606507301, + 0.007021377328783274, + 0.002604498527944088, + 0.010681935586035252, + 0.0036552096717059612, + 0.0036009284667670727, + 0.012284979224205017, + 0.007187332957983017, + 0.0025468156673014164, + 0.0033494289964437485, + 0.00361813697963953, + 0.0031145066022872925, + 0.004078186582773924, + 0.003272037487477064, + 0.005365618970245123, + 0.010392124764621258, + 0.002992413006722927, + 0.002957119606435299, + 0.010553468018770218, + 0.006237437948584557, + 0.010576393455266953, + 0.005586347077041864, + 0.009951791726052761, + 0.01741791144013405, + 0.0057143112644553185, + 0.007867802865803242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01254300493746996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640168234705925 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640168234705925, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640168234705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3051619231700897, + 0.039582423865795135, + 0.014412146992981434, + 0.04282372072339058, + 0.0580417700111866, + 0.35911232233047485, + 0.039073411375284195, + 0.016056636348366737, + 0.0768134668469429, + 0.048922084271907806 + ], + "credit_map": [ + 0.028236065059900284, + 0.018445605412125587, + 0.10070598870515823, + 0.01756296493113041, + 0.017373766750097275, + 0.018005920574069023, + 0.09004237502813339, + 0.014789262786507607, + 0.0032831644639372826, + 0.0033071881625801325, + 0.010368394665420055, + 0.004398443270474672, + 0.0027959474828094244, + 0.0029151795897632837, + 0.010089936666190624, + 0.002424175851047039, + 0.0019088907865807414, + 0.001766518922522664, + 0.0019678359385579824, + 0.0019257680978626013, + 0.0014682222390547395, + 0.0019472124986350536, + 0.001747164991684258, + 0.0016805335180833936, + 0.00929982028901577, + 0.004310520365834236, + 0.010756382718682289, + 0.0032773781567811966, + 0.003952506929636002, + 0.005657234229147434, + 0.002041315659880638, + 0.003528566099703312, + 0.011760479770600796, + 0.005608904641121626, + 0.01811695285141468, + 0.004792032763361931, + 0.005284782964736223, + 0.0049601346254348755, + 0.0038626196328550577, + 0.003655871609225869, + 0.04052778705954552, + 0.04771329462528229, + 0.04183207079768181, + 0.030874032527208328, + 0.03816326707601547, + 0.07732594013214111, + 0.0475873127579689, + 0.03508865833282471, + 0.004093787167221308, + 0.003940696828067303, + 0.0034916785079985857, + 0.003836099524050951, + 0.0050995247438549995, + 0.009922070428729057, + 0.00369097082875669, + 0.004998589400202036, + 0.002502439310774207, + 0.003352657426148653, + 0.001187576213851571, + 0.001064378535374999, + 0.002455857815220952, + 0.003412747522816062, + 0.0009932207176461816, + 0.0010877602035179734, + 0.007509481627494097, + 0.006950289476662874, + 0.0076375361531972885, + 0.004208680707961321, + 0.014257969334721565, + 0.027352137491106987, + 0.0039304280653595924, + 0.004966953303664923, + 0.006764566991478205, + 0.006593723315745592, + 0.006302305031567812, + 0.0035985943395644426, + 0.007919074036180973, + 0.009165506809949875, + 0.0032242315355688334, + 0.0053540864028036594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.008640168234705925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly overhead medium shot, then slowly pans forward to focus on a professional chef in a crisp white uniform and tall hat. He concentrates, wielding a gleaming stainless steel ladle, carefully pouring a ladleful of rich, amber-colored broth into a dark stoneware casserole heating on the stovetop. As the hot broth pours in, the meat and lush green vegetables already in the pot gently tumble in the bubbling broth. Puffs of white steam rise from the pot's rim, illuminating it with a warm yellow spotlight overhead, transforming into a visible, lingering cloud of aroma. The camera steadily pans forward, ultimately capturing an extreme close-up of the bubbling, oily broth within the casserole, each bubble visible as it bursts. The chef's profile is clearly outlined. He stares intently at the food in the pot, his eyes full of precise control of the heat. The whole picture, set against warm colors, creates a warm, quiet and craftsman-like cooking atmosphere, as if you can smell the tempting aroma through the screen.", + "seed": 26457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_9209_seed_26457_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": -0.05696762762963772, + "vlm_logprob": -0.05696762762963772, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05696762762963772, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18339082598686218 + ], + "opposite_logprob_tokens": [ + -1.8083908557891846 + ], + "logprob_reward": -0.18339082598686218, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18339082598686218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24917200207710266, + 0.37033045291900635, + 0.05064602568745613, + 0.1016656756401062, + 0.055759236216545105, + 0.04193568229675293, + 0.020843002945184708, + 0.029999440535902977, + 0.058287106454372406, + 0.0213614609092474 + ], + "credit_map": [ + 0.019910713657736778, + 0.046869050711393356, + 0.02049679309129715, + 0.04614149034023285, + 0.035651788115501404, + 0.029851915314793587, + 0.025888465344905853, + 0.024361757561564445, + 0.022118020802736282, + 0.04840714856982231, + 0.027470655739307404, + 0.0360570028424263, + 0.0708828940987587, + 0.07310857623815536, + 0.05707843601703644, + 0.03520767763257027, + 0.0025975224561989307, + 0.012833927758038044, + 0.0035022308584302664, + 0.005814247764647007, + 0.009339447133243084, + 0.007451143581420183, + 0.005308500025421381, + 0.0037990000564604998, + 0.011091767810285091, + 0.013907759450376034, + 0.009885088540613651, + 0.009496721439063549, + 0.022511322051286697, + 0.012909715063869953, + 0.014502160251140594, + 0.007361138239502907, + 0.006344253662973642, + 0.005962788593024015, + 0.007334759924560785, + 0.0050200545229017735, + 0.009693672880530357, + 0.010520484298467636, + 0.005561902187764645, + 0.005321316421031952, + 0.0030412746127694845, + 0.01419877540320158, + 0.0020473545882850885, + 0.002002217574045062, + 0.013559969142079353, + 0.003156947670504451, + 0.0024923591408878565, + 0.00143678009044379, + 0.001677843974903226, + 0.00255611096508801, + 0.002476769033819437, + 0.002042578998953104, + 0.005171706434339285, + 0.002307811053469777, + 0.002444393001496792, + 0.002165785990655422, + 0.0025487514212727547, + 0.004329422488808632, + 0.002621714724227786, + 0.003987386357039213, + 0.005877251271158457, + 0.0034529392141848803, + 0.004095832817256451, + 0.0030861401464790106, + 0.004737398587167263, + 0.010100265964865685, + 0.004128940403461456, + 0.006728298030793667, + 0.011685366742312908, + 0.009420695714652538, + 0.006422000005841255, + 0.005064135417342186, + 0.0017900113016366959, + 0.00408166041597724, + 0.0018406484741717577, + 0.001950635458342731, + 0.004211297258734703, + 0.003738376311957836, + 0.0016129440627992153, + 0.0021358865778893232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.18339082598686218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01610919088125229 + ], + "opposite_logprob_tokens": [ + -4.141108989715576 + ], + "logprob_reward": -0.01610919088125229, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01610919088125229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10440302640199661, + 0.2812628149986267, + 0.04161646589636803, + 0.09336435049772263, + 0.14091236889362335, + 0.06374096870422363, + 0.05511564761400223, + 0.10721393674612045, + 0.052785973995923996, + 0.05958446487784386 + ], + "credit_map": [ + 0.00953303650021553, + 0.018971091136336327, + 0.00979614071547985, + 0.014276394620537758, + 0.013500215485692024, + 0.013701111078262329, + 0.014179418794810772, + 0.01044560968875885, + 0.015361473895609379, + 0.08140572160482407, + 0.025581730529665947, + 0.04800078645348549, + 0.029997527599334717, + 0.031383197754621506, + 0.028191233053803444, + 0.021341130137443542, + 0.003050125204026699, + 0.004761017393320799, + 0.0036341636441648006, + 0.0051023634150624275, + 0.005668309982866049, + 0.009536290541291237, + 0.005778825376182795, + 0.0040853675454854965, + 0.0133203761652112, + 0.011063051410019398, + 0.016891274601221085, + 0.010574422776699066, + 0.013571458868682384, + 0.01092794630676508, + 0.008157212287187576, + 0.008858599700033665, + 0.01882844790816307, + 0.015115529298782349, + 0.023497959598898888, + 0.014344452880322933, + 0.020401019603013992, + 0.019572220742702484, + 0.018124403432011604, + 0.011028311215341091, + 0.005864775273948908, + 0.012160240672528744, + 0.005684015806764364, + 0.004658963996917009, + 0.012585987336933613, + 0.007551037706434727, + 0.008582014590501785, + 0.006653925403952599, + 0.004225519485771656, + 0.008957119658589363, + 0.004112676717340946, + 0.0052194371819496155, + 0.014136632904410362, + 0.006559661123901606, + 0.007711903192102909, + 0.004192691762000322, + 0.006958856713026762, + 0.01807478442788124, + 0.006664324086159468, + 0.009287961758673191, + 0.028433023020625114, + 0.012005255557596684, + 0.018140653148293495, + 0.0076490603387355804, + 0.004011711571365595, + 0.008149612694978714, + 0.003811750328168273, + 0.006348093040287495, + 0.010476337745785713, + 0.008141822181642056, + 0.007061297073960304, + 0.004785343538969755, + 0.005534665659070015, + 0.015075508505105972, + 0.003941403701901436, + 0.005219779908657074, + 0.009318125434219837, + 0.008874291554093361, + 0.004439027979969978, + 0.007181656081229448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.01610919088125229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868089407682419 + ], + "opposite_logprob_tokens": [ + -3.0486807823181152 + ], + "logprob_reward": -0.04868089407682419, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868089407682419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09204727411270142, + 0.08280536532402039, + 0.06007237359881401, + 0.043404217809438705, + 0.09086158871650696, + 0.3052361309528351, + 0.04027349129319191, + 0.08043680340051651, + 0.043523725122213364, + 0.16133905947208405 + ], + "credit_map": [ + 0.006219225935637951, + 0.020736997947096825, + 0.007588411681354046, + 0.019188905134797096, + 0.011510463431477547, + 0.007004386279731989, + 0.012214869260787964, + 0.0075840107165277, + 0.006482974160462618, + 0.01606973260641098, + 0.012254933826625347, + 0.007900461554527283, + 0.013953411020338535, + 0.010460724122822285, + 0.009448351338505745, + 0.006234781816601753, + 0.002857899060472846, + 0.015797188505530357, + 0.0023175603710114956, + 0.004905323032289743, + 0.01246182806789875, + 0.01201802771538496, + 0.004722591023892164, + 0.004991954658180475, + 0.005721052177250385, + 0.0051673296838998795, + 0.0063699171878397465, + 0.006371417548507452, + 0.005425650160759687, + 0.0038244035094976425, + 0.006393079645931721, + 0.0041313678957521915, + 0.010233089327812195, + 0.010378372855484486, + 0.011001009494066238, + 0.007996306754648685, + 0.016791436821222305, + 0.01776954159140587, + 0.00976097397506237, + 0.006930863950401545, + 0.022715019062161446, + 0.127498060464859, + 0.00420540664345026, + 0.006053542718291283, + 0.11675074696540833, + 0.01650017313659191, + 0.005305909551680088, + 0.006207277532666922, + 0.0036699159536510706, + 0.0071458094753324986, + 0.003709471318870783, + 0.0033578521106392145, + 0.008665218949317932, + 0.004385032225400209, + 0.005771584343165159, + 0.0035686090122908354, + 0.004457858856767416, + 0.010040623135864735, + 0.010720278136432171, + 0.013026639819145203, + 0.012944954447448254, + 0.008008312433958054, + 0.011489487253129482, + 0.009748652577400208, + 0.0036220720503479242, + 0.007144520524889231, + 0.003643900156021118, + 0.005224805790930986, + 0.009130993857979774, + 0.005638260394334793, + 0.004844220820814371, + 0.004274952225387096, + 0.013390458188951015, + 0.01922484114766121, + 0.01863955706357956, + 0.016049019992351532, + 0.03719571605324745, + 0.027727322652935982, + 0.01431040558964014, + 0.014801736921072006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.04868089407682419 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020586160942912102 + ], + "opposite_logprob_tokens": [ + -3.8955862522125244 + ], + "logprob_reward": -0.020586160942912102, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020586160942912102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07856986671686172, + 0.26617735624313354, + 0.07676608860492706, + 0.058277200907468796, + 0.0263198409229517, + 0.19046825170516968, + 0.02861352078616619, + 0.050816718488931656, + 0.14044789969921112, + 0.08354316651821136 + ], + "credit_map": [ + 0.006617962382733822, + 0.016621625050902367, + 0.00640304759144783, + 0.015203260816633701, + 0.012191659770905972, + 0.007050214800983667, + 0.009908700361847878, + 0.004573397804051638, + 0.01362240593880415, + 0.07406927645206451, + 0.018044864758849144, + 0.049176428467035294, + 0.03403395414352417, + 0.03281702101230621, + 0.024879658594727516, + 0.01953376643359661, + 0.004190697334706783, + 0.020282775163650513, + 0.005206144880503416, + 0.006872371304780245, + 0.015772590413689613, + 0.013300945982336998, + 0.005236528813838959, + 0.005904042162001133, + 0.005425744690001011, + 0.0074187638238072395, + 0.008022824302315712, + 0.005064711440354586, + 0.010790425352752209, + 0.008537739515304565, + 0.008298691362142563, + 0.004718303214758635, + 0.0032106360886245966, + 0.0028468340169638395, + 0.0038062676321715117, + 0.0032068213913589716, + 0.004073161166161299, + 0.0038175228983163834, + 0.003325817873701453, + 0.002032781718298793, + 0.01348444540053606, + 0.07831284403800964, + 0.0030999076552689075, + 0.0033118620049208403, + 0.07265804708003998, + 0.010678576305508614, + 0.004652612376958132, + 0.004269967321306467, + 0.002448635408654809, + 0.0045616161078214645, + 0.004074523691087961, + 0.004213184583932161, + 0.003972293343394995, + 0.0029147695749998093, + 0.0027165438514202833, + 0.003711956087499857, + 0.003934225998818874, + 0.007489471696317196, + 0.004118474666029215, + 0.004735137335956097, + 0.010966531932353973, + 0.004833819344639778, + 0.009532381780445576, + 0.005206676200032234, + 0.011671829968690872, + 0.018162105232477188, + 0.014935922808945179, + 0.027105700224637985, + 0.016774272546172142, + 0.018198788166046143, + 0.01444785576313734, + 0.01915142312645912, + 0.004468273371458054, + 0.014744073152542114, + 0.003095039399340749, + 0.004917833488434553, + 0.018377281725406647, + 0.016868598759174347, + 0.0067731840535998344, + 0.014298894442617893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 8 + ], + "selected_reward": -0.020586160942912102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607106626033783 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.01607106626033783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607106626033783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12188994139432907, + 0.35237717628479004, + 0.07208725064992905, + 0.03197173401713371, + 0.04281173273921013, + 0.22145740687847137, + 0.024044467136263847, + 0.028197573497891426, + 0.034226834774017334, + 0.07093584537506104 + ], + "credit_map": [ + 0.01196337305009365, + 0.026068154722452164, + 0.011557239107787609, + 0.023163774982094765, + 0.014877540059387684, + 0.008248018100857735, + 0.01878662407398224, + 0.007225220091640949, + 0.03320774435997009, + 0.10954291373491287, + 0.06270014494657516, + 0.0504106841981411, + 0.027819527313113213, + 0.01855522207915783, + 0.029947876930236816, + 0.02019309625029564, + 0.005189971067011356, + 0.009615025483071804, + 0.007229646202176809, + 0.009194090962409973, + 0.009850154630839825, + 0.013770469464361668, + 0.008746220730245113, + 0.008491677232086658, + 0.003702927613630891, + 0.003601349424570799, + 0.004705928266048431, + 0.0037450201343744993, + 0.00444745272397995, + 0.004343945067375898, + 0.00438291160389781, + 0.0030422010459005833, + 0.005189655814319849, + 0.004803122486919165, + 0.006520354654639959, + 0.004404235631227493, + 0.006788591388612986, + 0.00731992581859231, + 0.0044055478647351265, + 0.0033803016413003206, + 0.015602326951920986, + 0.09054568409919739, + 0.002575005404651165, + 0.0033594719134271145, + 0.08688133209943771, + 0.013244068250060081, + 0.003979496657848358, + 0.005270044784992933, + 0.0023396522738039494, + 0.0031404399778693914, + 0.002627922687679529, + 0.004067671950906515, + 0.004763116594403982, + 0.0025712940841913223, + 0.0026447977870702744, + 0.001889572711661458, + 0.001971850637346506, + 0.003760376712307334, + 0.0035254827234894037, + 0.004082933533936739, + 0.004100131802260876, + 0.0036894497461616993, + 0.003902295371517539, + 0.0031650548335164785, + 0.002893401775509119, + 0.005258872173726559, + 0.002627601381391287, + 0.0068648261949419975, + 0.004355981480330229, + 0.004426126834005117, + 0.0034611541777849197, + 0.004338875412940979, + 0.005332618020474911, + 0.021789545193314552, + 0.004508898593485355, + 0.0047386521473526955, + 0.012705438770353794, + 0.01489904336631298, + 0.003578499425202608, + 0.003383151488378644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.01607106626033783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly overhead medium shot, then slowly pans forward to focus on a professional chef in a crisp white uniform and tall hat. He concentrates, wielding a gleaming stainless steel ladle, carefully pouring a ladleful of rich, amber-colored broth into a dark stoneware casserole heating on the stovetop. As the hot broth pours in, the meat and lush green vegetables already in the pot gently tumble in the bubbling broth. Puffs of white steam rise from the pot's rim, illuminating it with a warm yellow spotlight overhead, transforming into a visible, lingering cloud of aroma. The camera steadily pans forward, ultimately capturing an extreme close-up of the bubbling, oily broth within the casserole, each bubble visible as it bursts. The chef's profile is clearly outlined. He stares intently at the food in the pot, his eyes full of precise control of the heat. The whole picture, set against warm colors, creates a warm, quiet and craftsman-like cooking atmosphere, as if you can smell the tempting aroma through the screen.", + "seed": 26458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_9209_seed_26458_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": -0.0839336789213121, + "vlm_logprob": -0.0839336789213121, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.0839336789213121, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25352156162261963 + ], + "opposite_logprob_tokens": [ + -1.5035215616226196 + ], + "logprob_reward": -0.25352156162261963, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25352156162261963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10676388442516327, + 0.19521473348140717, + 0.052894048392772675, + 0.2533605694770813, + 0.08131753653287888, + 0.09249527752399445, + 0.037915296852588654, + 0.06341399997472763, + 0.031196394935250282, + 0.08542825281620026 + ], + "credit_map": [ + 0.01124208141118288, + 0.015749957412481308, + 0.012759018689393997, + 0.01372721791267395, + 0.013918724842369556, + 0.010477744042873383, + 0.011623982340097427, + 0.017265165224671364, + 0.040321167558431625, + 0.06900687515735626, + 0.015964318066835403, + 0.020360110327601433, + 0.0102746756747365, + 0.015117181465029716, + 0.008176242932677269, + 0.015994157642126083, + 0.0048424480482935905, + 0.005688660778105259, + 0.004228429868817329, + 0.005446611903607845, + 0.004211245570331812, + 0.009020456112921238, + 0.006564329378306866, + 0.012891861610114574, + 0.025648634880781174, + 0.019974565133452415, + 0.02353704534471035, + 0.04030464217066765, + 0.022607972845435143, + 0.05789653956890106, + 0.025062372907996178, + 0.03832880035042763, + 0.009338542819023132, + 0.011130766943097115, + 0.009052306413650513, + 0.008550196886062622, + 0.012264726683497429, + 0.014863918535411358, + 0.008885826915502548, + 0.0072312443517148495, + 0.007114612031728029, + 0.007912267930805683, + 0.006199222523719072, + 0.013158776797354221, + 0.013412210159003735, + 0.018865685909986496, + 0.01663152128458023, + 0.009200980886816978, + 0.007854635827243328, + 0.006360467057675123, + 0.00304001709446311, + 0.0034804539754986763, + 0.004097576718777418, + 0.007048324681818485, + 0.002756856381893158, + 0.0032769665122032166, + 0.006283261813223362, + 0.006856648717075586, + 0.008129903115332127, + 0.00534472381696105, + 0.006728348322212696, + 0.007271407172083855, + 0.005802310071885586, + 0.016997400671243668, + 0.0025652344338595867, + 0.004187134560197592, + 0.003972773440182209, + 0.0035183955915272236, + 0.00508125452324748, + 0.005473140627145767, + 0.0026395991444587708, + 0.0037588647101074457, + 0.00639257300645113, + 0.02816786803305149, + 0.007950049825012684, + 0.003939022775739431, + 0.018448231741786003, + 0.014205687679350376, + 0.002484620315954089, + 0.003840206190943718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.25352156162261963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208357587456703 + ], + "opposite_logprob_tokens": [ + -2.8120834827423096 + ], + "logprob_reward": -0.06208357587456703, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208357587456703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06458223611116409, + 0.20241299271583557, + 0.058833882212638855, + 0.09484471380710602, + 0.0820009708404541, + 0.16097070276737213, + 0.022244080901145935, + 0.04190799966454506, + 0.02455744706094265, + 0.24764497578144073 + ], + "credit_map": [ + 0.00820712000131607, + 0.011492392979562283, + 0.009304631501436234, + 0.009251080453395844, + 0.006237925495952368, + 0.005333573557436466, + 0.007270922418683767, + 0.007484595291316509, + 0.042306412011384964, + 0.07609126716852188, + 0.01569022238254547, + 0.022143883630633354, + 0.010531937703490257, + 0.01319080125540495, + 0.007010919973254204, + 0.015447555109858513, + 0.006289650686085224, + 0.005215163808315992, + 0.004152944311499596, + 0.006604856811463833, + 0.006789441220462322, + 0.01066474337130785, + 0.007176804821938276, + 0.011940272524952888, + 0.010865605436265469, + 0.010875525884330273, + 0.00950684119015932, + 0.014761456288397312, + 0.008743563666939735, + 0.0180580522865057, + 0.010378447361290455, + 0.011655221693217754, + 0.008629958145320415, + 0.009142099879682064, + 0.015074390918016434, + 0.010944767855107784, + 0.00912622269243002, + 0.012022718787193298, + 0.009581223130226135, + 0.007479591760784388, + 0.013174457475543022, + 0.014166726730763912, + 0.01300484873354435, + 0.02423076517879963, + 0.020196720957756042, + 0.03449905663728714, + 0.02386671118438244, + 0.017831414937973022, + 0.002020699204877019, + 0.0043079424649477005, + 0.0013319285353645682, + 0.0016656782245263457, + 0.005197336431592703, + 0.004138065502047539, + 0.002024318091571331, + 0.0015581119805574417, + 0.004508429206907749, + 0.004066702909767628, + 0.004799363203346729, + 0.003249825444072485, + 0.0037646193522959948, + 0.005230782553553581, + 0.0039591616950929165, + 0.012329116463661194, + 0.0018700201762840152, + 0.00314119434915483, + 0.0026204970199614763, + 0.002882333705201745, + 0.004403980914503336, + 0.0040706247091293335, + 0.0028851686511188745, + 0.002683627186343074, + 0.009853916242718697, + 0.08834636211395264, + 0.004708935506641865, + 0.004504998214542866, + 0.10682576894760132, + 0.02468624711036682, + 0.004855146165937185, + 0.0038635889068245888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.06208357587456703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486704483628273 + ], + "opposite_logprob_tokens": [ + -3.048670530319214 + ], + "logprob_reward": -0.0486704483628273, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486704483628273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03329583629965782, + 0.05273168161511421, + 0.03405971825122833, + 0.06414871662855148, + 0.0471978522837162, + 0.03988835960626602, + 0.03147305175662041, + 0.05188900977373123, + 0.10933747887611389, + 0.5359783172607422 + ], + "credit_map": [ + 0.00729956803843379, + 0.010683153755962849, + 0.002726756501942873, + 0.0027998615987598896, + 0.0024810403119772673, + 0.0014638687716796994, + 0.001753258053213358, + 0.004088331013917923, + 0.00521697336807847, + 0.016832077875733376, + 0.004457613918930292, + 0.009093793109059334, + 0.005470911040902138, + 0.0035567244049161673, + 0.0024359587114304304, + 0.005667631980031729, + 0.0031491799745708704, + 0.0058802044950425625, + 0.0032689543440937996, + 0.0037589839193969965, + 0.002971823327243328, + 0.004029448144137859, + 0.004871688783168793, + 0.0061294338665902615, + 0.011771268211305141, + 0.009890533052384853, + 0.008121531456708908, + 0.010295182466506958, + 0.0033737581688910723, + 0.0069307745434343815, + 0.008082006126642227, + 0.00568365678191185, + 0.0056368689984083176, + 0.0059225126169621944, + 0.0050609842874109745, + 0.0046969447284936905, + 0.008645337074995041, + 0.005730850622057915, + 0.0067594763822853565, + 0.004744876176118851, + 0.003932580817490816, + 0.003692084224894643, + 0.003108485834673047, + 0.006676309276372194, + 0.005348410923033953, + 0.007577028591185808, + 0.005561238620430231, + 0.003992222249507904, + 0.0029360621701925993, + 0.005290478467941284, + 0.00196193833835423, + 0.004193812143057585, + 0.006425193976610899, + 0.0038464893586933613, + 0.003509258385747671, + 0.003309819148853421, + 0.007019260432571173, + 0.0065472666174173355, + 0.00872937124222517, + 0.005127669312059879, + 0.006193263456225395, + 0.004425838124006987, + 0.0036756261251866817, + 0.010170714929699898, + 0.008427368476986885, + 0.01376004982739687, + 0.008818739093840122, + 0.011229784227907658, + 0.023960625752806664, + 0.019931096583604813, + 0.009691739454865456, + 0.013518076390028, + 0.05051921308040619, + 0.21201713383197784, + 0.03516463190317154, + 0.01633678749203682, + 0.050556693226099014, + 0.14431758224964142, + 0.012135794386267662, + 0.014930485747754574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0486704483628273 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656027764081955 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.048656027764081955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656027764081955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08967873454093933, + 0.07897619158029556, + 0.02483869530260563, + 0.06214368715882301, + 0.0310305655002594, + 0.029519367963075638, + 0.02295529656112194, + 0.4019295871257782, + 0.08740761876106262, + 0.17152026295661926 + ], + "credit_map": [ + 0.01208022516220808, + 0.01853802241384983, + 0.011638001538813114, + 0.006726499646902084, + 0.007262682542204857, + 0.006202385760843754, + 0.006604164373129606, + 0.020626753568649292, + 0.005964495241641998, + 0.013880379498004913, + 0.01116052083671093, + 0.01612715981900692, + 0.008280357345938683, + 0.009321916848421097, + 0.004907479975372553, + 0.009333883412182331, + 0.0032727133948355913, + 0.003199546132236719, + 0.0025925305671989918, + 0.002500093076378107, + 0.001845689956098795, + 0.0035125070717185736, + 0.0034662955440580845, + 0.004449315834790468, + 0.006191665306687355, + 0.005713983904570341, + 0.007073858752846718, + 0.009817983955144882, + 0.008229336701333523, + 0.010783265344798565, + 0.006990919820964336, + 0.007342668250203133, + 0.003330445382744074, + 0.004808762576431036, + 0.002200688933953643, + 0.0029745555948466063, + 0.004011994693428278, + 0.007338039577007294, + 0.0036741208750754595, + 0.0026919571682810783, + 0.0027048250194638968, + 0.002450709929689765, + 0.002093428047373891, + 0.007648594677448273, + 0.00442949403077364, + 0.0031606804113835096, + 0.0040564751252532005, + 0.002975156996399164, + 0.0028975268360227346, + 0.004161609802395105, + 0.0022324882447719574, + 0.0026107472367584705, + 0.0033882679417729378, + 0.0029210802167654037, + 0.002561177359893918, + 0.002182398457080126, + 0.04398719593882561, + 0.031585145741701126, + 0.05008864030241966, + 0.027113022282719612, + 0.04115047678351402, + 0.04695100709795952, + 0.037816114723682404, + 0.1232379674911499, + 0.00852921698242426, + 0.012937444262206554, + 0.012923362664878368, + 0.010055117309093475, + 0.010805194266140461, + 0.017011886462569237, + 0.0076652271673083305, + 0.007480166386812925, + 0.011119420640170574, + 0.07530888170003891, + 0.004272784572094679, + 0.005236336495727301, + 0.03171638399362564, + 0.03555544465780258, + 0.003284434089437127, + 0.005026565864682198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.048656027764081955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736780982464552 + ], + "opposite_logprob_tokens": [ + -5.006736755371094 + ], + "logprob_reward": -0.006736780982464552, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736780982464552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14425405859947205, + 0.12202082574367523, + 0.08888012915849686, + 0.11733158677816391, + 0.055140022188425064, + 0.037436410784721375, + 0.03462901711463928, + 0.04122490808367729, + 0.0359644778072834, + 0.32311856746673584 + ], + "credit_map": [ + 0.024176914244890213, + 0.030871957540512085, + 0.020117657259106636, + 0.018940620124340057, + 0.012246866710484028, + 0.008519509807229042, + 0.010912228375673294, + 0.018468307331204414, + 0.02031688019633293, + 0.04541930556297302, + 0.008940402418375015, + 0.018250249326229095, + 0.009323439560830593, + 0.005968907848000526, + 0.00467679975554347, + 0.009124846197664738, + 0.009511981159448624, + 0.016134245321154594, + 0.009565409272909164, + 0.010468004271388054, + 0.006730316672474146, + 0.012173937633633614, + 0.011497106403112411, + 0.012799130752682686, + 0.0182404275983572, + 0.020387405529618263, + 0.013883847743272781, + 0.01441812701523304, + 0.009047728031873703, + 0.013316972181200981, + 0.016698727384209633, + 0.011338351294398308, + 0.0057242196053266525, + 0.007941207848489285, + 0.004074530676007271, + 0.00524971354752779, + 0.00982057023793459, + 0.011468415148556232, + 0.006282536778599024, + 0.004578824620693922, + 0.003825219813734293, + 0.0041543287225067616, + 0.002693293383345008, + 0.0069231330417096615, + 0.0045586759224534035, + 0.007882733829319477, + 0.004020879976451397, + 0.0033781470265239477, + 0.004301831126213074, + 0.003838599193841219, + 0.003143654903396964, + 0.004806061740964651, + 0.005441973451524973, + 0.004467233549803495, + 0.00404710415750742, + 0.00458255922421813, + 0.00514944177120924, + 0.006626698654145002, + 0.004904320929199457, + 0.004382087849080563, + 0.006275412626564503, + 0.003634453983977437, + 0.0030194073915481567, + 0.007233084179461002, + 0.002929016947746277, + 0.005276645999401808, + 0.0032161064445972443, + 0.00453248480334878, + 0.0051963492296636105, + 0.007375125773251057, + 0.004045342560857534, + 0.0033934053499251604, + 0.02093968167901039, + 0.095135897397995, + 0.011281011626124382, + 0.010161031037569046, + 0.05368901789188385, + 0.11341497302055359, + 0.007527002599090338, + 0.010969975031912327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.006736780982464552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly overhead medium shot, then slowly pans forward to focus on a professional chef in a crisp white uniform and tall hat. He concentrates, wielding a gleaming stainless steel ladle, carefully pouring a ladleful of rich, amber-colored broth into a dark stoneware casserole heating on the stovetop. As the hot broth pours in, the meat and lush green vegetables already in the pot gently tumble in the bubbling broth. Puffs of white steam rise from the pot's rim, illuminating it with a warm yellow spotlight overhead, transforming into a visible, lingering cloud of aroma. The camera steadily pans forward, ultimately capturing an extreme close-up of the bubbling, oily broth within the casserole, each bubble visible as it bursts. The chef's profile is clearly outlined. He stares intently at the food in the pot, his eyes full of precise control of the heat. The whole picture, set against warm colors, creates a warm, quiet and craftsman-like cooking atmosphere, as if you can smell the tempting aroma through the screen.", + "seed": 26459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_5/index_9209_seed_26459_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": -0.028858570381999016, + "vlm_logprob": -0.028858570381999016, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.028858570381999016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03178735449910164 + ], + "opposite_logprob_tokens": [ + -3.531787395477295 + ], + "logprob_reward": -0.03178735449910164, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03178735449910164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18196353316307068, + 0.07199395447969437, + 0.46815118193626404, + 0.11537744104862213, + 0.027316104620695114, + 0.02701718732714653, + 0.055442240089178085, + 0.02301269955933094, + 0.011004387401044369, + 0.018721280619502068 + ], + "credit_map": [ + 0.013835291378200054, + 0.031077861785888672, + 0.030735185369849205, + 0.017605016008019447, + 0.04076400026679039, + 0.016668541356921196, + 0.017691493034362793, + 0.013586149550974369, + 0.007266082800924778, + 0.01189072709530592, + 0.005084832198917866, + 0.005593786481767893, + 0.013093744404613972, + 0.00997344683855772, + 0.00983232818543911, + 0.00925900973379612, + 0.047965023666620255, + 0.08907811343669891, + 0.05714140087366104, + 0.07199067622423172, + 0.0711013451218605, + 0.04652077704668045, + 0.0378619022667408, + 0.04649196192622185, + 0.009454664774239063, + 0.021077381446957588, + 0.00858880765736103, + 0.012024537660181522, + 0.023328205570578575, + 0.010932429693639278, + 0.01701996847987175, + 0.012951444834470749, + 0.001885353121906519, + 0.0020763406064361334, + 0.0024688648991286755, + 0.0021740177180618048, + 0.004586501512676477, + 0.0062299929559230804, + 0.0043158335611224174, + 0.0035792007111012936, + 0.002397296018898487, + 0.0030083158053457737, + 0.006598576437681913, + 0.003200765233486891, + 0.002941835206001997, + 0.0035030983854085207, + 0.002603042172268033, + 0.0027642585337162018, + 0.006799119990319014, + 0.004923871252685785, + 0.006043783389031887, + 0.00894701573997736, + 0.0076010385528206825, + 0.005886163096874952, + 0.010601739399135113, + 0.004639504011720419, + 0.002617300720885396, + 0.0029306274373084307, + 0.0031116323079913855, + 0.003016719361767173, + 0.002777884714305401, + 0.0023232256062328815, + 0.0036357210483402014, + 0.0025995883625000715, + 0.001288069412112236, + 0.001618488458916545, + 0.0011370712891221046, + 0.001512274146080017, + 0.0013625548454001546, + 0.0010148095898330212, + 0.001842060824856162, + 0.0012290594168007374, + 0.003591199405491352, + 0.0029905391857028008, + 0.0017601397121325135, + 0.0015366242732852697, + 0.0034142632503062487, + 0.0019884391222149134, + 0.0012112018885090947, + 0.0022288740146905184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.03178735449910164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038139987736940384 + ], + "opposite_logprob_tokens": [ + -3.288140058517456 + ], + "logprob_reward": -0.038139987736940384, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038139987736940384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20849215984344482, + 0.24645808339118958, + 0.03106086514890194, + 0.03588482365012169, + 0.026825735345482826, + 0.12960031628608704, + 0.11671799421310425, + 0.07953327149152756, + 0.06199170649051666, + 0.06343512237071991 + ], + "credit_map": [ + 0.01608663983643055, + 0.05647755786776543, + 0.02221645973622799, + 0.013330744579434395, + 0.05698729306459427, + 0.01631876640021801, + 0.015742763876914978, + 0.011331942863762379, + 0.024108337238430977, + 0.03563637286424637, + 0.019934402778744698, + 0.020058663561940193, + 0.04072944447398186, + 0.03631686419248581, + 0.03695204481482506, + 0.032721955329179764, + 0.004105083644390106, + 0.003460085252299905, + 0.0067947362549602985, + 0.0033208741806447506, + 0.0034394743852317333, + 0.003447480034083128, + 0.0027738430071622133, + 0.0037192869931459427, + 0.002467111684381962, + 0.00537237199023366, + 0.004067661706358194, + 0.0028044653590768576, + 0.009491750970482826, + 0.003470241092145443, + 0.004636185709387064, + 0.0035750356037169695, + 0.0022514171432703733, + 0.0026934598572552204, + 0.003858136245980859, + 0.0031767801847308874, + 0.004092975985258818, + 0.00381072866730392, + 0.003224707208573818, + 0.003717528423294425, + 0.010280968621373177, + 0.013065706938505173, + 0.029684582725167274, + 0.019574327394366264, + 0.01701935939490795, + 0.017100123688578606, + 0.012552031315863132, + 0.010323219932615757, + 0.009275687858462334, + 0.012992496602237225, + 0.013099719770252705, + 0.03458603471517563, + 0.01244072150439024, + 0.015224425122141838, + 0.01345264632254839, + 0.005646254401654005, + 0.010202857665717602, + 0.007813474163413048, + 0.012848300859332085, + 0.012160763144493103, + 0.010539202019572258, + 0.006667254026979208, + 0.012221927754580975, + 0.007079496514052153, + 0.006739744916558266, + 0.0067177326418459415, + 0.007279964629560709, + 0.007427970878779888, + 0.006589872296899557, + 0.006767026148736477, + 0.012588012963533401, + 0.00788138434290886, + 0.010067139752209187, + 0.011958952061831951, + 0.005764067638665438, + 0.0060317786410450935, + 0.012115144170820713, + 0.006169353146106005, + 0.004365246742963791, + 0.006963436491787434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.038139987736940384 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029817067086696625 + ], + "opposite_logprob_tokens": [ + -3.5298171043395996 + ], + "logprob_reward": -0.029817067086696625, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029817067086696625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06834544986486435, + 0.03710765019059181, + 0.042911987751722336, + 0.06036648526787758, + 0.013643482699990273, + 0.06283945590257645, + 0.37112540006637573, + 0.04910631105303764, + 0.15300561487674713, + 0.14154809713363647 + ], + "credit_map": [ + 0.006365817505866289, + 0.009355591610074043, + 0.009471401572227478, + 0.005988932214677334, + 0.016911374405026436, + 0.006715590134263039, + 0.008115805685520172, + 0.005420938599854708, + 0.003266502870246768, + 0.005575132090598345, + 0.0028332569636404514, + 0.0031562414951622486, + 0.006263669114559889, + 0.005724179092794657, + 0.005335539113730192, + 0.004953132476657629, + 0.005746437236666679, + 0.005732934456318617, + 0.008589467965066433, + 0.006835303269326687, + 0.0050173597410321236, + 0.0030881070997565985, + 0.0035068588331341743, + 0.004395522177219391, + 0.005433300044387579, + 0.008863179013133049, + 0.008935784921050072, + 0.006746039260178804, + 0.006604616530239582, + 0.006294839549809694, + 0.00917481817305088, + 0.008313914760947227, + 0.001558735384605825, + 0.001573258196003735, + 0.0016559642972424626, + 0.0011919147800654173, + 0.002333841985091567, + 0.002068255329504609, + 0.001738797058351338, + 0.0015227171825245023, + 0.005838950630277395, + 0.009560259990394115, + 0.012518486939370632, + 0.0091122230514884, + 0.006952701602131128, + 0.007549019996076822, + 0.006077984347939491, + 0.005229840520769358, + 0.048170629888772964, + 0.040020786225795746, + 0.04114062711596489, + 0.0669863298535347, + 0.05283256247639656, + 0.030703293159604073, + 0.0640212669968605, + 0.027249926701188087, + 0.005036919843405485, + 0.006138389930129051, + 0.005328564904630184, + 0.007324389647692442, + 0.006010457407683134, + 0.004317448008805513, + 0.010526095516979694, + 0.004424049519002438, + 0.018264828249812126, + 0.014445381239056587, + 0.019873185083270073, + 0.017929254099726677, + 0.017285043373703957, + 0.014640911482274532, + 0.03514218330383301, + 0.015424846671521664, + 0.02224215678870678, + 0.03044578991830349, + 0.01291849184781313, + 0.015543578192591667, + 0.023611238226294518, + 0.01703670807182789, + 0.01085344236344099, + 0.008896701969206333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.029817067086696625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197370693087578 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197370693087578, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197370693087578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07250498980283737, + 0.08734817057847977, + 0.05418825522065163, + 0.0688067227602005, + 0.051059748977422714, + 0.1149551123380661, + 0.426962286233902, + 0.05326873064041138, + 0.04013344645500183, + 0.030772510915994644 + ], + "credit_map": [ + 0.007012845017015934, + 0.015092357061803341, + 0.011968601495027542, + 0.007312936242669821, + 0.014676080085337162, + 0.005078219808638096, + 0.0059201717376708984, + 0.0054437825456261635, + 0.007722977548837662, + 0.013735579326748848, + 0.007604282349348068, + 0.008544202893972397, + 0.013708080165088177, + 0.012745186686515808, + 0.01180841401219368, + 0.011479453183710575, + 0.004004110582172871, + 0.008609751239418983, + 0.007246929686516523, + 0.004764980636537075, + 0.006671782582998276, + 0.006954630836844444, + 0.006646488327533007, + 0.009289584122598171, + 0.007135732099413872, + 0.015188043005764484, + 0.007572532165795565, + 0.009334227070212364, + 0.00895934458822012, + 0.006146891973912716, + 0.008676920086145401, + 0.005793038289994001, + 0.0059401970356702805, + 0.0048904684372246265, + 0.007984736934304237, + 0.00656325276941061, + 0.005992658901959658, + 0.006338620092719793, + 0.009041091427206993, + 0.004308726172894239, + 0.008767064660787582, + 0.01187217514961958, + 0.026045840233564377, + 0.0172555074095726, + 0.01089180912822485, + 0.018154505640268326, + 0.011193634942173958, + 0.01077458169311285, + 0.06269094347953796, + 0.049595121294260025, + 0.05462716892361641, + 0.06571897119283676, + 0.05149383097887039, + 0.032412804663181305, + 0.07739046961069107, + 0.033032990992069244, + 0.006687176879495382, + 0.0055834827944636345, + 0.007478326093405485, + 0.00982539914548397, + 0.006121876183897257, + 0.0035728418733924627, + 0.008994192816317081, + 0.005005440209060907, + 0.005072435364127159, + 0.004335320089012384, + 0.003129993798211217, + 0.004217426758259535, + 0.004018811509013176, + 0.004380757454782724, + 0.01068483293056488, + 0.004293870180845261, + 0.006204567849636078, + 0.005736250896006823, + 0.002428776351734996, + 0.0024172666016966105, + 0.005073666572570801, + 0.003380879759788513, + 0.0022723772563040257, + 0.0032587284222245216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.018197370693087578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a professional chef wearing a crisp white uniform and a tall hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chef holding a gleaming stainless steel ladle and pouring amber-colored broth into a dark stoneware casserole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show meat and lush green vegetables tumbling in the bubbling broth inside the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a slightly overhead medium shot and slowly pan forward to an extreme close-up of the broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are puffs of white steam rising from the pot's rim, illuminated by a warm yellow spotlight overhead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351071894168854 + ], + "opposite_logprob_tokens": [ + -3.651350975036621 + ], + "logprob_reward": -0.026351071894168854, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351071894168854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16780202090740204, + 0.10154205560684204, + 0.04881960153579712, + 0.09168409556150436, + 0.043432947248220444, + 0.07873323559761047, + 0.11211124807596207, + 0.10767442733049393, + 0.1952814757823944, + 0.0529189258813858 + ], + "credit_map": [ + 0.01409158855676651, + 0.038298558443784714, + 0.029148900881409645, + 0.01376713439822197, + 0.038839519023895264, + 0.01059431117027998, + 0.012428509071469307, + 0.010633501224219799, + 0.010351952165365219, + 0.012444681487977505, + 0.018534362316131592, + 0.013006572611629963, + 0.00982939638197422, + 0.013801051303744316, + 0.013465113006532192, + 0.01010892540216446, + 0.005956181790679693, + 0.0061270506121218204, + 0.007013783790171146, + 0.005761015694588423, + 0.005496948957443237, + 0.003495874349027872, + 0.006768302991986275, + 0.008200444281101227, + 0.008898855186998844, + 0.013108903542160988, + 0.018148476257920265, + 0.011938921175897121, + 0.010704469867050648, + 0.0074026985093951225, + 0.01192519348114729, + 0.009556571021676064, + 0.004682608880102634, + 0.005621221847832203, + 0.005434046499431133, + 0.004471797961741686, + 0.00519939698278904, + 0.0058818235993385315, + 0.00856363121420145, + 0.0035784190986305475, + 0.007408270612359047, + 0.010374770499765873, + 0.014032946899533272, + 0.009841843508183956, + 0.011230559088289738, + 0.013104707933962345, + 0.0064779845997691154, + 0.006262160837650299, + 0.011030315421521664, + 0.012555220164358616, + 0.01228544395416975, + 0.029789285734295845, + 0.01015503890812397, + 0.018068920820951462, + 0.011749030090868473, + 0.00647799763828516, + 0.012693137861788273, + 0.013249026611447334, + 0.013477598316967487, + 0.01905214414000511, + 0.01495271548628807, + 0.010629136115312576, + 0.013084600679576397, + 0.010536066256463528, + 0.02202947437763214, + 0.023549627512693405, + 0.029708756133913994, + 0.0183867234736681, + 0.02748952805995941, + 0.02408444695174694, + 0.035782210528850555, + 0.014250699430704117, + 0.010316533967852592, + 0.01015230268239975, + 0.005119629204273224, + 0.004184381105005741, + 0.009493253193795681, + 0.00517302704975009, + 0.0035518608056008816, + 0.004927936941385269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.026351071894168854 + } + ], + "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/0000022/group_6/index_tensor([8300, 8300])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_tensor([8300, 8300])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cbf45cc58c09f0784044bf32b2894b44fcb88ba9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_tensor([8300, 8300])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a medium shot at eye level, the camera moves smoothly to the right, tracking a newspaper as it glides almost weightlessly across a cityscape. The paper, a slightly yellowed broadsheet with frayed edges, billows and twirls in the breeze, its fibrous texture visible as it momentarily catches the light, revealing a bold black headline and a faint impression of a faded monochrome photograph. Its graceful journey is emphasized by a natural motion blur that softens its edges and background, focusing the viewer's attention on its fluid, dance-like movement as it undulates. The scene is set on a quiet city street corner during golden hour, where warm, low-angle sunlight flows between brick buildings and casts long, dramatic shadows across the wet asphalt. The predominant color palette is warm and cinematic, with the golden light illuminating the dust in the air and making the newspaper the brightest element in the frame, creating an overall atmosphere of quiet, poetic solitude that captures a fleeting moment of unexpected beauty.", + "seed": 14645, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14645_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": -0.07746291672810912, + "vlm_logprob": -0.07746291672810912, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.07746291672810912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17055463790893555 + ], + "opposite_logprob_tokens": [ + -2.0455546379089355 + ], + "logprob_reward": -0.17055463790893555, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17055463790893555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29507607221603394, + 0.07832660526037216, + 0.07038254290819168, + 0.18605457246303558, + 0.09942881017923355, + 0.06685472279787064, + 0.034291427582502365, + 0.027279235422611237, + 0.061561278998851776, + 0.080744668841362 + ], + "credit_map": [ + 0.05458977818489075, + 0.025465810671448708, + 0.017023207619786263, + 0.014912671409547329, + 0.02912749908864498, + 0.017322612926363945, + 0.11561103165149689, + 0.0210234597325325, + 0.007042220328003168, + 0.016887513920664787, + 0.007795034442096949, + 0.0071649993769824505, + 0.009741239249706268, + 0.006970737129449844, + 0.01163026038557291, + 0.011094602756202221, + 0.008610451593995094, + 0.009790265932679176, + 0.012220962904393673, + 0.009460321627557278, + 0.006998424883931875, + 0.007746806833893061, + 0.008437750861048698, + 0.007117558736354113, + 0.0306363757699728, + 0.011969279497861862, + 0.0161614790558815, + 0.014168011955916882, + 0.01691783219575882, + 0.013297727331519127, + 0.06496526300907135, + 0.01793859526515007, + 0.009513722732663155, + 0.00865140650421381, + 0.006999594159424305, + 0.010772579349577427, + 0.011926824226975441, + 0.01579187996685505, + 0.025877535343170166, + 0.009895261377096176, + 0.005309718661010265, + 0.00725560775026679, + 0.0054083517752587795, + 0.011347012594342232, + 0.00840784516185522, + 0.007301318924874067, + 0.012549251317977905, + 0.00927562266588211, + 0.0023849839344620705, + 0.0026248209178447723, + 0.0024722216185182333, + 0.004836500622332096, + 0.00280204089358449, + 0.007733130361884832, + 0.004480474162846804, + 0.006957257632166147, + 0.0030249522533267736, + 0.00252667092718184, + 0.0030045367311686277, + 0.00394629780203104, + 0.0034625502303242683, + 0.0038379954639822245, + 0.0037195126060396433, + 0.0037567189428955317, + 0.004865431692451239, + 0.00574768241494894, + 0.004623815882951021, + 0.008791579864919186, + 0.010240419767796993, + 0.011461673304438591, + 0.005815599579364061, + 0.010015075094997883, + 0.007550222799181938, + 0.00731274951249361, + 0.005431820638477802, + 0.0062632085755467415, + 0.008293229155242443, + 0.022142205387353897, + 0.006657790392637253, + 0.017093438655138016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.17055463790893555 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003468781709671 + ], + "opposite_logprob_tokens": [ + -2.350346803665161 + ], + "logprob_reward": -0.1003468781709671, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003468781709671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25682100653648376, + 0.057974375784397125, + 0.037243545055389404, + 0.30190715193748474, + 0.19516050815582275, + 0.04605131596326828, + 0.020914366468787193, + 0.028806954622268677, + 0.02755139395594597, + 0.027569396421313286 + ], + "credit_map": [ + 0.05434431508183479, + 0.01261829398572445, + 0.009440570138394833, + 0.012772549875080585, + 0.02743993140757084, + 0.01487525925040245, + 0.10956401377916336, + 0.015766069293022156, + 0.0077733928337693214, + 0.018993576988577843, + 0.005146161187440157, + 0.004089566878974438, + 0.004641958978027105, + 0.003943655174225569, + 0.008800034411251545, + 0.004586031194776297, + 0.004477428272366524, + 0.006170184817165136, + 0.004761932417750359, + 0.004358314909040928, + 0.0035157278180122375, + 0.0032661783043295145, + 0.006912885699421167, + 0.0037808918859809637, + 0.03254404664039612, + 0.031319279223680496, + 0.03402843698859215, + 0.032933179289102554, + 0.031364984810352325, + 0.06637588143348694, + 0.03494312986731529, + 0.03839823976159096, + 0.013323347084224224, + 0.012513307854533195, + 0.008040853776037693, + 0.010328160598874092, + 0.00859745405614376, + 0.017624327912926674, + 0.06224256753921509, + 0.06249048560857773, + 0.0033896879758685827, + 0.004406764637678862, + 0.0027986159548163414, + 0.005589894950389862, + 0.0051194289699196815, + 0.006006424780935049, + 0.009041168726980686, + 0.009699331596493721, + 0.0020690266974270344, + 0.0023486469872295856, + 0.001258490956388414, + 0.004512193147093058, + 0.0022893026471138, + 0.003791163442656398, + 0.002465516095981002, + 0.002180026611313224, + 0.0024196647573262453, + 0.0033920472487807274, + 0.0020123948343098164, + 0.0030733775347471237, + 0.00524023175239563, + 0.005188103765249252, + 0.0032261896412819624, + 0.004254946950823069, + 0.002956368960440159, + 0.0029335857834666967, + 0.0034686897415667772, + 0.004350967239588499, + 0.0032889104913920164, + 0.004175597336143255, + 0.00266924942843616, + 0.0037080254405736923, + 0.0015977059956640005, + 0.0019475214648991823, + 0.0012741569662466645, + 0.0026071779429912567, + 0.002183752367272973, + 0.007990394718945026, + 0.001935065258294344, + 0.008033622987568378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.1003468781709671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04205726459622383, + 0.18917417526245117, + 0.07883831113576889, + 0.338631808757782, + 0.23565436899662018, + 0.024665122851729393, + 0.014163353480398655, + 0.008659121580421925, + 0.023262951523065567, + 0.04489355906844139 + ], + "credit_map": [ + 0.009014870040118694, + 0.005410144105553627, + 0.004287043586373329, + 0.0031244803685694933, + 0.007625178899616003, + 0.0031384839676320553, + 0.004543014336377382, + 0.004914045333862305, + 0.0238724946975708, + 0.06227606534957886, + 0.014322110451757908, + 0.011246574111282825, + 0.016819367185235023, + 0.015734601765871048, + 0.02449440024793148, + 0.020408539101481438, + 0.009858573786914349, + 0.007661279756575823, + 0.013495400547981262, + 0.010709384456276894, + 0.007771805860102177, + 0.006708158180117607, + 0.014381233602762222, + 0.008252459578216076, + 0.03422204405069351, + 0.03487404063344002, + 0.03531978651881218, + 0.034372370690107346, + 0.03344392403960228, + 0.06826256960630417, + 0.05649587884545326, + 0.04164116084575653, + 0.019966669380664825, + 0.020125064998865128, + 0.016663648188114166, + 0.022046735510230064, + 0.019137324765324593, + 0.02094687893986702, + 0.06410710513591766, + 0.05266091227531433, + 0.0016478430479764938, + 0.0023108769673854113, + 0.0017287269001826644, + 0.002835866529494524, + 0.003128765383735299, + 0.00311090424656868, + 0.0054692672565579414, + 0.004432870540767908, + 0.0015754541382193565, + 0.0015821553533896804, + 0.0011309967376291752, + 0.0010951379081234336, + 0.0018800239777192473, + 0.0025596243795007467, + 0.0011931691551581025, + 0.0031467913649976254, + 0.0006665369146503508, + 0.0010625888826325536, + 0.0008194160182029009, + 0.0009197547915391624, + 0.0013148117577657104, + 0.001267047831788659, + 0.0011395213659852743, + 0.0014694436686113477, + 0.002348906360566616, + 0.0026010479778051376, + 0.0020930066239088774, + 0.0033930442295968533, + 0.0025996204931288958, + 0.0033562725875526667, + 0.002670461544767022, + 0.004200591240078211, + 0.0037833615206182003, + 0.004659998696297407, + 0.003060870571061969, + 0.003978702239692211, + 0.004858261905610561, + 0.010723799467086792, + 0.003489819588139653, + 0.010338744148612022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.007636047434061766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898228615522385 + ], + "opposite_logprob_tokens": [ + -2.578982353210449 + ], + "logprob_reward": -0.07898228615522385, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898228615522385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09613174945116043, + 0.10072994977235794, + 0.11988924443721771, + 0.052660539746284485, + 0.3313501179218292, + 0.029002197086811066, + 0.04481565207242966, + 0.030356908217072487, + 0.09473751485347748, + 0.10032621026039124 + ], + "credit_map": [ + 0.018367478623986244, + 0.008677314035594463, + 0.007019607815891504, + 0.005517978221178055, + 0.0076978872530162334, + 0.008702713996171951, + 0.031840626150369644, + 0.008308141492307186, + 0.012257302179932594, + 0.03235529735684395, + 0.007630887906998396, + 0.006099632475525141, + 0.00927793700248003, + 0.00855755340307951, + 0.012894915416836739, + 0.011656420305371284, + 0.014501042664051056, + 0.013081020675599575, + 0.02098742686212063, + 0.017011621966958046, + 0.011602791026234627, + 0.011089623905718327, + 0.018729163333773613, + 0.012886553071439266, + 0.0039137606509029865, + 0.004897512961179018, + 0.004400600679218769, + 0.004317330662161112, + 0.005975637119263411, + 0.015479757450520992, + 0.0064332857728004456, + 0.0072426563128829, + 0.025729596614837646, + 0.02671976201236248, + 0.01991298608481884, + 0.029099389910697937, + 0.023345021530985832, + 0.022461747750639915, + 0.09719756990671158, + 0.0868840366601944, + 0.003453907324001193, + 0.0027920713182538748, + 0.0038272435776889324, + 0.0023024356923997402, + 0.002727220533415675, + 0.003801891580224037, + 0.004673266317695379, + 0.00542416051030159, + 0.0035125843714922667, + 0.001935478183440864, + 0.0022585480473935604, + 0.001848712912760675, + 0.0035811110865324736, + 0.013855665922164917, + 0.0021582022309303284, + 0.01566535234451294, + 0.0017875728663057089, + 0.0023225126788020134, + 0.0023112231865525246, + 0.002742534037679434, + 0.004835639148950577, + 0.007562389597296715, + 0.0023796907626092434, + 0.006415345706045628, + 0.008265461772680283, + 0.01030812133103609, + 0.007520279847085476, + 0.014153198339045048, + 0.010895796120166779, + 0.014127890579402447, + 0.011880272999405861, + 0.017586497589945793, + 0.009768573567271233, + 0.012927521020174026, + 0.007225248962640762, + 0.010187739506363869, + 0.011313002556562424, + 0.022965285927057266, + 0.007656223140656948, + 0.018282612785696983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.07898228615522385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794733971357346 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794733971357346, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794733971357346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0725298523902893, + 0.2335180640220642, + 0.04384498670697212, + 0.0964851900935173, + 0.3061874210834503, + 0.031143715605139732, + 0.05794356018304825, + 0.01955539919435978, + 0.018892226740717888, + 0.11989954113960266 + ], + "credit_map": [ + 0.01188601553440094, + 0.0069277225993573666, + 0.005823719780892134, + 0.004977331962436438, + 0.008974726311862469, + 0.011862698942422867, + 0.014182955957949162, + 0.007894689217209816, + 0.030276022851467133, + 0.08114498853683472, + 0.016428537666797638, + 0.014408107846975327, + 0.019848976284265518, + 0.016968553885817528, + 0.030530642718076706, + 0.02391224168241024, + 0.005620144307613373, + 0.006531454622745514, + 0.007504088804125786, + 0.005915168207138777, + 0.0038086636923253536, + 0.004829420242458582, + 0.0051637087017297745, + 0.004472334869205952, + 0.008000542409718037, + 0.008314086124300957, + 0.007643637713044882, + 0.009671276435256004, + 0.011473665945231915, + 0.027121860533952713, + 0.01172773726284504, + 0.012532386928796768, + 0.022084763273596764, + 0.02167435735464096, + 0.015591531991958618, + 0.022786444053053856, + 0.01783418096601963, + 0.02523750066757202, + 0.09278302639722824, + 0.08819562196731567, + 0.0024678127374500036, + 0.0029790126718580723, + 0.002303828252479434, + 0.004479914903640747, + 0.004220691043883562, + 0.004307536408305168, + 0.007396141067147255, + 0.002988777356222272, + 0.004757589660584927, + 0.004793993663042784, + 0.0035689296200871468, + 0.008643774315714836, + 0.0059844800271093845, + 0.01419563964009285, + 0.003923070151358843, + 0.012076086364686489, + 0.0015700326766818762, + 0.0019095424795523286, + 0.0016940042842179537, + 0.0024314941838383675, + 0.0020586501341313124, + 0.003802195657044649, + 0.0024459795095026493, + 0.003643499920144677, + 0.0016330064972862601, + 0.0022348270285874605, + 0.0018680697539821267, + 0.0024440118577331305, + 0.0037633851170539856, + 0.0028316190000623465, + 0.0019443202763795853, + 0.0021729867439717054, + 0.008171509020030499, + 0.011108835227787495, + 0.00674669723957777, + 0.010073576122522354, + 0.01005452312529087, + 0.033767543733119965, + 0.007497966289520264, + 0.03247889503836632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.029794733971357346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot at eye level, the camera moves smoothly to the right, tracking a newspaper as it glides almost weightlessly across a cityscape. The paper, a slightly yellowed broadsheet with frayed edges, billows and twirls in the breeze, its fibrous texture visible as it momentarily catches the light, revealing a bold black headline and a faint impression of a faded monochrome photograph. Its graceful journey is emphasized by a natural motion blur that softens its edges and background, focusing the viewer's attention on its fluid, dance-like movement as it undulates. The scene is set on a quiet city street corner during golden hour, where warm, low-angle sunlight flows between brick buildings and casts long, dramatic shadows across the wet asphalt. The predominant color palette is warm and cinematic, with the golden light illuminating the dust in the air and making the newspaper the brightest element in the frame, creating an overall atmosphere of quiet, poetic solitude that captures a fleeting moment of unexpected beauty.", + "seed": 14646, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14646_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": -0.10184730365872383, + "vlm_logprob": -0.10184730365872383, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.10184730365872383, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37382519245147705 + ], + "opposite_logprob_tokens": [ + -1.248825192451477 + ], + "logprob_reward": -0.37382519245147705, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37382519245147705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11386103928089142, + 0.05074233561754227, + 0.05981044843792915, + 0.23486016690731049, + 0.041920363903045654, + 0.043402787297964096, + 0.11419584602117538, + 0.2991195619106293, + 0.03198784962296486, + 0.010099602863192558 + ], + "credit_map": [ + 0.010690093040466309, + 0.01978692039847374, + 0.007727291435003281, + 0.01382820587605238, + 0.019354859367012978, + 0.013424104079604149, + 0.01563453860580921, + 0.013415032997727394, + 0.005264555104076862, + 0.010209541767835617, + 0.0031457587610930204, + 0.0038098073564469814, + 0.006703756283968687, + 0.0070341178216040134, + 0.006620706059038639, + 0.007954096421599388, + 0.005975452251732349, + 0.00964772142469883, + 0.005035775247961283, + 0.0069129555486142635, + 0.009737599641084671, + 0.006755203008651733, + 0.006456326227635145, + 0.009289417415857315, + 0.01889736019074917, + 0.04659042879939079, + 0.0171621423214674, + 0.023240363225340843, + 0.032422807067632675, + 0.040321506559848785, + 0.026669073849916458, + 0.029556507244706154, + 0.0033301382791250944, + 0.007977752014994621, + 0.0027892242651432753, + 0.004312620498239994, + 0.006404468789696693, + 0.005154014099389315, + 0.005394520703703165, + 0.006557627581059933, + 0.0028099212795495987, + 0.005451940465718508, + 0.0031650103628635406, + 0.0073904781602323055, + 0.007460149470716715, + 0.006245604250580072, + 0.006424581632018089, + 0.004455102141946554, + 0.006936382036656141, + 0.013937989249825478, + 0.006271119229495525, + 0.006489034276455641, + 0.005885590799152851, + 0.026121877133846283, + 0.009479093365371227, + 0.03907477483153343, + 0.026406902819871902, + 0.054232675582170486, + 0.021207407116889954, + 0.033062439411878586, + 0.04075334221124649, + 0.04389648139476776, + 0.041404545307159424, + 0.03815576806664467, + 0.0033649387769401073, + 0.00410419562831521, + 0.002556098625063896, + 0.0029132941272109747, + 0.0062254080548882484, + 0.004371133632957935, + 0.003925790078938007, + 0.004526991862803698, + 0.0009038468706421554, + 0.001568308100104332, + 0.0007848748937249184, + 0.0008110792259685695, + 0.0011119961272925138, + 0.0019341122824698687, + 0.0014496129006147385, + 0.0015357735101133585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.37382519245147705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043101537972688675 + ], + "opposite_logprob_tokens": [ + -3.1681015491485596 + ], + "logprob_reward": -0.043101537972688675, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043101537972688675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17913484573364258, + 0.08332055062055588, + 0.08493147045373917, + 0.12662506103515625, + 0.10383636504411697, + 0.04343368113040924, + 0.09584739059209824, + 0.11922736465930939, + 0.11736112087965012, + 0.046282146126031876 + ], + "credit_map": [ + 0.01710064336657524, + 0.02064322680234909, + 0.012702031061053276, + 0.0197149608284235, + 0.0264730267226696, + 0.026578066870570183, + 0.02667185105383396, + 0.029251037165522575, + 0.008704124018549919, + 0.016438225284218788, + 0.0056783235631883144, + 0.007056542206555605, + 0.010622977279126644, + 0.010437516495585442, + 0.011641699820756912, + 0.012741148471832275, + 0.006968525238335133, + 0.015327042900025845, + 0.005980303511023521, + 0.009360195137560368, + 0.015881923958659172, + 0.009991142898797989, + 0.010371902026236057, + 0.011050433851778507, + 0.009924853220582008, + 0.023902758955955505, + 0.009503278881311417, + 0.01512770727276802, + 0.01677776500582695, + 0.021261269226670265, + 0.014039899222552776, + 0.01608753576874733, + 0.00867408886551857, + 0.01819007284939289, + 0.005085894837975502, + 0.008323191665112972, + 0.017942819744348526, + 0.01761891320347786, + 0.014380461536347866, + 0.013620913960039616, + 0.0038932389579713345, + 0.007132370490580797, + 0.005207928363233805, + 0.0067698475904762745, + 0.006061697844415903, + 0.005279775708913803, + 0.004618339706212282, + 0.004470484331250191, + 0.009379768744111061, + 0.014992041513323784, + 0.008544153533875942, + 0.008853905834257603, + 0.009139386005699635, + 0.0159181859344244, + 0.014237107709050179, + 0.01478283666074276, + 0.009336714632809162, + 0.02300341986119747, + 0.009285656735301018, + 0.01590120606124401, + 0.0190016757696867, + 0.012983807362616062, + 0.013669774867594242, + 0.0160451028496027, + 0.009967262856662273, + 0.014295443892478943, + 0.0062178666703403, + 0.008628370240330696, + 0.0316525362432003, + 0.01479562558233738, + 0.01514560729265213, + 0.01665840856730938, + 0.0035381990019232035, + 0.006997165735810995, + 0.0025956002064049244, + 0.005930730141699314, + 0.005007777828723192, + 0.0067980303429067135, + 0.006271961145102978, + 0.00914268009364605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.043101537972688675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203056693077087 + ], + "opposite_logprob_tokens": [ + -4.264203071594238 + ], + "logprob_reward": -0.014203056693077087, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203056693077087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12467692792415619, + 0.05471299961209297, + 0.22368355095386505, + 0.03817799314856529, + 0.06574689596891403, + 0.046647436916828156, + 0.10229315608739853, + 0.1477416753768921, + 0.12817111611366272, + 0.06814827769994736 + ], + "credit_map": [ + 0.011180014349520206, + 0.0183379128575325, + 0.007525659631937742, + 0.016577864065766335, + 0.02278861403465271, + 0.016278594732284546, + 0.013242349959909916, + 0.018745914101600647, + 0.005622438155114651, + 0.012960112653672695, + 0.007824692875146866, + 0.004667145665735006, + 0.007204999681562185, + 0.005300172604620457, + 0.005865495651960373, + 0.005267936736345291, + 0.022260908037424088, + 0.05976272374391556, + 0.0141132902354002, + 0.024536047130823135, + 0.04111219197511673, + 0.015565433539450169, + 0.025584300979971886, + 0.020748646929860115, + 0.0023040869273245335, + 0.00723234424367547, + 0.00262522348202765, + 0.00739670917391777, + 0.004438101314008236, + 0.005106558557599783, + 0.004166220314800739, + 0.004908745642751455, + 0.006898428313434124, + 0.010623641312122345, + 0.00803933385759592, + 0.005112044978886843, + 0.00782030075788498, + 0.008926224894821644, + 0.009474081918597221, + 0.008852836675941944, + 0.003998665604740381, + 0.008345273323357105, + 0.0040641664527356625, + 0.005509801208972931, + 0.00812502857297659, + 0.006585943512618542, + 0.005223668646067381, + 0.004794882610440254, + 0.01123468391597271, + 0.01798425056040287, + 0.006712508853524923, + 0.009649498388171196, + 0.010290354490280151, + 0.013941947370767593, + 0.01782279647886753, + 0.014657106250524521, + 0.012417959980666637, + 0.022786874324083328, + 0.008592995814979076, + 0.017760010436177254, + 0.03176847845315933, + 0.019393453374505043, + 0.01901278831064701, + 0.016009097918868065, + 0.013493791222572327, + 0.015323895029723644, + 0.009078394621610641, + 0.010612383484840393, + 0.026004573330283165, + 0.01739727519452572, + 0.015447140671312809, + 0.02081364579498768, + 0.0043813008815050125, + 0.007868844084441662, + 0.004896529950201511, + 0.010490156710147858, + 0.008844975382089615, + 0.012881607748568058, + 0.008591176941990852, + 0.01019368041306734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.014203056693077087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204469874501228 + ], + "opposite_logprob_tokens": [ + -2.812044620513916 + ], + "logprob_reward": -0.06204469874501228, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204469874501228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11175903677940369, + 0.04661113768815994, + 0.26561957597732544, + 0.046379558742046356, + 0.06568529456853867, + 0.07529756426811218, + 0.24254047870635986, + 0.0491207055747509, + 0.06636736541986465, + 0.03061925806105137 + ], + "credit_map": [ + 0.010473079048097134, + 0.01633608713746071, + 0.011264649219810963, + 0.011128732934594154, + 0.013584855943918228, + 0.016516655683517456, + 0.016253598034381866, + 0.016201384365558624, + 0.004015882965177298, + 0.009837519377470016, + 0.005221168044954538, + 0.003594276960939169, + 0.006919518578797579, + 0.004419095814228058, + 0.0063390061259269714, + 0.006264665629714727, + 0.025683147832751274, + 0.04599757120013237, + 0.015231789089739323, + 0.0259696152061224, + 0.05331636592745781, + 0.030968716368079185, + 0.031532786786556244, + 0.03691959008574486, + 0.003975764848291874, + 0.008252520114183426, + 0.0035543558187782764, + 0.0070254141464829445, + 0.006162568926811218, + 0.006401585415005684, + 0.006098361685872078, + 0.004908986389636993, + 0.0066496338695287704, + 0.013063397258520126, + 0.007676653563976288, + 0.0059609985910356045, + 0.008193415589630604, + 0.007286637090146542, + 0.007970036938786507, + 0.008884517475962639, + 0.0056542893871665, + 0.009824822656810284, + 0.006927952636033297, + 0.013177826069295406, + 0.012007731944322586, + 0.008957711979746819, + 0.012008574791252613, + 0.006738652940839529, + 0.022832700982689857, + 0.0441892147064209, + 0.01748259738087654, + 0.019217750057578087, + 0.024256765842437744, + 0.04796528071165085, + 0.02543766051530838, + 0.04115850850939751, + 0.0046927547082304955, + 0.006417697761207819, + 0.003688369644805789, + 0.006399017758667469, + 0.008156029507517815, + 0.006293791811913252, + 0.007516639307141304, + 0.005956405773758888, + 0.006037443410605192, + 0.0078033157624304295, + 0.004629708360880613, + 0.005297298543155193, + 0.01335109956562519, + 0.009413696825504303, + 0.00833414401859045, + 0.01150065753608942, + 0.0024457669351249933, + 0.003820941084995866, + 0.0019232574850320816, + 0.003807531204074621, + 0.0028845774941146374, + 0.005210593342781067, + 0.00482479203492403, + 0.0057017989456653595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.06204469874501228 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606203243136406 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606203243136406, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606203243136406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1590883582830429, + 0.0941043421626091, + 0.098558209836483, + 0.03110218606889248, + 0.03458508476614952, + 0.06170839071273804, + 0.12808699905872345, + 0.22657012939453125, + 0.08677509427070618, + 0.07942121475934982 + ], + "credit_map": [ + 0.01666964590549469, + 0.021445784717798233, + 0.014269107021391392, + 0.015214387327432632, + 0.016397099941968918, + 0.021161524578928947, + 0.026942266151309013, + 0.026988545432686806, + 0.009329277090728283, + 0.021428795531392097, + 0.004932680167257786, + 0.007565007079392672, + 0.012694770470261574, + 0.007983455434441566, + 0.01758386567234993, + 0.012586499564349651, + 0.00833433959633112, + 0.015068456530570984, + 0.0070543233305215836, + 0.009775360114872456, + 0.018241992220282555, + 0.015569365583360195, + 0.009534095413982868, + 0.014980271458625793, + 0.0030825238209217787, + 0.005076992325484753, + 0.0023439438082277775, + 0.004804471507668495, + 0.005512801930308342, + 0.003988143987953663, + 0.0033872139174491167, + 0.0029060959350317717, + 0.0033842732664197683, + 0.005848696455359459, + 0.0024268035776913166, + 0.0042987409979105, + 0.004903009627014399, + 0.004968065302819014, + 0.0039883870631456375, + 0.004767108242958784, + 0.00458846241235733, + 0.01224242988973856, + 0.003886380698531866, + 0.006075720302760601, + 0.012293960899114609, + 0.00893232598900795, + 0.007141848560422659, + 0.006547261029481888, + 0.008751201443374157, + 0.015504911541938782, + 0.006470119580626488, + 0.006115999538451433, + 0.007524251472204924, + 0.029127832502126694, + 0.010843152180314064, + 0.043749529868364334, + 0.01844303496181965, + 0.03904804587364197, + 0.01580704376101494, + 0.03209145739674568, + 0.0335351824760437, + 0.03297209367156029, + 0.02658764086663723, + 0.028085634112358093, + 0.010684873908758163, + 0.009607557207345963, + 0.005999329034239054, + 0.007557339034974575, + 0.011249929666519165, + 0.012946409173309803, + 0.011519418098032475, + 0.017210233956575394, + 0.006362235639244318, + 0.00822386797517538, + 0.0060304077342152596, + 0.010193162597715855, + 0.006555146537721157, + 0.012410246767103672, + 0.013660822995007038, + 0.01598532125353813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01606203243136406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot at eye level, the camera moves smoothly to the right, tracking a newspaper as it glides almost weightlessly across a cityscape. The paper, a slightly yellowed broadsheet with frayed edges, billows and twirls in the breeze, its fibrous texture visible as it momentarily catches the light, revealing a bold black headline and a faint impression of a faded monochrome photograph. Its graceful journey is emphasized by a natural motion blur that softens its edges and background, focusing the viewer's attention on its fluid, dance-like movement as it undulates. The scene is set on a quiet city street corner during golden hour, where warm, low-angle sunlight flows between brick buildings and casts long, dramatic shadows across the wet asphalt. The predominant color palette is warm and cinematic, with the golden light illuminating the dust in the air and making the newspaper the brightest element in the frame, creating an overall atmosphere of quiet, poetic solitude that captures a fleeting moment of unexpected beauty.", + "seed": 14647, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14647_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": -0.10910479240119457, + "vlm_logprob": -0.10910479240119457, + "vlm_yes_no_margin": 3.374999988079071, + "vlm_token_credit": -0.10910479240119457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12864771485328674 + ], + "opposite_logprob_tokens": [ + -2.378647804260254 + ], + "logprob_reward": -0.12864771485328674, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12864771485328674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0419047474861145, + 0.08431128412485123, + 0.05190502852201462, + 0.035150885581970215, + 0.1587371975183487, + 0.0494941882789135, + 0.11083775758743286, + 0.43120598793029785, + 0.023590466007590294, + 0.012862492352724075 + ], + "credit_map": [ + 0.003433257108554244, + 0.006613192148506641, + 0.004009891301393509, + 0.004151246976107359, + 0.004818638786673546, + 0.006751452572643757, + 0.006507819052785635, + 0.005619252100586891, + 0.00409590033814311, + 0.010317192412912846, + 0.005683218594640493, + 0.005416924133896828, + 0.018747886642813683, + 0.008097773417830467, + 0.020227549597620964, + 0.011724835261702538, + 0.004032961558550596, + 0.0014787359395995736, + 0.002471530344337225, + 0.012448630295693874, + 0.003693021135404706, + 0.0015599941834807396, + 0.020932378247380257, + 0.005287775304168463, + 0.004097534343600273, + 0.004854349885135889, + 0.002004483947530389, + 0.004484127275645733, + 0.005043041426688433, + 0.0018668007105588913, + 0.004128065891563892, + 0.008672480471432209, + 0.014878661371767521, + 0.01804243214428425, + 0.008147403597831726, + 0.0463712178170681, + 0.019783923402428627, + 0.009579222649335861, + 0.0291779525578022, + 0.012756370939314365, + 0.004246498458087444, + 0.010704809799790382, + 0.003141174092888832, + 0.005851127207279205, + 0.007138053420931101, + 0.0059513370506465435, + 0.006771809887140989, + 0.005689376965165138, + 0.018777217715978622, + 0.014998028986155987, + 0.005248481873422861, + 0.011627058498561382, + 0.006374632474035025, + 0.00794018805027008, + 0.035752322524785995, + 0.010119817219674587, + 0.03512102738022804, + 0.021442575380206108, + 0.008064065128564835, + 0.03643522784113884, + 0.11603643000125885, + 0.0200972743332386, + 0.15603260695934296, + 0.03797672316431999, + 0.0027166474610567093, + 0.0038081191014498472, + 0.0014605894684791565, + 0.004299489315599203, + 0.0035753250122070312, + 0.002302193082869053, + 0.0030019800178706646, + 0.0024261188227683306, + 0.0019889601971954107, + 0.0011282082414254546, + 0.001209464157000184, + 0.0017749915132299066, + 0.0009515307028777897, + 0.0021147034130990505, + 0.002366399159654975, + 0.0013282339787110686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.12864771485328674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705116510391235 + ], + "opposite_logprob_tokens": [ + -1.1370511054992676 + ], + "logprob_reward": -0.38705116510391235, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38705116510391235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04117581248283386, + 0.04023568704724312, + 0.14494891464710236, + 0.03949715569615364, + 0.03972403705120087, + 0.032138142734766006, + 0.05862674117088318, + 0.5301340222358704, + 0.03570465371012688, + 0.03781482204794884 + ], + "credit_map": [ + 0.0027484146412461996, + 0.006005279254168272, + 0.0037400927394628525, + 0.0029141693376004696, + 0.004591450560837984, + 0.008580555208027363, + 0.00738329766318202, + 0.005212555639445782, + 0.003063992829993367, + 0.004123983439058065, + 0.003917026333510876, + 0.0026404557283967733, + 0.00723903626203537, + 0.0037084855139255524, + 0.008581830188632011, + 0.006960876286029816, + 0.010824320837855339, + 0.004232722334563732, + 0.004685004707425833, + 0.030847473070025444, + 0.009938233532011509, + 0.0037859748117625713, + 0.06152728572487831, + 0.01910790614783764, + 0.003717856714501977, + 0.00491748470813036, + 0.0019608191214501858, + 0.005959809757769108, + 0.008451822213828564, + 0.0024995028506964445, + 0.004756658338010311, + 0.007233200594782829, + 0.0038022000808268785, + 0.004244827665388584, + 0.003041590563952923, + 0.004892487078905106, + 0.006608748342841864, + 0.00231162435375154, + 0.009711600840091705, + 0.005110957659780979, + 0.004269008059054613, + 0.004672154318541288, + 0.0023973886854946613, + 0.005048519466072321, + 0.003286061342805624, + 0.002675962634384632, + 0.0055627706460654736, + 0.004226275719702244, + 0.006914958823472261, + 0.006344886496663094, + 0.003385246265679598, + 0.0040650744922459126, + 0.006266758311539888, + 0.003963491413742304, + 0.01770402118563652, + 0.009982306510210037, + 0.042563844472169876, + 0.02225366048514843, + 0.008219574578106403, + 0.047278597950935364, + 0.13543973863124847, + 0.02262130193412304, + 0.20368404686450958, + 0.048073235899209976, + 0.0035425524692982435, + 0.0049933986738324165, + 0.0022552849259227514, + 0.008745312690734863, + 0.002737169386819005, + 0.0016610274324193597, + 0.006522437557578087, + 0.005247469060122967, + 0.004464621189981699, + 0.003265827428549528, + 0.007498604711145163, + 0.004473051521927118, + 0.0030564027838408947, + 0.004436451941728592, + 0.006331390701234341, + 0.004288471303880215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.38705116510391235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551127932965755 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551127932965755, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551127932965755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0709087997674942, + 0.052530862390995026, + 0.0712311714887619, + 0.04904356971383095, + 0.20316694676876068, + 0.057604625821113586, + 0.028852662071585655, + 0.40107741951942444, + 0.03729252517223358, + 0.02829142101109028 + ], + "credit_map": [ + 0.004028200171887875, + 0.009054436348378658, + 0.0039917281828820705, + 0.004017333034425974, + 0.004800379276275635, + 0.01585943065583706, + 0.012126821093261242, + 0.017030466347932816, + 0.003662414848804474, + 0.006094355136156082, + 0.003826995613053441, + 0.004024269059300423, + 0.010079252533614635, + 0.006157979369163513, + 0.012309430167078972, + 0.006376163102686405, + 0.004859515931457281, + 0.0029588129837065935, + 0.0025735911913216114, + 0.007328926585614681, + 0.007492298725992441, + 0.002480837982147932, + 0.02412661910057068, + 0.01941056363284588, + 0.005078163463622332, + 0.005844468250870705, + 0.0024924040772020817, + 0.007137764245271683, + 0.009641948156058788, + 0.002563111251220107, + 0.006290930788964033, + 0.009994777850806713, + 0.018953124061226845, + 0.023596150800585747, + 0.010730724781751633, + 0.05114245414733887, + 0.02834297902882099, + 0.012658311054110527, + 0.040391117334365845, + 0.01735207997262478, + 0.007495904341340065, + 0.010441158898174763, + 0.0033989090006798506, + 0.005842448677867651, + 0.00818527489900589, + 0.0043669831939041615, + 0.009670938365161419, + 0.008203002624213696, + 0.0026740976609289646, + 0.0040607512928545475, + 0.0015745065174996853, + 0.002870689844712615, + 0.0017846429254859686, + 0.0019362710881978273, + 0.00973780918866396, + 0.004213892389088869, + 0.032117221504449844, + 0.016116047278046608, + 0.006435174960643053, + 0.036113761365413666, + 0.10236109048128128, + 0.017165374010801315, + 0.15406189858913422, + 0.036706820130348206, + 0.0032779111061245203, + 0.00440820911899209, + 0.0025921736378222704, + 0.006602215580642223, + 0.005044454243034124, + 0.003752652322873473, + 0.007343485485762358, + 0.004271420184522867, + 0.004205969162285328, + 0.00213799811899662, + 0.0028057079762220383, + 0.003041371703147888, + 0.0022135269828140736, + 0.0025562280789017677, + 0.006908874958753586, + 0.004421743098646402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.012551127932965755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640049956738949 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008640049956738949, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640049956738949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12783248722553253, + 0.08198557794094086, + 0.10886461287736893, + 0.04010738059878349, + 0.45307523012161255, + 0.028696130961179733, + 0.02968134731054306, + 0.034861285239458084, + 0.029673289507627487, + 0.06522267311811447 + ], + "credit_map": [ + 0.0073711504228413105, + 0.017676740884780884, + 0.008827573619782925, + 0.007269434165209532, + 0.011114613153040409, + 0.027528177946805954, + 0.015851296484470367, + 0.03219349682331085, + 0.006320883985608816, + 0.009379000402987003, + 0.0082408981397748, + 0.007700970396399498, + 0.01251987274736166, + 0.010784797370433807, + 0.017411354929208755, + 0.009627796709537506, + 0.011650675907731056, + 0.005713305436074734, + 0.006982837338000536, + 0.028805039823055267, + 0.010958435013890266, + 0.005330509040504694, + 0.012492560781538486, + 0.026931243017315865, + 0.008457856252789497, + 0.00668370584025979, + 0.0038861604407429695, + 0.005353397689759731, + 0.00471610389649868, + 0.0033483842853456736, + 0.003952937666326761, + 0.003708834759891033, + 0.045552413910627365, + 0.05948987975716591, + 0.02546437829732895, + 0.1194368377327919, + 0.059255894273519516, + 0.029389603063464165, + 0.07974635064601898, + 0.03473985940217972, + 0.00383850303478539, + 0.004833455663174391, + 0.0016136830672621727, + 0.002905568340793252, + 0.004381177015602589, + 0.001900108065456152, + 0.00573475519195199, + 0.0034888796508312225, + 0.005214085336774588, + 0.004115039948374033, + 0.0018003538716584444, + 0.004196240566670895, + 0.0017851692391559482, + 0.002236491534858942, + 0.006049447227269411, + 0.004284516908228397, + 0.005364907905459404, + 0.0040691932663321495, + 0.0018713060999289155, + 0.004542167764157057, + 0.002602634485810995, + 0.0019264492439106107, + 0.010571547783911228, + 0.003913079388439655, + 0.003968472126871347, + 0.0032434207387268543, + 0.0026803682558238506, + 0.004610550589859486, + 0.003836889285594225, + 0.0042931921780109406, + 0.0034004119224846363, + 0.0036399844102561474, + 0.0067531331442296505, + 0.005008911248296499, + 0.006385708227753639, + 0.008086263202130795, + 0.0074971807189285755, + 0.005609764717519283, + 0.016735730692744255, + 0.009145982563495636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.008640049956738949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633904159069061 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.008633904159069061, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633904159069061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08957677334547043, + 0.046056196093559265, + 0.10997360199689865, + 0.03792572021484375, + 0.11690501123666763, + 0.12494442611932755, + 0.04963652044534683, + 0.3527202308177948, + 0.042331263422966, + 0.029930245131254196 + ], + "credit_map": [ + 0.006989837158471346, + 0.012739782221615314, + 0.006868244148790836, + 0.00634658383205533, + 0.006383643019944429, + 0.021452123299241066, + 0.015322326682507992, + 0.013474234379827976, + 0.004135466180741787, + 0.004495060537010431, + 0.0048437900841236115, + 0.003727186704054475, + 0.007771056145429611, + 0.0059759472496807575, + 0.007959461770951748, + 0.007148224860429764, + 0.008191671222448349, + 0.00338596454821527, + 0.0033559566363692284, + 0.01913275010883808, + 0.006921649910509586, + 0.0031773098744452, + 0.04724457487463951, + 0.01856372132897377, + 0.003461625659838319, + 0.0046306815929710865, + 0.0020745820365846157, + 0.005261786747723818, + 0.007722139824181795, + 0.0031391733791679144, + 0.0050508081912994385, + 0.006584927439689636, + 0.010764505714178085, + 0.014599206857383251, + 0.006264954339712858, + 0.02146006003022194, + 0.023960663005709648, + 0.008089205250144005, + 0.01753031462430954, + 0.014236102811992168, + 0.011798509396612644, + 0.015715742483735085, + 0.007400793489068747, + 0.01253563817590475, + 0.01841917261481285, + 0.009483851492404938, + 0.02998017519712448, + 0.019610540941357613, + 0.008131466805934906, + 0.0071567208506166935, + 0.0032107059378176928, + 0.00812661275267601, + 0.0034811063669621944, + 0.0033960321452468634, + 0.009034212678670883, + 0.007099667564034462, + 0.02840239927172661, + 0.014051136560738087, + 0.005484732333570719, + 0.0321173295378685, + 0.09003159403800964, + 0.01528453454375267, + 0.13540473580360413, + 0.03194373473525047, + 0.004699042532593012, + 0.004415949806571007, + 0.0027658359613269567, + 0.010794911533594131, + 0.00287329638376832, + 0.0022116664331406355, + 0.009002032689750195, + 0.00556853087618947, + 0.0036659985780715942, + 0.002867973642423749, + 0.0035409019328653812, + 0.005134901963174343, + 0.002272140234708786, + 0.0030374859925359488, + 0.005885695572942495, + 0.0035251497756689787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.008633904159069061 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a medium shot at eye level, the camera moves smoothly to the right, tracking a newspaper as it glides almost weightlessly across a cityscape. The paper, a slightly yellowed broadsheet with frayed edges, billows and twirls in the breeze, its fibrous texture visible as it momentarily catches the light, revealing a bold black headline and a faint impression of a faded monochrome photograph. Its graceful journey is emphasized by a natural motion blur that softens its edges and background, focusing the viewer's attention on its fluid, dance-like movement as it undulates. The scene is set on a quiet city street corner during golden hour, where warm, low-angle sunlight flows between brick buildings and casts long, dramatic shadows across the wet asphalt. The predominant color palette is warm and cinematic, with the golden light illuminating the dust in the air and making the newspaper the brightest element in the frame, creating an overall atmosphere of quiet, poetic solitude that captures a fleeting moment of unexpected beauty.", + "seed": 14648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_6/index_8300_seed_14648_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": -0.08755701510235667, + "vlm_logprob": -0.08755701510235667, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.08755701510235667, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35398948192596436 + ], + "opposite_logprob_tokens": [ + -1.3539894819259644 + ], + "logprob_reward": -0.35398948192596436, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.35398948192596436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07045672833919525, + 0.08652184158563614, + 0.08453153818845749, + 0.047323111444711685, + 0.5093297362327576, + 0.023704769089818, + 0.026596779003739357, + 0.03297285735607147, + 0.051154617220163345, + 0.0674079954624176 + ], + "credit_map": [ + 0.008474485017359257, + 0.0075493380427360535, + 0.009956415742635727, + 0.009945648722350597, + 0.008998681791126728, + 0.007068756967782974, + 0.008966555818915367, + 0.009496845304965973, + 0.006761408410966396, + 0.006778632756322622, + 0.014313353225588799, + 0.010734496638178825, + 0.012207184918224812, + 0.011416896246373653, + 0.01367614883929491, + 0.010633719153702259, + 0.007269599474966526, + 0.004301067907363176, + 0.019763881340622902, + 0.0051441011019051075, + 0.01299263071268797, + 0.011852129362523556, + 0.008243221789598465, + 0.014964899979531765, + 0.00329593732021749, + 0.0022548765409737825, + 0.008019288070499897, + 0.0035348564852029085, + 0.0063691772520542145, + 0.006951119285076857, + 0.008720719255506992, + 0.00817713513970375, + 0.03714124858379364, + 0.1221088320016861, + 0.17974470555782318, + 0.019738560542464256, + 0.037210699170827866, + 0.03941665589809418, + 0.015257933177053928, + 0.05871109664440155, + 0.0019839818123728037, + 0.002330861985683441, + 0.003246435197070241, + 0.004117484204471111, + 0.002334390301257372, + 0.00257013994269073, + 0.003127660835161805, + 0.003993813879787922, + 0.001733344979584217, + 0.0027485452592372894, + 0.004382242448627949, + 0.0030911730136722326, + 0.0035813678987324238, + 0.0044039334170520306, + 0.002649859990924597, + 0.004006309434771538, + 0.0031737566459923983, + 0.002959704725071788, + 0.0038089423906058073, + 0.004171077627688646, + 0.0034029162488877773, + 0.00558496592566371, + 0.003524268278852105, + 0.006347223185002804, + 0.003946548793464899, + 0.0035866268444806337, + 0.005891014356166124, + 0.009109815582633018, + 0.008266631513834, + 0.006640967912971973, + 0.006597913336008787, + 0.007115097716450691, + 0.0045845662243664265, + 0.004118305165320635, + 0.012051569297909737, + 0.009537220932543278, + 0.010123522952198982, + 0.00978919304907322, + 0.01115468144416809, + 0.006048926152288914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.35398948192596436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498724430799484 + ], + "opposite_logprob_tokens": [ + -2.9299871921539307 + ], + "logprob_reward": -0.05498724430799484, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498724430799484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14990545809268951, + 0.08178924024105072, + 0.15206386148929596, + 0.03596910089254379, + 0.3439423739910126, + 0.04751674458384514, + 0.02234242483973503, + 0.058332547545433044, + 0.05778355151414871, + 0.05035465955734253 + ], + "credit_map": [ + 0.007936986163258553, + 0.024874206632375717, + 0.009140241891145706, + 0.04566974937915802, + 0.02212025783956051, + 0.01212853379547596, + 0.01594003289937973, + 0.012095450423657894, + 0.0044716233387589455, + 0.010357532650232315, + 0.007962649688124657, + 0.011549551039934158, + 0.012102259323000908, + 0.008760021068155766, + 0.017886746674776077, + 0.008698851801455021, + 0.008519638329744339, + 0.00736597366631031, + 0.05368119850754738, + 0.009519509971141815, + 0.029924461618065834, + 0.0209417175501585, + 0.009631830267608166, + 0.012479528784751892, + 0.0029733041301369667, + 0.0034078164026141167, + 0.00454599829390645, + 0.0030037788674235344, + 0.005438774358481169, + 0.004380438011139631, + 0.007184635382145643, + 0.005034353584051132, + 0.027204599231481552, + 0.09213124215602875, + 0.1132713034749031, + 0.01237860694527626, + 0.027396151795983315, + 0.02610435150563717, + 0.011601600795984268, + 0.03385452926158905, + 0.004299902357161045, + 0.005099795758724213, + 0.004765552468597889, + 0.00483236787840724, + 0.006191592197865248, + 0.007656332105398178, + 0.004552264232188463, + 0.010118932463228703, + 0.0018037629779428244, + 0.0027236791793257, + 0.00273143220692873, + 0.0035628341138362885, + 0.0031256992369890213, + 0.0027002415154129267, + 0.002810835372656584, + 0.0028839395381510258, + 0.007178459316492081, + 0.005807605106383562, + 0.009422438219189644, + 0.00562669150531292, + 0.006646972149610519, + 0.00891545508056879, + 0.007350123021751642, + 0.007384798023849726, + 0.003795488504692912, + 0.0022267133463174105, + 0.00498906709253788, + 0.0064200665801763535, + 0.004949713591486216, + 0.01580534130334854, + 0.00434472830966115, + 0.01525243278592825, + 0.0042953211814165115, + 0.002272315789014101, + 0.010122325271368027, + 0.009765099734067917, + 0.005374446976929903, + 0.006614961661398411, + 0.0071281553246080875, + 0.00478203734382987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.05498724430799484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645722642540932 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645722642540932, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645722642540932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27656498551368713, + 0.03682440146803856, + 0.21907423436641693, + 0.08106192201375961, + 0.11598099768161774, + 0.04449713975191116, + 0.026649517938494682, + 0.0994824692606926, + 0.030018271878361702, + 0.06984607875347137 + ], + "credit_map": [ + 0.013222549110651016, + 0.03737605735659599, + 0.015484933741390705, + 0.0764160305261612, + 0.07333950698375702, + 0.024316441267728806, + 0.019492555409669876, + 0.016916928812861443, + 0.0027555800043046474, + 0.004705109167844057, + 0.004296494647860527, + 0.0065169683657586575, + 0.006077507045120001, + 0.003846612526103854, + 0.005323241930454969, + 0.0033028870820999146, + 0.013402906246483326, + 0.0111327413469553, + 0.06431403756141663, + 0.015819046646356583, + 0.05213649570941925, + 0.02088809199631214, + 0.02149173989892006, + 0.019889170303940773, + 0.006614848040044308, + 0.005680868402123451, + 0.010544921271502972, + 0.007021973840892315, + 0.014038988389074802, + 0.011049131862819195, + 0.012488983571529388, + 0.013622207567095757, + 0.010484063066542149, + 0.027537086978554726, + 0.01699153706431389, + 0.007295905612409115, + 0.016423853114247322, + 0.012091687880456448, + 0.012646233662962914, + 0.012510620057582855, + 0.0029798389878124, + 0.0038440709467977285, + 0.006142794154584408, + 0.008262185379862785, + 0.005670739337801933, + 0.005215170793235302, + 0.006917431019246578, + 0.005464909132570028, + 0.0024503152817487717, + 0.003605228615924716, + 0.003509192494675517, + 0.003281953977420926, + 0.004043201450258493, + 0.002616953570395708, + 0.003654462518170476, + 0.003488209331408143, + 0.007052794098854065, + 0.00955683272331953, + 0.01079557929188013, + 0.011801257729530334, + 0.010026448406279087, + 0.018456613644957542, + 0.011496486142277718, + 0.02029646560549736, + 0.00252264691516757, + 0.0014915494248270988, + 0.003069370985031128, + 0.004422302357852459, + 0.0034559261985123158, + 0.006178851705044508, + 0.0027658557519316673, + 0.006111766677349806, + 0.005471480544656515, + 0.0038602494169026613, + 0.019279969856142998, + 0.011533462442457676, + 0.006862274836748838, + 0.00878735352307558, + 0.007791424170136452, + 0.006259863264858723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.008645722642540932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539356015622616 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539356015622616, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539356015622616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23302502930164337, + 0.04517899453639984, + 0.07605274766683578, + 0.10040336847305298, + 0.3611171841621399, + 0.016854356974363327, + 0.009558954276144505, + 0.04648255556821823, + 0.04353119432926178, + 0.06779557466506958 + ], + "credit_map": [ + 0.01316784881055355, + 0.03399363160133362, + 0.01326434500515461, + 0.06461663544178009, + 0.043463416397571564, + 0.02409960702061653, + 0.019312690943479538, + 0.02110685408115387, + 0.0038464819081127644, + 0.005053413100540638, + 0.006093144882470369, + 0.005761192180216312, + 0.006035292521119118, + 0.005065449979156256, + 0.007679966278374195, + 0.0056440569460392, + 0.00678997952491045, + 0.005855257157236338, + 0.022563427686691284, + 0.005899016745388508, + 0.011612473987042904, + 0.010950918309390545, + 0.006730067078024149, + 0.0056516132317483425, + 0.008700328879058361, + 0.015505633316934109, + 0.014094059355556965, + 0.01031426340341568, + 0.019277114421129227, + 0.010366801172494888, + 0.01119216438382864, + 0.010953002609312534, + 0.029263736680150032, + 0.09548325091600418, + 0.12265142798423767, + 0.013159372843801975, + 0.026640718802809715, + 0.027908625081181526, + 0.00782602746039629, + 0.0381840355694294, + 0.0015355059877038002, + 0.0017200486036017537, + 0.002665401203557849, + 0.0014401893131434917, + 0.0016482386272400618, + 0.0025667224545031786, + 0.0017637662822380662, + 0.0035144865978509188, + 0.00126702431589365, + 0.0008235111017711461, + 0.0015387211460620165, + 0.0011093274224549532, + 0.0011452847393229604, + 0.0014755841111764312, + 0.0009465236798860133, + 0.0012529792729765177, + 0.0034996827598661184, + 0.005650119390338659, + 0.005536288488656282, + 0.006800551433116198, + 0.005371835082769394, + 0.007409184705466032, + 0.0050884271040558815, + 0.007126469165086746, + 0.0027075668331235647, + 0.0012592568527907133, + 0.004119741264730692, + 0.004655095282942057, + 0.002909056143835187, + 0.013220650143921375, + 0.0021015615202486515, + 0.012558270245790482, + 0.005908589344471693, + 0.003636552719399333, + 0.01388856302946806, + 0.009105191566050053, + 0.007257452700287104, + 0.008615382015705109, + 0.011566045694053173, + 0.007817805744707584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.012539356015622616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a medium shot at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move smoothly to the right while tracking a newspaper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the newspaper a slightly yellowed broadsheet with frayed edges and visible fibrous texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a quiet city street corner during golden hour with warm, low-angle sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the newspaper glide across the wet asphalt while billowing and twirling in the breeze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623270619660616 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623270619660616, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623270619660616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11032021790742874, + 0.026935754343867302, + 0.08805202692747116, + 0.0648966059088707, + 0.24062371253967285, + 0.043976783752441406, + 0.02802235446870327, + 0.12509264051914215, + 0.12894026935100555, + 0.14313960075378418 + ], + "credit_map": [ + 0.00926232524216175, + 0.01304483413696289, + 0.009626999497413635, + 0.018045660108327866, + 0.017000488936901093, + 0.015050857327878475, + 0.01567726582288742, + 0.012611792422831059, + 0.002327862661331892, + 0.0025516406167298555, + 0.0037472734693437815, + 0.0030345330014824867, + 0.004028114955872297, + 0.0036555922124534845, + 0.00469242362305522, + 0.002898313570767641, + 0.006543557625263929, + 0.00540798669680953, + 0.011122399009764194, + 0.008566062897443771, + 0.014721919782459736, + 0.01354834996163845, + 0.014904098585247993, + 0.013237648643553257, + 0.004617821425199509, + 0.0028257230296730995, + 0.008951587602496147, + 0.00510902376845479, + 0.008223353885114193, + 0.010018525645136833, + 0.010951142758131027, + 0.014199423603713512, + 0.014060286805033684, + 0.0489325225353241, + 0.08473223447799683, + 0.010012662969529629, + 0.019187452271580696, + 0.019305327907204628, + 0.012226919643580914, + 0.03216630592942238, + 0.0031156234908849, + 0.003836745163425803, + 0.005973349325358868, + 0.006198730319738388, + 0.0050752717070281506, + 0.005686746910214424, + 0.00543132983148098, + 0.008658988401293755, + 0.003044558921828866, + 0.002655566204339266, + 0.003806276014074683, + 0.0030608479864895344, + 0.004238077439367771, + 0.004505487158894539, + 0.002738261828199029, + 0.003973278682678938, + 0.013381394557654858, + 0.014768845401704311, + 0.015308219008147717, + 0.013899865560233593, + 0.01391550898551941, + 0.0210257638245821, + 0.016808059066534042, + 0.01598498597741127, + 0.005984639283269644, + 0.003698544343933463, + 0.007040699012577534, + 0.008819757960736752, + 0.008407803252339363, + 0.04461856186389923, + 0.007609033025801182, + 0.042761240154504776, + 0.010759011842310429, + 0.007106549106538296, + 0.03915916383266449, + 0.02455037273466587, + 0.011405988596379757, + 0.01847507432103157, + 0.017552515491843224, + 0.014130919240415096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.007623270619660616 + } + ], + "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/0000022/group_7/index_tensor([980, 980])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_tensor([980, 980])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7ae2b2eec620d281d5ad49255d587245b8032e37 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_tensor([980, 980])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up opens on a weathered wooden windowsill, where a single, paper-thin flake of dried, pale blue paint clings precariously to the splintered surface. The camera remains still, capturing the fragile, curled texture of the paint flake, its surface a miniature landscape of fine cracks, before it separates from the aging wood. A gentle draft, carrying dust particles illuminated by a beam of warm afternoon sunlight, disturbs the flake. It sways for a moment, then gently separates. As the flake begins its downward drift, the camera initiates a slow, precise vertical tilt, following the object's descent from the ledge. The flake tumbles and swirls in a slow, unpredictable pattern, its descent determined by air resistance, its pale surface catching and reflecting golden light. The camera tracks its entire journey until it settles with weightless silence onto the dark, dusty floor surface below, creating a tranquil and contemplative atmosphere focused on a quiet moment of decay.", + "seed": 30193, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30193_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": -0.010952620208263398, + "vlm_logprob": -0.010952620208263398, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.010952620208263398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016703277826309204 + ], + "opposite_logprob_tokens": [ + -4.266703128814697 + ], + "logprob_reward": -0.016703277826309204, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016703277826309204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10888244956731796, + 0.05420972779393196, + 0.05736992135643959, + 0.03508598729968071, + 0.04566027224063873, + 0.4677555561065674, + 0.05329253152012825, + 0.025310350582003593, + 0.03476661071181297, + 0.11766664683818817 + ], + "credit_map": [ + 0.010197499766945839, + 0.011259122751653194, + 0.02033049799501896, + 0.014081337489187717, + 0.010493192821741104, + 0.013288868591189384, + 0.009944613091647625, + 0.01928730495274067, + 0.007782783824950457, + 0.008204042911529541, + 0.007659157272428274, + 0.005465308669954538, + 0.007182248868048191, + 0.006291449535638094, + 0.0056053767912089825, + 0.006019354797899723, + 0.008115977980196476, + 0.008016887120902538, + 0.006716496776789427, + 0.00531025230884552, + 0.006984925828874111, + 0.0071939267218112946, + 0.007394567131996155, + 0.007636878173798323, + 0.0037915578577667475, + 0.004698409698903561, + 0.003511116374284029, + 0.004622474312782288, + 0.005221324972808361, + 0.00471719354391098, + 0.004786698147654533, + 0.0037372070364654064, + 0.003996725659817457, + 0.005739332176744938, + 0.0051451693288981915, + 0.005401211325079203, + 0.0046112388372421265, + 0.008804709650576115, + 0.006365572568029165, + 0.0055963085032999516, + 0.03212912380695343, + 0.033147405833005905, + 0.14491458237171173, + 0.06155921891331673, + 0.036477893590927124, + 0.03182283788919449, + 0.044023096561431885, + 0.08368135988712311, + 0.004142924677580595, + 0.006048537325114012, + 0.009012537077069283, + 0.00699000945314765, + 0.004625601228326559, + 0.010733642615377903, + 0.007602595258504152, + 0.0041366764344275, + 0.003011398483067751, + 0.003441987093538046, + 0.0027093407697975636, + 0.0032014173921197653, + 0.002888277405872941, + 0.0037305825389921665, + 0.0031898915767669678, + 0.0031374539248645306, + 0.0032084507402032614, + 0.006105001084506512, + 0.0035192063078284264, + 0.0029818196780979633, + 0.005884950514882803, + 0.0051111094653606415, + 0.004324986599385738, + 0.003631082596257329, + 0.010337541811168194, + 0.014544137753546238, + 0.015616420656442642, + 0.009316162206232548, + 0.017069920897483826, + 0.03102702461183071, + 0.008150109089910984, + 0.011605317704379559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.016703277826309204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962205585092306 + ], + "opposite_logprob_tokens": [ + -5.130962371826172 + ], + "logprob_reward": -0.005962205585092306, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962205585092306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07647751271724701, + 0.10980597883462906, + 0.11174998432397842, + 0.05274337902665138, + 0.046640582382678986, + 0.3169359862804413, + 0.06326626986265182, + 0.03320891410112381, + 0.06221899762749672, + 0.12695243954658508 + ], + "credit_map": [ + 0.007056409027427435, + 0.008399716578423977, + 0.010397211648523808, + 0.009446775540709496, + 0.009961447678506374, + 0.011487525887787342, + 0.007848134264349937, + 0.011880295351147652, + 0.015291507355868816, + 0.016671378165483475, + 0.014952504076063633, + 0.01115958672016859, + 0.015742111951112747, + 0.013266186229884624, + 0.00956655852496624, + 0.013156146742403507, + 0.008639582432806492, + 0.011142374947667122, + 0.010248567909002304, + 0.01581910438835621, + 0.016072632744908333, + 0.015941310673952103, + 0.018043285235762596, + 0.015843121334910393, + 0.005412606988102198, + 0.0066943601705133915, + 0.0050118714570999146, + 0.006650097668170929, + 0.007896301336586475, + 0.007714921608567238, + 0.008151352405548096, + 0.005211871117353439, + 0.005609515123069286, + 0.006777470000088215, + 0.003930883482098579, + 0.004664143081754446, + 0.006478412076830864, + 0.005340903997421265, + 0.008306118659675121, + 0.005533137358725071, + 0.022337356582283974, + 0.024016592651605606, + 0.09485860168933868, + 0.04003486409783363, + 0.02816172130405903, + 0.02235962077975273, + 0.029994888231158257, + 0.055172353982925415, + 0.005519670434296131, + 0.007751269731670618, + 0.01264914683997631, + 0.007467948831617832, + 0.005534903146326542, + 0.010671171359717846, + 0.007972504012286663, + 0.005699655506759882, + 0.003883759258314967, + 0.0038116525392979383, + 0.004492630250751972, + 0.0053359209559857845, + 0.0030093069653958082, + 0.0034477903973311186, + 0.0051645259372889996, + 0.004063325934112072, + 0.0035082201939076185, + 0.007654558401554823, + 0.005510056857019663, + 0.003446009708568454, + 0.009576039388775826, + 0.01771855540573597, + 0.00535973533987999, + 0.009445820935070515, + 0.008177750743925571, + 0.016537975519895554, + 0.012882247567176819, + 0.010211090557277203, + 0.017227156087756157, + 0.04193852096796036, + 0.010677971877157688, + 0.009299736469984055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.005962205585092306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197137862443924 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197137862443924, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197137862443924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06012123450636864, + 0.1143205314874649, + 0.14179235696792603, + 0.05801177769899368, + 0.059050627052783966, + 0.40976250171661377, + 0.04866515100002289, + 0.041039664298295975, + 0.03039337508380413, + 0.036842767149209976 + ], + "credit_map": [ + 0.006781487260013819, + 0.005750279873609543, + 0.007393519859761, + 0.006800364702939987, + 0.007478725165128708, + 0.01046922616660595, + 0.005559591576457024, + 0.009888038970530033, + 0.016922198235988617, + 0.01557307131588459, + 0.01848495751619339, + 0.011976011097431183, + 0.015492783859372139, + 0.011795190162956715, + 0.008906042203307152, + 0.015170274302363396, + 0.015895120799541473, + 0.02028513140976429, + 0.017956459894776344, + 0.016214272007346153, + 0.014800310134887695, + 0.020882872864603996, + 0.019366499036550522, + 0.016391698271036148, + 0.005908196792006493, + 0.006907358765602112, + 0.0078382957726717, + 0.007609885185956955, + 0.00951199047267437, + 0.00765652023255825, + 0.005914448760449886, + 0.006665081717073917, + 0.005184472072869539, + 0.007612661458551884, + 0.005528449080884457, + 0.0065209330059587955, + 0.009474489837884903, + 0.00815941859036684, + 0.009517105296254158, + 0.007053096778690815, + 0.027912722900509834, + 0.026822522282600403, + 0.1314370334148407, + 0.053171683102846146, + 0.03478288650512695, + 0.023282570764422417, + 0.03559083864092827, + 0.07676223665475845, + 0.004715852905064821, + 0.008813458494842052, + 0.005848579574376345, + 0.005302551668137312, + 0.0065086884424090385, + 0.007708893157541752, + 0.006537032779306173, + 0.003230096073821187, + 0.007462671957910061, + 0.005733401980251074, + 0.004322950728237629, + 0.004705897998064756, + 0.004396485630422831, + 0.004760526120662689, + 0.00526484614238143, + 0.004392883740365505, + 0.001907162251882255, + 0.0038622335996478796, + 0.0024426444433629513, + 0.0017250339733436704, + 0.00422520050778985, + 0.008460154756903648, + 0.0020489960443228483, + 0.005721948575228453, + 0.0026138985995203257, + 0.0043953098356723785, + 0.004261592868715525, + 0.004355081357061863, + 0.00470245536416769, + 0.009241022169589996, + 0.0042253718711435795, + 0.0030480357818305492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.018197137862443924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043788228183984756, + 0.024395033717155457, + 0.08612438291311264, + 0.05048718675971031, + 0.07608801126480103, + 0.3948412537574768, + 0.1809670478105545, + 0.032099463045597076, + 0.04486720263957977, + 0.06634220480918884 + ], + "credit_map": [ + 0.00486999424174428, + 0.005157787352800369, + 0.006157223135232925, + 0.005324505269527435, + 0.006422617007046938, + 0.0061877137050032616, + 0.003890091786161065, + 0.005778298247605562, + 0.002994038164615631, + 0.003711058758199215, + 0.00283115659840405, + 0.0026570516638457775, + 0.003906684461981058, + 0.0027670981362462044, + 0.0030811831820756197, + 0.0024467611219733953, + 0.0069819348864257336, + 0.008040381595492363, + 0.0074859983287751675, + 0.012367001734673977, + 0.01386230904608965, + 0.012146104127168655, + 0.013016949407756329, + 0.012223705649375916, + 0.006238649133592844, + 0.007282853592187166, + 0.006107978988438845, + 0.005730113014578819, + 0.007639904040843248, + 0.006010903045535088, + 0.005961532238870859, + 0.0055152494460344315, + 0.0076988437213003635, + 0.010278357192873955, + 0.007238222751766443, + 0.008493788540363312, + 0.011241119354963303, + 0.01327113900333643, + 0.01038405578583479, + 0.007482484448701143, + 0.02856440655887127, + 0.0293524581938982, + 0.1185467392206192, + 0.053094811737537384, + 0.03667837008833885, + 0.02516535110771656, + 0.036133773624897, + 0.06730535626411438, + 0.014765267260372639, + 0.016371458768844604, + 0.041771162301301956, + 0.023823687806725502, + 0.013158082962036133, + 0.02640802226960659, + 0.028047991916537285, + 0.016621381044387817, + 0.003474349156022072, + 0.0030636920128017664, + 0.003921942785382271, + 0.0053648273460567, + 0.003325536148622632, + 0.004079931881278753, + 0.005236255470663309, + 0.0036329294089227915, + 0.0032752156257629395, + 0.006924155633896589, + 0.003150584176182747, + 0.0028300394769757986, + 0.007707496173679829, + 0.010571557097136974, + 0.004373567178845406, + 0.006034585181623697, + 0.006735236383974552, + 0.007230113726109266, + 0.007003610488027334, + 0.008617499843239784, + 0.009443262591958046, + 0.014670305885374546, + 0.007943125441670418, + 0.004699049051851034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258894059807062 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005258894059807062, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258894059807062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047872766852378845, + 0.045419421046972275, + 0.06297092884778976, + 0.07885119318962097, + 0.07688837498426437, + 0.10041598230600357, + 0.302362859249115, + 0.05027352273464203, + 0.09118495881557465, + 0.14375998079776764 + ], + "credit_map": [ + 0.005401067901402712, + 0.004240377806127071, + 0.00941285490989685, + 0.006042772438377142, + 0.005964800249785185, + 0.006160887889564037, + 0.0039232936687767506, + 0.006726710591465235, + 0.00473213754594326, + 0.0063253603875637054, + 0.0058935778215527534, + 0.005609120242297649, + 0.008550834842026234, + 0.004945336375385523, + 0.0038869651034474373, + 0.005476087331771851, + 0.006811968982219696, + 0.006778905168175697, + 0.0068384441547095776, + 0.0071306428872048855, + 0.008024094626307487, + 0.007864629849791527, + 0.009748735465109348, + 0.009773503057658672, + 0.008408976718783379, + 0.008576279506087303, + 0.009109599515795708, + 0.012604644522070885, + 0.009718580171465874, + 0.010242673568427563, + 0.011423742398619652, + 0.008766698651015759, + 0.00655910512432456, + 0.008050534874200821, + 0.011157129891216755, + 0.01059682760387659, + 0.01048280205577612, + 0.010278724133968353, + 0.00981514248996973, + 0.00994810275733471, + 0.008701634593307972, + 0.00753980316221714, + 0.02674749493598938, + 0.012433132156729698, + 0.011083555407822132, + 0.008897410705685616, + 0.009475077502429485, + 0.015537872910499573, + 0.025054926052689552, + 0.028350669890642166, + 0.08467129617929459, + 0.03786173835396767, + 0.017961042001843452, + 0.034357987344264984, + 0.04086156561970711, + 0.033243656158447266, + 0.004061699379235506, + 0.007486905436962843, + 0.006125906948000193, + 0.0072120241820812225, + 0.0067510041408240795, + 0.00568211916834116, + 0.008404092863202095, + 0.004549768753349781, + 0.00830321665853262, + 0.014749333262443542, + 0.008594516664743423, + 0.005225354805588722, + 0.01204258669167757, + 0.019043579697608948, + 0.009005294181406498, + 0.014221071265637875, + 0.01531251985579729, + 0.014357051812112331, + 0.01925557106733322, + 0.022265685722231865, + 0.01882348395884037, + 0.021807285025715828, + 0.015717914327979088, + 0.016220472753047943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.005258894059807062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up opens on a weathered wooden windowsill, where a single, paper-thin flake of dried, pale blue paint clings precariously to the splintered surface. The camera remains still, capturing the fragile, curled texture of the paint flake, its surface a miniature landscape of fine cracks, before it separates from the aging wood. A gentle draft, carrying dust particles illuminated by a beam of warm afternoon sunlight, disturbs the flake. It sways for a moment, then gently separates. As the flake begins its downward drift, the camera initiates a slow, precise vertical tilt, following the object's descent from the ledge. The flake tumbles and swirls in a slow, unpredictable pattern, its descent determined by air resistance, its pale surface catching and reflecting golden light. The camera tracks its entire journey until it settles with weightless silence onto the dark, dusty floor surface below, creating a tranquil and contemplative atmosphere focused on a quiet moment of decay.", + "seed": 30194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30194_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": -0.009120947401970624, + "vlm_logprob": -0.009120947401970624, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.009120947401970624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013902139849960804 + ], + "opposite_logprob_tokens": [ + -4.763902187347412 + ], + "logprob_reward": -0.013902139849960804, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013902139849960804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2555767297744751, + 0.06185712665319443, + 0.11301703006029129, + 0.04833916574716568, + 0.05925094336271286, + 0.07236131280660629, + 0.05383250489830971, + 0.043683990836143494, + 0.06294024735689163, + 0.22914089262485504 + ], + "credit_map": [ + 0.02715970389544964, + 0.01976090855896473, + 0.04951021820306778, + 0.03833942115306854, + 0.021093670278787613, + 0.029346216470003128, + 0.025493796914815903, + 0.04487280920147896, + 0.005148011259734631, + 0.006001546047627926, + 0.012038065120577812, + 0.007838965393602848, + 0.005542382597923279, + 0.006953624077141285, + 0.007823482155799866, + 0.010511047206819057, + 0.008043279871344566, + 0.00672762468457222, + 0.033318448811769485, + 0.03492690250277519, + 0.005366248078644276, + 0.008753779344260693, + 0.00686821062117815, + 0.00901254266500473, + 0.005831559654325247, + 0.004690659698098898, + 0.005875108297914267, + 0.006203501019626856, + 0.006471326109021902, + 0.006835400126874447, + 0.006662109866738319, + 0.005769496783614159, + 0.005262176971882582, + 0.007257436867803335, + 0.010551664978265762, + 0.009008126333355904, + 0.008887249045073986, + 0.0072927456349134445, + 0.006432726979255676, + 0.004558823071420193, + 0.00903609674423933, + 0.008318471722304821, + 0.012360761873424053, + 0.010050640441477299, + 0.008900676853954792, + 0.009466760791838169, + 0.00870792381465435, + 0.005519981496036053, + 0.006091362796723843, + 0.007976104505360126, + 0.007946604862809181, + 0.004517115652561188, + 0.008451088331639767, + 0.0065637980587780476, + 0.008400271646678448, + 0.0038861590437591076, + 0.003113596234470606, + 0.004286415409296751, + 0.004120863974094391, + 0.004090846050530672, + 0.00945369154214859, + 0.006961706560105085, + 0.008012835867702961, + 0.003644032869488001, + 0.006841891910880804, + 0.00892439391463995, + 0.005569683853536844, + 0.006194141693413258, + 0.009026122279465199, + 0.010049923323094845, + 0.01180353295058012, + 0.004530561622232199, + 0.02470356971025467, + 0.03152425214648247, + 0.03136180713772774, + 0.01918090134859085, + 0.02019571326673031, + 0.0327829085290432, + 0.027821065858006477, + 0.04157068952918053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.013902139849960804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00527063338086009 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.00527063338086009, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527063338086009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11408910155296326, + 0.12184707075357437, + 0.10228603333234787, + 0.10580017417669296, + 0.14927536249160767, + 0.10735928267240524, + 0.10062579065561295, + 0.01960931159555912, + 0.05429926887154579, + 0.12480860948562622 + ], + "credit_map": [ + 0.0123878950253129, + 0.00910958182066679, + 0.015487892553210258, + 0.01439289003610611, + 0.022990917786955833, + 0.010466750711202621, + 0.017274994403123856, + 0.011978179216384888, + 0.007873719558119774, + 0.010213078930974007, + 0.022472012788057327, + 0.021180490031838417, + 0.013657324947416782, + 0.01155981793999672, + 0.01689745858311653, + 0.01799316518008709, + 0.007195387501269579, + 0.00830827746540308, + 0.027656054124236107, + 0.026243243366479874, + 0.009409310296177864, + 0.006185551639646292, + 0.00915192998945713, + 0.008136274293065071, + 0.011005522683262825, + 0.012546561658382416, + 0.014451133087277412, + 0.016988877207040787, + 0.01567227579653263, + 0.011231856420636177, + 0.016791608184576035, + 0.007112345192581415, + 0.010435769334435463, + 0.014286027289927006, + 0.03957144170999527, + 0.03203925862908363, + 0.019033780321478844, + 0.012296031229197979, + 0.013590315356850624, + 0.008022739551961422, + 0.010379621759057045, + 0.009324978105723858, + 0.023001600056886673, + 0.01845008134841919, + 0.010792848654091358, + 0.01463413331657648, + 0.01210458017885685, + 0.00867143739014864, + 0.007665598765015602, + 0.011983982287347317, + 0.014333263970911503, + 0.010594521649181843, + 0.017008159309625626, + 0.013363279402256012, + 0.018425902351737022, + 0.007251082919538021, + 0.0018191828858107328, + 0.002119573298841715, + 0.001577409217134118, + 0.0016195481875911355, + 0.004485800862312317, + 0.002825265983119607, + 0.0033362023532390594, + 0.0018263306701555848, + 0.006377002224326134, + 0.007629768922924995, + 0.005203906912356615, + 0.0057430630549788475, + 0.007452362682670355, + 0.008581491187214851, + 0.008509919978678226, + 0.004801756702363491, + 0.015807926654815674, + 0.018025202676653862, + 0.006305006332695484, + 0.008721087127923965, + 0.04080597683787346, + 0.014139476232230663, + 0.01114464271813631, + 0.009859290905296803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.00527063338086009 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01254853792488575 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.01254853792488575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01254853792488575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09913238883018494, + 0.10607772320508957, + 0.07574838399887085, + 0.04009902477264404, + 0.04148588329553604, + 0.5100258588790894, + 0.04248584434390068, + 0.022733038291335106, + 0.033536508679389954, + 0.028675373643636703 + ], + "credit_map": [ + 0.008130568079650402, + 0.00847566407173872, + 0.015721172094345093, + 0.009523820132017136, + 0.010432085022330284, + 0.014174509793519974, + 0.016113659366965294, + 0.01656089723110199, + 0.007413033861666918, + 0.008731136098504066, + 0.0199043620377779, + 0.017723916098475456, + 0.0075987037271261215, + 0.010561461560428143, + 0.013945179060101509, + 0.020199915394186974, + 0.004713411442935467, + 0.004884180147200823, + 0.023849917575716972, + 0.0234779492020607, + 0.0027491827495396137, + 0.005281205754727125, + 0.004642270505428314, + 0.006150259170681238, + 0.004063697066158056, + 0.003454140154644847, + 0.00513342535123229, + 0.006209893152117729, + 0.0038910778239369392, + 0.006964805070310831, + 0.005305706989020109, + 0.005076276138424873, + 0.003345152363181114, + 0.004034373909235001, + 0.009152772836387157, + 0.008671272546052933, + 0.0040465425699949265, + 0.005141600035130978, + 0.003328913589939475, + 0.0037652484606951475, + 0.05034613236784935, + 0.047919027507305145, + 0.12145980447530746, + 0.09177865087985992, + 0.044124241918325424, + 0.061191245913505554, + 0.051408495754003525, + 0.04179822653532028, + 0.004294998943805695, + 0.00583777017891407, + 0.00496193952858448, + 0.004034591838717461, + 0.007408874575048685, + 0.004266920033842325, + 0.0064611430279910564, + 0.005219601560384035, + 0.002040772931650281, + 0.002525113755837083, + 0.0016957706538960338, + 0.0017993263900279999, + 0.006354817654937506, + 0.003295348258689046, + 0.0028546794783324003, + 0.0021672050934284925, + 0.004326087422668934, + 0.0033435996156185865, + 0.003151518991217017, + 0.0034528779797255993, + 0.006134675815701485, + 0.0044141351245343685, + 0.005807170178741217, + 0.0029064398258924484, + 0.0037588842678815126, + 0.005148281343281269, + 0.003392683807760477, + 0.00326335895806551, + 0.002444598590955138, + 0.003308429615572095, + 0.0036093713715672493, + 0.0037497624289244413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.01254853792488575 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10691450536251068, + 0.048996374011039734, + 0.08073053508996964, + 0.051433827728033066, + 0.07868920266628265, + 0.0939982607960701, + 0.09852193295955658, + 0.07152257859706879, + 0.18658776581287384, + 0.1826050877571106 + ], + "credit_map": [ + 0.008841248229146004, + 0.009753961116075516, + 0.016384407877922058, + 0.018770109862089157, + 0.011256380006670952, + 0.011576679535210133, + 0.01671678200364113, + 0.013614935800433159, + 0.004428541753441095, + 0.005203814245760441, + 0.005639762617647648, + 0.0066495356149971485, + 0.007513838354498148, + 0.005330706015229225, + 0.008112339302897453, + 0.006117841694504023, + 0.006132436450570822, + 0.009905772283673286, + 0.018945865333080292, + 0.018099533393979073, + 0.006090058479458094, + 0.008581787347793579, + 0.006208567414432764, + 0.006766516249626875, + 0.0042580366134643555, + 0.003750690957531333, + 0.007686370052397251, + 0.01138152927160263, + 0.0058903475292027, + 0.006241397466510534, + 0.006841801572591066, + 0.005383649840950966, + 0.007239694241434336, + 0.006096035707741976, + 0.016127152368426323, + 0.017527518793940544, + 0.007334643974900246, + 0.009168868884444237, + 0.008613581769168377, + 0.006581701338291168, + 0.012144290842115879, + 0.007830851711332798, + 0.017401574179530144, + 0.013368174433708191, + 0.01239107083529234, + 0.008806096389889717, + 0.013972424902021885, + 0.008083779364824295, + 0.010489520616829395, + 0.015504833310842514, + 0.013581984676420689, + 0.010499244555830956, + 0.014864722266793251, + 0.010520298965275288, + 0.013020524755120277, + 0.010040807537734509, + 0.005283355247229338, + 0.007808875758200884, + 0.006614577956497669, + 0.006635002326220274, + 0.013899754732847214, + 0.012271282263100147, + 0.012690979987382889, + 0.006318746600300074, + 0.022823458537459373, + 0.018929501995444298, + 0.017546582967042923, + 0.01991499960422516, + 0.03556246683001518, + 0.02391641028225422, + 0.0334927998483181, + 0.014401542022824287, + 0.013710454106330872, + 0.029906369745731354, + 0.024848848581314087, + 0.021203292533755302, + 0.01689867116510868, + 0.025258293375372887, + 0.02922634407877922, + 0.021552812308073044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.009783650748431683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099775105714798 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004099775105714798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099775105714798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10664481669664383, + 0.041688114404678345, + 0.03496332839131355, + 0.0448424406349659, + 0.08502205461263657, + 0.13291312754154205, + 0.11873600631952286, + 0.07910217344760895, + 0.11223168671131134, + 0.243856281042099 + ], + "credit_map": [ + 0.01152697391808033, + 0.006945296656340361, + 0.018706589937210083, + 0.01625492237508297, + 0.006473571062088013, + 0.014861415140330791, + 0.012094506062567234, + 0.019781537353992462, + 0.003711662720888853, + 0.0033944989554584026, + 0.00812140479683876, + 0.005437775049358606, + 0.004444981925189495, + 0.004587260540574789, + 0.0051589226350188255, + 0.0068316105753183365, + 0.0035161785781383514, + 0.004630099982023239, + 0.0060639819130301476, + 0.00562743004411459, + 0.00404774397611618, + 0.0038453435990959406, + 0.004032581578940153, + 0.003199970815330744, + 0.004147280473262072, + 0.0039832270704209805, + 0.006042336113750935, + 0.005819849669933319, + 0.004070722032338381, + 0.006692254450172186, + 0.00806587003171444, + 0.006020901259034872, + 0.0061032832600176334, + 0.006457258947193623, + 0.021688589826226234, + 0.021225199103355408, + 0.008055955171585083, + 0.009009972214698792, + 0.006626014597713947, + 0.0058557759039103985, + 0.013684178702533245, + 0.012922454625368118, + 0.03173765540122986, + 0.025227943435311317, + 0.01146958488970995, + 0.014742283150553703, + 0.013349329121410847, + 0.00977969542145729, + 0.010537412017583847, + 0.013283001258969307, + 0.010932897217571735, + 0.010770301334559917, + 0.021360889077186584, + 0.013292876072227955, + 0.021381104364991188, + 0.017177527770400047, + 0.006100401747971773, + 0.008089855313301086, + 0.006464877165853977, + 0.005978653207421303, + 0.023347744718194008, + 0.013982599601149559, + 0.009241322055459023, + 0.005896723363548517, + 0.013097627088427544, + 0.013527086935937405, + 0.01010811235755682, + 0.011805425398051739, + 0.023031923919916153, + 0.013327959924936295, + 0.018434900790452957, + 0.008898649364709854, + 0.015805236995220184, + 0.055861104279756546, + 0.021246984601020813, + 0.015726352110505104, + 0.06487501412630081, + 0.019600210711359978, + 0.02781451679766178, + 0.022926855832338333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.004099775105714798 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up opens on a weathered wooden windowsill, where a single, paper-thin flake of dried, pale blue paint clings precariously to the splintered surface. The camera remains still, capturing the fragile, curled texture of the paint flake, its surface a miniature landscape of fine cracks, before it separates from the aging wood. A gentle draft, carrying dust particles illuminated by a beam of warm afternoon sunlight, disturbs the flake. It sways for a moment, then gently separates. As the flake begins its downward drift, the camera initiates a slow, precise vertical tilt, following the object's descent from the ledge. The flake tumbles and swirls in a slow, unpredictable pattern, its descent determined by air resistance, its pale surface catching and reflecting golden light. The camera tracks its entire journey until it settles with weightless silence onto the dark, dusty floor surface below, creating a tranquil and contemplative atmosphere focused on a quiet moment of decay.", + "seed": 30195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30195_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": -0.00933855187613517, + "vlm_logprob": -0.00933855187613517, + "vlm_yes_no_margin": 5.075, + "vlm_token_credit": -0.00933855187613517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017821071669459343 + ], + "opposite_logprob_tokens": [ + -4.517820835113525 + ], + "logprob_reward": -0.017821071669459343, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017821071669459343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07988516986370087, + 0.036690421402454376, + 0.013686430640518665, + 0.04071645811200142, + 0.040327318012714386, + 0.2689027786254883, + 0.13150447607040405, + 0.2615572214126587, + 0.0735093206167221, + 0.05322037637233734 + ], + "credit_map": [ + 0.005716479849070311, + 0.02024848945438862, + 0.005762738641351461, + 0.0071969483979046345, + 0.013771423138678074, + 0.013012626208364964, + 0.007760053500533104, + 0.006416411604732275, + 0.0019260377157479525, + 0.004622954875230789, + 0.002296709455549717, + 0.003773544682189822, + 0.008684876374900341, + 0.0035748023074120283, + 0.007692948915064335, + 0.004118549171835184, + 0.0007465244852937758, + 0.0021227116230875254, + 0.0008233806001953781, + 0.0013337129494175315, + 0.002941213781014085, + 0.001960471738129854, + 0.002329664770513773, + 0.0014287508092820644, + 0.0029407558031380177, + 0.008422308601439, + 0.0031529595144093037, + 0.003601997159421444, + 0.0036578031722456217, + 0.01006915420293808, + 0.005502606742084026, + 0.003368874080479145, + 0.0027362520340830088, + 0.008489958941936493, + 0.004561048466712236, + 0.004143183119595051, + 0.0030473442748188972, + 0.008374777622520924, + 0.004485171753913164, + 0.004489582497626543, + 0.007143157068639994, + 0.014917118474841118, + 0.01184473279863596, + 0.013162171468138695, + 0.09780304878950119, + 0.09120181947946548, + 0.01947087049484253, + 0.013359852135181427, + 0.007338026538491249, + 0.017820654436945915, + 0.012890668585896492, + 0.013945898972451687, + 0.025479445233941078, + 0.017664579674601555, + 0.022532936185598373, + 0.01383225992321968, + 0.010249003767967224, + 0.06681235879659653, + 0.004980186000466347, + 0.02221783995628357, + 0.038315724581480026, + 0.08178436756134033, + 0.010392693802714348, + 0.026805056259036064, + 0.004054556135088205, + 0.003903714008629322, + 0.009544135071337223, + 0.025180287659168243, + 0.0062057748436927795, + 0.006270324811339378, + 0.005846466403454542, + 0.012504065409302711, + 0.0047787874937057495, + 0.008246629498898983, + 0.006234492640942335, + 0.008668613620102406, + 0.0069305854849517345, + 0.005520247854292393, + 0.00880653876811266, + 0.004034482873976231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.017821071669459343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004113071598112583 + ], + "opposite_logprob_tokens": [ + -5.50411319732666 + ], + "logprob_reward": -0.004113071598112583, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004113071598112583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04961841553449631, + 0.056454405188560486, + 0.04304800182580948, + 0.0847146064043045, + 0.06758327037096024, + 0.1633637398481369, + 0.08836236596107483, + 0.08581137657165527, + 0.2608017325401306, + 0.100242018699646 + ], + "credit_map": [ + 0.0038413831498473883, + 0.00781805906444788, + 0.002802301663905382, + 0.005379176698625088, + 0.011178087443113327, + 0.0063662403263151646, + 0.007194381672888994, + 0.0050387904047966, + 0.0030510001815855503, + 0.007894144393503666, + 0.003615747904404998, + 0.005389897618442774, + 0.014747655019164085, + 0.00728499423712492, + 0.009316515177488327, + 0.005154459271579981, + 0.0023977162782102823, + 0.0066513074561953545, + 0.002395044080913067, + 0.00378177547827363, + 0.008891399018466473, + 0.007429139222949743, + 0.007063498254865408, + 0.004438122734427452, + 0.0035656339023262262, + 0.028329873457551003, + 0.0034666566643863916, + 0.004769885912537575, + 0.006901377346366644, + 0.02511528693139553, + 0.007769118063151836, + 0.004796783439815044, + 0.004331361036747694, + 0.012775189243257046, + 0.006150907836854458, + 0.0065531739965081215, + 0.007265512831509113, + 0.013804770074784756, + 0.00927713792771101, + 0.007425224874168634, + 0.004190005827695131, + 0.009432420134544373, + 0.007742296904325485, + 0.009018700569868088, + 0.05793815106153488, + 0.05346405506134033, + 0.012548120692372322, + 0.0090299928560853, + 0.004561737645417452, + 0.012901168316602707, + 0.010234366171061993, + 0.010460726916790009, + 0.017840834334492683, + 0.0100272037088871, + 0.01285733375698328, + 0.009479000233113766, + 0.003039048984646797, + 0.004802525043487549, + 0.005005890037864447, + 0.015354504808783531, + 0.011844157241284847, + 0.012559305876493454, + 0.015891043469309807, + 0.017314910888671875, + 0.008894547820091248, + 0.01816864125430584, + 0.028170939534902573, + 0.0774211436510086, + 0.03427237644791603, + 0.021489368751645088, + 0.020822899416089058, + 0.0515618659555912, + 0.006570181809365749, + 0.014398721978068352, + 0.009824426844716072, + 0.014738948084414005, + 0.017005274072289467, + 0.010982588864862919, + 0.01803247444331646, + 0.008689412847161293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.004113071598112583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607787050306797 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.01607787050306797, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607787050306797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03382464870810509, + 0.030039189383387566, + 0.0259893499314785, + 0.11069050431251526, + 0.03567110374569893, + 0.13718892633914948, + 0.07090640813112259, + 0.10688523948192596, + 0.287998229265213, + 0.16080637276172638 + ], + "credit_map": [ + 0.002574144396930933, + 0.006480096839368343, + 0.0027855627704411745, + 0.0033576174173504114, + 0.005405138246715069, + 0.005406852811574936, + 0.004006655886769295, + 0.003808580804616213, + 0.0019463230855762959, + 0.006257242988795042, + 0.0020436139311641455, + 0.0025888364762067795, + 0.005311077460646629, + 0.0035846971441060305, + 0.005367959383875132, + 0.0029394396115094423, + 0.001705619739368558, + 0.0036592278629541397, + 0.0019288943149149418, + 0.0028551819268614054, + 0.004489096347242594, + 0.003478684928268194, + 0.004809947684407234, + 0.0030626978259533644, + 0.0063159530982375145, + 0.040113192051649094, + 0.006429673172533512, + 0.006042596884071827, + 0.005898274481296539, + 0.034434977918863297, + 0.005554444156587124, + 0.0059013995341956615, + 0.0029389820992946625, + 0.006565062329173088, + 0.004412671085447073, + 0.0035141808912158012, + 0.003548289416357875, + 0.006215008907020092, + 0.004758534021675587, + 0.003718372667208314, + 0.0031851299572736025, + 0.008440842851996422, + 0.0061432067304849625, + 0.007673005107790232, + 0.04580966383218765, + 0.04666420444846153, + 0.011805377900600433, + 0.007467498537153006, + 0.003827609121799469, + 0.010634233243763447, + 0.00763546722009778, + 0.00787793193012476, + 0.013951543718576431, + 0.008668283931910992, + 0.010843008756637573, + 0.007468333933502436, + 0.0038911188021302223, + 0.032420918345451355, + 0.0030332955066114664, + 0.007585153449326754, + 0.014083732850849628, + 0.03199397400021553, + 0.006743084639310837, + 0.007133956532925367, + 0.013979746028780937, + 0.025455417111516, + 0.03633687645196915, + 0.06289298832416534, + 0.03605325520038605, + 0.02922303043305874, + 0.037269629538059235, + 0.04678728058934212, + 0.009728490374982357, + 0.03929111361503601, + 0.01608429104089737, + 0.027168013155460358, + 0.024865075945854187, + 0.018653104081749916, + 0.016109969466924667, + 0.008906301110982895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.01607787050306797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04884960129857063, + 0.039679184556007385, + 0.01942441612482071, + 0.04001069441437721, + 0.0557306744158268, + 0.17511816322803497, + 0.12407181411981583, + 0.17177842557430267, + 0.14323359727859497, + 0.18210342526435852 + ], + "credit_map": [ + 0.002813121769577265, + 0.00739253219217062, + 0.0027026026509702206, + 0.005265713203698397, + 0.011584889143705368, + 0.006302591413259506, + 0.008205278776586056, + 0.004582872614264488, + 0.0033670428674668074, + 0.00724511593580246, + 0.0036627030931413174, + 0.0041470034047961235, + 0.006872065365314484, + 0.003942260053008795, + 0.006684803403913975, + 0.0037581881042569876, + 0.0011758761247619987, + 0.002690842840820551, + 0.0014970365446060896, + 0.002115886425599456, + 0.004197217524051666, + 0.0023840980138629675, + 0.0032790633849799633, + 0.0020843965467065573, + 0.0016922190552577376, + 0.010863298550248146, + 0.0025555910542607307, + 0.0025124673265963793, + 0.0036089494824409485, + 0.010977205820381641, + 0.00497809424996376, + 0.002822869224473834, + 0.004406488034874201, + 0.010665863752365112, + 0.0037731339689344168, + 0.00581184821203351, + 0.006080534774810076, + 0.00953290518373251, + 0.009075691923499107, + 0.0063842060044407845, + 0.004802478477358818, + 0.010068343952298164, + 0.008958791382610798, + 0.00986363086849451, + 0.05878159776329994, + 0.05845224857330322, + 0.014604934491217136, + 0.009586142376065254, + 0.006782740820199251, + 0.019532376900315285, + 0.011889391578733921, + 0.013024171814322472, + 0.024705179035663605, + 0.012573258951306343, + 0.021966557949781418, + 0.013598132878541946, + 0.006893228739500046, + 0.056384287774562836, + 0.005021339748054743, + 0.01249648816883564, + 0.019363107159733772, + 0.03971029072999954, + 0.019070519134402275, + 0.012839174829423428, + 0.008543932810425758, + 0.01115199364721775, + 0.017539510503411293, + 0.03445272147655487, + 0.023043077439069748, + 0.017674565315246582, + 0.017223455011844635, + 0.013604342006146908, + 0.008849444799125195, + 0.03347375616431236, + 0.018087415024638176, + 0.021281149238348007, + 0.03715120628476143, + 0.018573012202978134, + 0.03408927470445633, + 0.01059817522764206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.006736426148563623 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019443194614723325 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019443194614723325, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019443194614723325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04529033973813057, + 0.03302251175045967, + 0.016438104212284088, + 0.06290385127067566, + 0.06685131788253784, + 0.1032203957438469, + 0.06665477156639099, + 0.22488076984882355, + 0.2544056177139282, + 0.1263323575258255 + ], + "credit_map": [ + 0.0026973444037139416, + 0.00935972947627306, + 0.0028087173122912645, + 0.004703991115093231, + 0.006679257843643427, + 0.0066351876594126225, + 0.007992864586412907, + 0.004413245245814323, + 0.0027128171641379595, + 0.006351925898343325, + 0.0031566584948450327, + 0.003520327154546976, + 0.0050843399949371815, + 0.0035117287188768387, + 0.005097122862935066, + 0.0035875902976840734, + 0.0011982492869719863, + 0.0032195502426475286, + 0.0011446718126535416, + 0.0013339859433472157, + 0.002540149725973606, + 0.0024380795657634735, + 0.002711737994104624, + 0.001851679407991469, + 0.0035095466300845146, + 0.009975424967706203, + 0.003305742284283042, + 0.006166654638946056, + 0.010247278027236462, + 0.011765372939407825, + 0.011392662301659584, + 0.006541166454553604, + 0.003873895388096571, + 0.014209169894456863, + 0.00718712667003274, + 0.007061600685119629, + 0.00598802138119936, + 0.014746633358299732, + 0.007594993803650141, + 0.0061898790299892426, + 0.0033710526768118143, + 0.007797921076416969, + 0.0038483471143990755, + 0.0062841204926371574, + 0.03737800568342209, + 0.030542587861418724, + 0.00842223409563303, + 0.0055761318653821945, + 0.00309390970505774, + 0.012295166961848736, + 0.005795716308057308, + 0.007184080313891172, + 0.014739973470568657, + 0.006121636368334293, + 0.010798508301377296, + 0.0066257803700864315, + 0.008195084519684315, + 0.01987297087907791, + 0.008144119754433632, + 0.030839473009109497, + 0.036003950983285904, + 0.055045537650585175, + 0.023782769218087196, + 0.04299687221646309, + 0.008433357812464237, + 0.020724698901176453, + 0.02568841353058815, + 0.06030857935547829, + 0.03370652347803116, + 0.02705489844083786, + 0.02847583219408989, + 0.05001332610845566, + 0.007662028539925814, + 0.01669640652835369, + 0.016444921493530273, + 0.01308187935501337, + 0.01597808301448822, + 0.0152324503287673, + 0.02870539017021656, + 0.012531207874417305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0019443194614723325 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up opens on a weathered wooden windowsill, where a single, paper-thin flake of dried, pale blue paint clings precariously to the splintered surface. The camera remains still, capturing the fragile, curled texture of the paint flake, its surface a miniature landscape of fine cracks, before it separates from the aging wood. A gentle draft, carrying dust particles illuminated by a beam of warm afternoon sunlight, disturbs the flake. It sways for a moment, then gently separates. As the flake begins its downward drift, the camera initiates a slow, precise vertical tilt, following the object's descent from the ledge. The flake tumbles and swirls in a slow, unpredictable pattern, its descent determined by air resistance, its pale surface catching and reflecting golden light. The camera tracks its entire journey until it settles with weightless silence onto the dark, dusty floor surface below, creating a tranquil and contemplative atmosphere focused on a quiet moment of decay.", + "seed": 30196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_7/index_980_seed_30196_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": -0.03446094812825322, + "vlm_logprob": -0.03446094812825322, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03446094812825322, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05065440386533737 + ], + "opposite_logprob_tokens": [ + -3.175654411315918 + ], + "logprob_reward": -0.05065440386533737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05065440386533737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17877094447612762, + 0.12452054768800735, + 0.029129749163985252, + 0.043360479176044464, + 0.12116577476263046, + 0.09976273775100708, + 0.11860653758049011, + 0.10426357388496399, + 0.06263205409049988, + 0.11778759956359863 + ], + "credit_map": [ + 0.016829464584589005, + 0.02230052836239338, + 0.021577034145593643, + 0.018833013251423836, + 0.02034905180335045, + 0.0262166578322649, + 0.02812388353049755, + 0.02454131469130516, + 0.01695432886481285, + 0.013619945384562016, + 0.00791126862168312, + 0.007843234576284885, + 0.03199422359466553, + 0.01770128309726715, + 0.010949631221592426, + 0.017546644434332848, + 0.0032914786133915186, + 0.0038103011902421713, + 0.0023470937740057707, + 0.004348542541265488, + 0.004752061329782009, + 0.003653113730251789, + 0.004061825107783079, + 0.0028653324116021395, + 0.004377260338515043, + 0.003991992212831974, + 0.003297962946817279, + 0.007093765307217836, + 0.007751864846795797, + 0.0059220087714493275, + 0.005289187654852867, + 0.005636436864733696, + 0.014501236379146576, + 0.021944427862763405, + 0.010099252685904503, + 0.01745193637907505, + 0.01984453946352005, + 0.013413477689027786, + 0.011178084649145603, + 0.012732813134789467, + 0.009596035815775394, + 0.01390144508332014, + 0.008096139878034592, + 0.015822364017367363, + 0.010475176386535168, + 0.012461916543543339, + 0.017363235354423523, + 0.012046423740684986, + 0.013103842735290527, + 0.015734711661934853, + 0.009475072845816612, + 0.010325295850634575, + 0.014226006343960762, + 0.018204016610980034, + 0.01116126962006092, + 0.02637631632387638, + 0.011207729578018188, + 0.01469912938773632, + 0.00783335231244564, + 0.012511979788541794, + 0.01538064330816269, + 0.015832653269171715, + 0.013672880828380585, + 0.01312519982457161, + 0.0069724260829389095, + 0.008397126570343971, + 0.0041313027031719685, + 0.00618240050971508, + 0.008023497648537159, + 0.01093912310898304, + 0.008467289619147778, + 0.009518887847661972, + 0.011121584102511406, + 0.011700121685862541, + 0.008693291805684566, + 0.012126903980970383, + 0.017293719574809074, + 0.01831398345530033, + 0.015197350643575191, + 0.023340655490756035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05065440386533737 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005975714419037104 + ], + "opposite_logprob_tokens": [ + -5.130975723266602 + ], + "logprob_reward": -0.005975714419037104, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005975714419037104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07347079366445541, + 0.11821910738945007, + 0.054775137454271317, + 0.07369959354400635, + 0.08655206859111786, + 0.3580819368362427, + 0.09101555496454239, + 0.0668773204088211, + 0.04406172037124634, + 0.033246781677007675 + ], + "credit_map": [ + 0.005602783989161253, + 0.005997828673571348, + 0.005276309326291084, + 0.00597191508859396, + 0.006536247208714485, + 0.012988312169909477, + 0.011483350768685341, + 0.01961403153836727, + 0.011383074335753918, + 0.01663987897336483, + 0.01018667034804821, + 0.01583823747932911, + 0.013917638920247555, + 0.014450074173510075, + 0.020452484488487244, + 0.015351037494838238, + 0.004640492144972086, + 0.006173747591674328, + 0.0033891203347593546, + 0.005110605154186487, + 0.00840518157929182, + 0.008807760663330555, + 0.006368506699800491, + 0.01187971793115139, + 0.008462004363536835, + 0.006090099457651377, + 0.0050702630542218685, + 0.008600115776062012, + 0.015270086005330086, + 0.011631444096565247, + 0.007165009155869484, + 0.01141055952757597, + 0.00997335184365511, + 0.015534126199781895, + 0.006681826431304216, + 0.014442135579884052, + 0.01293637603521347, + 0.00953806284815073, + 0.008091730065643787, + 0.009354449808597565, + 0.030734160915017128, + 0.03264018893241882, + 0.03221212327480316, + 0.06189895048737526, + 0.028912439942359924, + 0.04508965089917183, + 0.07743676006793976, + 0.04915763437747955, + 0.01077620591968298, + 0.012521195225417614, + 0.007066152058541775, + 0.01088587287813425, + 0.014832017943263054, + 0.011739876121282578, + 0.008492322638630867, + 0.01470190566033125, + 0.006748660933226347, + 0.008662399835884571, + 0.004958907142281532, + 0.007907530292868614, + 0.009439373388886452, + 0.009817885234951973, + 0.007540460675954819, + 0.011802100576460361, + 0.005177951883524656, + 0.005664822645485401, + 0.003738120198249817, + 0.004357676021754742, + 0.00653457548469305, + 0.005658687558025122, + 0.006800119299441576, + 0.006129763089120388, + 0.003789653303101659, + 0.0033703348599374294, + 0.001979396678507328, + 0.0028625871054828167, + 0.0052868700586259365, + 0.00514473719522357, + 0.0043885186314582825, + 0.006424679420888424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.005975714419037104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901213318109512 + ], + "opposite_logprob_tokens": [ + -2.579012155532837 + ], + "logprob_reward": -0.07901213318109512, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901213318109512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10103399306535721, + 0.11072993278503418, + 0.09306953102350235, + 0.08759459853172302, + 0.14703939855098724, + 0.23884449899196625, + 0.06420541554689407, + 0.05710418149828911, + 0.0430738627910614, + 0.05730452015995979 + ], + "credit_map": [ + 0.011544870212674141, + 0.013871868140995502, + 0.010670887306332588, + 0.010796083137392998, + 0.010872878134250641, + 0.01501085702329874, + 0.014012428931891918, + 0.014254136942327023, + 0.013358961790800095, + 0.012762730941176414, + 0.007269959431141615, + 0.009800649248063564, + 0.026705684140324593, + 0.014563250355422497, + 0.012263834476470947, + 0.014004873111844063, + 0.008733612485229969, + 0.009529904462397099, + 0.0049619898200035095, + 0.007416492328047752, + 0.017739055678248405, + 0.015560472384095192, + 0.009554432705044746, + 0.019573578611016273, + 0.00827611144632101, + 0.008426747284829617, + 0.0058031948283314705, + 0.01544435415416956, + 0.012541405856609344, + 0.013098593801259995, + 0.009301351383328438, + 0.014702847227454185, + 0.01885383203625679, + 0.028618188574910164, + 0.01111945603042841, + 0.022114450111985207, + 0.020113639533519745, + 0.016151295974850655, + 0.014270992018282413, + 0.01579754240810871, + 0.03262766823172569, + 0.039780791848897934, + 0.01972552202641964, + 0.02961314469575882, + 0.029468849301338196, + 0.030840512365102768, + 0.021195419132709503, + 0.03559263050556183, + 0.0071328808553516865, + 0.009950386360287666, + 0.004812250379472971, + 0.007649821694940329, + 0.010497642681002617, + 0.008963788859546185, + 0.006306873634457588, + 0.008891778998076916, + 0.005119228269904852, + 0.008267420344054699, + 0.004056872334331274, + 0.006265176460146904, + 0.007658614311367273, + 0.009532527066767216, + 0.006091126706451178, + 0.01011322345584631, + 0.00504108564928174, + 0.003953421488404274, + 0.003483233042061329, + 0.00372122460976243, + 0.005450545810163021, + 0.0074979220516979694, + 0.005226258188486099, + 0.008700179867446423, + 0.006795868743211031, + 0.006059900391846895, + 0.0036557777784764767, + 0.004463464021682739, + 0.008028156124055386, + 0.009130867198109627, + 0.006874477956444025, + 0.012296013534069061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.07901213318109512 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607259176671505 + ], + "opposite_logprob_tokens": [ + -4.141072750091553 + ], + "logprob_reward": -0.01607259176671505, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607259176671505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18522967398166656, + 0.06792964041233063, + 0.023752888664603233, + 0.08820550888776779, + 0.09838906675577164, + 0.35409536957740784, + 0.08943813294172287, + 0.02540600672364235, + 0.05321984365582466, + 0.014333898201584816 + ], + "credit_map": [ + 0.014815455302596092, + 0.020283741876482964, + 0.014522681944072247, + 0.017094580456614494, + 0.019079778343439102, + 0.029467381536960602, + 0.035871073603630066, + 0.03409497067332268, + 0.00797873456031084, + 0.008819782175123692, + 0.005054600071161985, + 0.008577670902013779, + 0.011920166201889515, + 0.008483760058879852, + 0.009557695128023624, + 0.007537230849266052, + 0.0020206859335303307, + 0.0021855595987290144, + 0.0017979443073272705, + 0.002111118985339999, + 0.0038209613412618637, + 0.0038987360894680023, + 0.0029152825009077787, + 0.0050025987438857555, + 0.008529131300747395, + 0.006840829737484455, + 0.005971075966954231, + 0.011690041981637478, + 0.013938036747276783, + 0.012951909564435482, + 0.010596880689263344, + 0.0176876001060009, + 0.010054241865873337, + 0.01297054160386324, + 0.009969096630811691, + 0.018191834911704063, + 0.012042982503771782, + 0.0112860556691885, + 0.011758580803871155, + 0.012115725316107273, + 0.028561320155858994, + 0.029218418523669243, + 0.0331723652780056, + 0.06780029833316803, + 0.03116440214216709, + 0.04335111752152443, + 0.07446157932281494, + 0.04636584594845772, + 0.00977132935076952, + 0.013401485048234463, + 0.010482036508619785, + 0.012067507021129131, + 0.012979350052773952, + 0.009823218919336796, + 0.01035962626338005, + 0.010553576983511448, + 0.002728153020143509, + 0.0037235564086586237, + 0.001830469467677176, + 0.003030634019523859, + 0.0033108515199273825, + 0.004222608171403408, + 0.0026177174877375364, + 0.003942017909139395, + 0.004709802567958832, + 0.004790517967194319, + 0.003862827317789197, + 0.004613928031176329, + 0.006323426030576229, + 0.009100714698433876, + 0.006490867119282484, + 0.013327759690582752, + 0.0016023090574890375, + 0.0013279950944706798, + 0.0009205921669490635, + 0.0011442643590271473, + 0.002212916500866413, + 0.0021591808181256056, + 0.0015714546898379922, + 0.0033951853401958942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.01607259176671505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 an extreme close-up of a weathered wooden windowsill?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a single, paper-thin flake of dried, pale blue paint visible clinging to the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dust particles illuminated by a beam of warm afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera initiate a slow, precise vertical tilt to follow the paint flake's descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the paint flake tumble and swirl in a slow, unpredictable pattern before settling on the dark, dusty floor?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058989740908146 + ], + "opposite_logprob_tokens": [ + -3.895589828491211 + ], + "logprob_reward": -0.02058989740908146, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058989740908146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047968532890081406, + 0.04652496799826622, + 0.034135688096284866, + 0.06931094080209732, + 0.06618411839008331, + 0.4951735734939575, + 0.11151144653558731, + 0.07164470851421356, + 0.03556319326162338, + 0.02198277786374092 + ], + "credit_map": [ + 0.004319579806178808, + 0.005802573636174202, + 0.0038450686261057854, + 0.004355451557785273, + 0.0042614261619746685, + 0.007640706840902567, + 0.008196743205189705, + 0.009546990506350994, + 0.004485464189201593, + 0.004919401369988918, + 0.002764597302302718, + 0.0071787447668612, + 0.005203482694923878, + 0.0062895482406020164, + 0.007766292430460453, + 0.007917438633739948, + 0.004326911643147469, + 0.0031733140349388123, + 0.0027146439533680677, + 0.003103505354374647, + 0.008329312317073345, + 0.00405711866915226, + 0.0034053726121783257, + 0.00502551207318902, + 0.007101239636540413, + 0.005675268825143576, + 0.004815222229808569, + 0.010450558736920357, + 0.012096249498426914, + 0.00900267157703638, + 0.010433940216898918, + 0.009735792875289917, + 0.006231673993170261, + 0.0089729530736804, + 0.00563195301219821, + 0.011900622397661209, + 0.008270283229649067, + 0.008639628067612648, + 0.008750750683248043, + 0.00778626324608922, + 0.04202055558562279, + 0.044038981199264526, + 0.04532453417778015, + 0.08951824903488159, + 0.03967265784740448, + 0.06090446561574936, + 0.10789590328931808, + 0.06579825282096863, + 0.011854820884764194, + 0.014492813497781754, + 0.007587776053696871, + 0.011098050512373447, + 0.019778477028012276, + 0.01489246916025877, + 0.013245346955955029, + 0.018561700358986855, + 0.008536229841411114, + 0.010067490860819817, + 0.005784338340163231, + 0.008612734265625477, + 0.011304696090519428, + 0.00963375810533762, + 0.00852419063448906, + 0.009181277826428413, + 0.003200618317350745, + 0.004338376224040985, + 0.0026800069026649, + 0.004298797342926264, + 0.005328664090484381, + 0.005305455066263676, + 0.005344092845916748, + 0.005067185964435339, + 0.002873852849006653, + 0.0027389591559767723, + 0.001495376811362803, + 0.0022639587987214327, + 0.003721807384863496, + 0.0029311676044017076, + 0.0024822347331792116, + 0.003475422505289316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.02058989740908146 + } + ], + "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/0000022/group_8/index_tensor([3707, 3707])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_tensor([3707, 3707])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d66456f296f677b8b6824c655ca60ba4e597423c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_tensor([3707, 3707])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic drone shot begins with a high-angle, sweeping view, capturing the vastness of the valley before slowly and steadily panning forward. Under the soft, golden light of dusk, hundreds of deer converge into a vast torrent, a living, surging river, flowing along the valley floor, frost-stained ochre. This vast herd, their thick winter coats gradating from deep reddish-brown to light tan, moves in a synchronized, rhythmic cadence. The massive antlers of the males rise and fall within the herd, like dead logs floating on a river. As the camera zooms in, the vast collective scene breaks down into crisp individual details: some deer can be seen bounding gracefully over small mounds, while others graze on hay without breaking their rhythm. Their white breath condenses into wisps of mist in the cold, crisp air. The soft setting sun outlines a shimmering golden outline on the edges of the deer's fur and casts long, dynamic shadows behind them. The whole picture is shrouded in a tranquil, majestic and primitive atmosphere full of vitality, showing the awe-inspiring grandeur of nature.", + "seed": 28674, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_3707_seed_28674_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": -0.07069767857901751, + "vlm_logprob": -0.07069767857901751, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.07069767857901751, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024909352883696556 + ], + "opposite_logprob_tokens": [ + -4.024909496307373 + ], + "logprob_reward": -0.024909352883696556, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024909352883696556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1519022136926651, + 0.07343752682209015, + 0.09418860822916031, + 0.21902337670326233, + 0.12796398997306824, + 0.07623488456010818, + 0.09154827892780304, + 0.0659075677394867, + 0.056838225573301315, + 0.04295535013079643 + ], + "credit_map": [ + 0.013622008264064789, + 0.01904771476984024, + 0.023083386942744255, + 0.02033071033656597, + 0.01972491294145584, + 0.01777222752571106, + 0.021573662757873535, + 0.016747578978538513, + 0.0064606680534780025, + 0.010665006935596466, + 0.009630726650357246, + 0.008503357879817486, + 0.011607948690652847, + 0.008971238508820534, + 0.009945289231836796, + 0.0076532852835953236, + 0.008697240613400936, + 0.007908721454441547, + 0.022659381851553917, + 0.008973716758191586, + 0.008535915985703468, + 0.02396010234951973, + 0.0073476554825901985, + 0.0061058709397912025, + 0.028504671528935432, + 0.01854635775089264, + 0.033904965966939926, + 0.01500790100544691, + 0.026731139048933983, + 0.06605339050292969, + 0.014224923215806484, + 0.016050012782216072, + 0.025457315146923065, + 0.025392591953277588, + 0.010913575999438763, + 0.013323558494448662, + 0.01163160428404808, + 0.013740894384682178, + 0.010116091929376125, + 0.0173883568495512, + 0.00873014610260725, + 0.010909083299338818, + 0.009095398709177971, + 0.00826200656592846, + 0.010273919440805912, + 0.013074074871838093, + 0.007173466496169567, + 0.008716784417629242, + 0.011099332012236118, + 0.01285657286643982, + 0.01156790554523468, + 0.009890107437968254, + 0.009981171227991581, + 0.011205107904970646, + 0.011797952465713024, + 0.013150126673281193, + 0.008315850049257278, + 0.008701355196535587, + 0.008106681518256664, + 0.0062216902151703835, + 0.011430102400481701, + 0.011424049735069275, + 0.006456073839217424, + 0.005251756869256496, + 0.004220990464091301, + 0.008013400249183178, + 0.005707289557904005, + 0.0071311211213469505, + 0.00701809162274003, + 0.009162676520645618, + 0.01120769139379263, + 0.004376960918307304, + 0.00442960299551487, + 0.004306608345359564, + 0.009415487758815289, + 0.009931462816894054, + 0.004466686863452196, + 0.003281960031017661, + 0.004767249338328838, + 0.002356288954615593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.024909352883696556 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003737986087799 + ], + "opposite_logprob_tokens": [ + -2.695037364959717 + ], + "logprob_reward": -0.07003737986087799, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003737986087799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09082955121994019, + 0.10111583769321442, + 0.08314045518636703, + 0.4179900288581848, + 0.04965776577591896, + 0.12294168025255203, + 0.02083774283528328, + 0.023662589490413666, + 0.05643618106842041, + 0.03338809683918953 + ], + "credit_map": [ + 0.008082203567028046, + 0.011815249919891357, + 0.016224661841988564, + 0.011005307547748089, + 0.008708241395652294, + 0.00839171838015318, + 0.01703147031366825, + 0.009570707567036152, + 0.008671000599861145, + 0.01532662007957697, + 0.01380582433193922, + 0.010712779127061367, + 0.01608743704855442, + 0.010091437958180904, + 0.01561871450394392, + 0.010802028700709343, + 0.01031623687595129, + 0.008071791380643845, + 0.015044206753373146, + 0.005471372045576572, + 0.0071899439208209515, + 0.027046507224440575, + 0.006125772371888161, + 0.0038746369536966085, + 0.07606056332588196, + 0.04690678045153618, + 0.03504667803645134, + 0.03449026122689247, + 0.05823645368218422, + 0.10009056329727173, + 0.034116022288799286, + 0.03304272145032883, + 0.007481289096176624, + 0.0060977935791015625, + 0.0052316561341285706, + 0.006277446169406176, + 0.005118152592331171, + 0.0075360857881605625, + 0.00489821145310998, + 0.007017134688794613, + 0.014585809782147408, + 0.019368330016732216, + 0.012530063278973103, + 0.015289979986846447, + 0.015716833993792534, + 0.016049746423959732, + 0.012054968625307083, + 0.017345955595374107, + 0.0022295168600976467, + 0.0036057678516954184, + 0.002267194213345647, + 0.0028034590650349855, + 0.0024909924250096083, + 0.002309046220034361, + 0.0025971971917897463, + 0.0025345711037516594, + 0.0028004846535623074, + 0.003730075666680932, + 0.0027626827359199524, + 0.0026736916042864323, + 0.0034321267157793045, + 0.003445818554610014, + 0.003021021606400609, + 0.0017966905143111944, + 0.005844803526997566, + 0.008835804648697376, + 0.0064069898799061775, + 0.007113962899893522, + 0.008301220834255219, + 0.007339152507483959, + 0.008276417851448059, + 0.004317833110690117, + 0.003924891352653503, + 0.003319571493193507, + 0.007509533781558275, + 0.007068133447319269, + 0.0026700575836002827, + 0.0020163198933005333, + 0.004404352977871895, + 0.0024752386379987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.07003737986087799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554735839366913 + ], + "opposite_logprob_tokens": [ + -1.6005473136901855 + ], + "logprob_reward": -0.22554735839366913, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554735839366913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12398514896631241, + 0.0790630578994751, + 0.14966771006584167, + 0.10848430544137955, + 0.1719992309808731, + 0.15833528339862823, + 0.10081586986780167, + 0.03318742290139198, + 0.043545857071876526, + 0.030916091054677963 + ], + "credit_map": [ + 0.013111980631947517, + 0.01850203424692154, + 0.021252701058983803, + 0.014837335795164108, + 0.011447983793914318, + 0.01156365405768156, + 0.013500332832336426, + 0.01976912096142769, + 0.005987341981381178, + 0.010318656452000141, + 0.012033076956868172, + 0.010161074809730053, + 0.01147811021655798, + 0.009892228990793228, + 0.010089120827615261, + 0.0091034434735775, + 0.029052894562482834, + 0.012105464935302734, + 0.015053206123411655, + 0.008879389613866806, + 0.013561142608523369, + 0.052082836627960205, + 0.010191681794822216, + 0.008741105906665325, + 0.023802489042282104, + 0.009909048676490784, + 0.023191867396235466, + 0.009899817407131195, + 0.00983318779617548, + 0.014966394752264023, + 0.007959229871630669, + 0.008922271430492401, + 0.019085928797721863, + 0.041396718472242355, + 0.017441727221012115, + 0.019150303676724434, + 0.016241811215877533, + 0.018179260194301605, + 0.014826729893684387, + 0.025676749646663666, + 0.015726754441857338, + 0.04535544291138649, + 0.011305890046060085, + 0.013678528368473053, + 0.01495992112904787, + 0.017805635929107666, + 0.013631903566420078, + 0.02587120421230793, + 0.010915792547166348, + 0.015867605805397034, + 0.010856048204004765, + 0.009678100235760212, + 0.009808726608753204, + 0.012701638974249363, + 0.015403290279209614, + 0.015584668144583702, + 0.002889145165681839, + 0.0046777366660535336, + 0.006486669648438692, + 0.003293829271569848, + 0.005167263559997082, + 0.005601284559816122, + 0.002775287488475442, + 0.002296207007020712, + 0.0051998235285282135, + 0.005440277047455311, + 0.006106084678322077, + 0.006127055734395981, + 0.0067826369777321815, + 0.004361474886536598, + 0.005622674245387316, + 0.0039058304391801357, + 0.0023310899268835783, + 0.0022974770981818438, + 0.006111550610512495, + 0.007415378000587225, + 0.003586604492738843, + 0.002053431700915098, + 0.0048757437616586685, + 0.0022448129020631313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.22554735839366913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190310439094901 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190310439094901, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190310439094901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11685612052679062, + 0.06220876798033714, + 0.0851091518998146, + 0.11590253561735153, + 0.09242431074380875, + 0.1316719800233841, + 0.07800211757421494, + 0.051155272871255875, + 0.1724434494972229, + 0.09422636777162552 + ], + "credit_map": [ + 0.014216025359928608, + 0.01554061658680439, + 0.019789719954133034, + 0.015362474136054516, + 0.012162663042545319, + 0.011482438072562218, + 0.012097084894776344, + 0.016205091029405594, + 0.0057242377661168575, + 0.012376138009130955, + 0.008406391367316246, + 0.006169445812702179, + 0.010357835330069065, + 0.008177448995411396, + 0.005804016720503569, + 0.005193248391151428, + 0.01977805607020855, + 0.009749599732458591, + 0.005471089389175177, + 0.006514491979032755, + 0.0069288089871406555, + 0.026732610538601875, + 0.0050440700724720955, + 0.004890415817499161, + 0.012837880291044712, + 0.013544119894504547, + 0.014132105745375156, + 0.010167861357331276, + 0.0156655702739954, + 0.028586288914084435, + 0.010232388973236084, + 0.010736302472651005, + 0.022693736478686333, + 0.029197974130511284, + 0.006837835069745779, + 0.007596124429255724, + 0.006686169654130936, + 0.007136133499443531, + 0.005312186665832996, + 0.006964140571653843, + 0.017509428784251213, + 0.04075571894645691, + 0.010964230634272099, + 0.011723674833774567, + 0.013318939134478569, + 0.01200922392308712, + 0.007582792080938816, + 0.017807956784963608, + 0.007703881245106459, + 0.011446441523730755, + 0.009207754395902157, + 0.007506542839109898, + 0.007726206909865141, + 0.007553919218480587, + 0.012267024256289005, + 0.014590339735150337, + 0.007747832685709, + 0.005855221301317215, + 0.009652281180024147, + 0.00769155565649271, + 0.004369615111500025, + 0.005110696889460087, + 0.006101484876126051, + 0.004626581445336342, + 0.015185683965682983, + 0.02667192742228508, + 0.015016001649200916, + 0.018393151462078094, + 0.01846250705420971, + 0.035060200840234756, + 0.02975490875542164, + 0.013899056240916252, + 0.00876564346253872, + 0.008075492456555367, + 0.01340928953140974, + 0.017669133841991425, + 0.011000867933034897, + 0.008129986003041267, + 0.01821736805140972, + 0.00895857997238636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.003190310439094901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803991317749023 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.029803991317749023, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803991317749023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1251872181892395, + 0.06805305182933807, + 0.08788701146841049, + 0.1356266438961029, + 0.0683971494436264, + 0.08245941251516342, + 0.12513764202594757, + 0.17186680436134338, + 0.07479174435138702, + 0.06059331074357033 + ], + "credit_map": [ + 0.010937773622572422, + 0.01879514940083027, + 0.01721745729446411, + 0.018431494012475014, + 0.015691984444856644, + 0.014828533865511417, + 0.01412068959325552, + 0.015164132229983807, + 0.00600390275940299, + 0.00748165650293231, + 0.008668452501296997, + 0.008278450928628445, + 0.012793246656656265, + 0.009548095054924488, + 0.008367321453988552, + 0.006911925505846739, + 0.009339148178696632, + 0.007310043089091778, + 0.022640328854322433, + 0.00815736036747694, + 0.006909125484526157, + 0.020630935207009315, + 0.007085053250193596, + 0.005815011914819479, + 0.025089334696531296, + 0.015590643510222435, + 0.014378572814166546, + 0.012586004100739956, + 0.016359835863113403, + 0.028108052909374237, + 0.012195338495075703, + 0.011318860575556755, + 0.009726080112159252, + 0.017376268282532692, + 0.004932479001581669, + 0.006154803559184074, + 0.005683002527803183, + 0.008939371444284916, + 0.005755364429205656, + 0.00982977170497179, + 0.010395586490631104, + 0.019792573526501656, + 0.007747205439954996, + 0.008601046167314053, + 0.007905169390141964, + 0.009237506426870823, + 0.009151557460427284, + 0.009628763422369957, + 0.011027161031961441, + 0.02255195379257202, + 0.0131066320464015, + 0.013998632319271564, + 0.012918107211589813, + 0.009867615066468716, + 0.020773211494088173, + 0.020894328132271767, + 0.023502497002482414, + 0.022450704127550125, + 0.028293516486883163, + 0.017857534810900688, + 0.02509184554219246, + 0.02311919815838337, + 0.017001818865537643, + 0.014549676328897476, + 0.0070786066353321075, + 0.010713406838476658, + 0.007280999328941107, + 0.00954557303339243, + 0.00818687491118908, + 0.012658840045332909, + 0.011814783327281475, + 0.007512655574828386, + 0.006594979204237461, + 0.00539428973570466, + 0.00798310898244381, + 0.01161669846624136, + 0.008963894098997116, + 0.0049420492723584175, + 0.010212164372205734, + 0.00488611776381731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.029803991317749023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot begins with a high-angle, sweeping view, capturing the vastness of the valley before slowly and steadily panning forward. Under the soft, golden light of dusk, hundreds of deer converge into a vast torrent, a living, surging river, flowing along the valley floor, frost-stained ochre. This vast herd, their thick winter coats gradating from deep reddish-brown to light tan, moves in a synchronized, rhythmic cadence. The massive antlers of the males rise and fall within the herd, like dead logs floating on a river. As the camera zooms in, the vast collective scene breaks down into crisp individual details: some deer can be seen bounding gracefully over small mounds, while others graze on hay without breaking their rhythm. Their white breath condenses into wisps of mist in the cold, crisp air. The soft setting sun outlines a shimmering golden outline on the edges of the deer's fur and casts long, dynamic shadows behind them. The whole picture is shrouded in a tranquil, majestic and primitive atmosphere full of vitality, showing the awe-inspiring grandeur of nature.", + "seed": 28675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_3707_seed_28675_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": -0.040184934809803965, + "vlm_logprob": -0.040184934809803965, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.040184934809803965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03391522541642189 + ], + "opposite_logprob_tokens": [ + -3.6589152812957764 + ], + "logprob_reward": -0.03391522541642189, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03391522541642189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08700089156627655, + 0.05793224647641182, + 0.0924607515335083, + 0.16678909957408905, + 0.15724778175354004, + 0.10879944264888763, + 0.0955556184053421, + 0.08830186724662781, + 0.12208492308855057, + 0.02382734604179859 + ], + "credit_map": [ + 0.009572802111506462, + 0.010948818176984787, + 0.010217331349849701, + 0.010794823057949543, + 0.010924193076789379, + 0.010517182759940624, + 0.011354166083037853, + 0.012671577744185925, + 0.005478718318045139, + 0.008443458937108517, + 0.00698721781373024, + 0.00628173490986228, + 0.0075355879962444305, + 0.00686054490506649, + 0.009529788047075272, + 0.006815197877585888, + 0.008127563633024693, + 0.011340380646288395, + 0.0105079161003232, + 0.01072637178003788, + 0.01661147177219391, + 0.008586971089243889, + 0.014218258671462536, + 0.0123418178409338, + 0.016411708667874336, + 0.014461686834692955, + 0.015598868019878864, + 0.025179244577884674, + 0.01759055070579052, + 0.01643146015703678, + 0.03296026214957237, + 0.028155304491519928, + 0.017977656796574593, + 0.02029247209429741, + 0.01364672090858221, + 0.02148318849503994, + 0.017082475125789642, + 0.015090938657522202, + 0.0294660534709692, + 0.02220827154815197, + 0.011722707189619541, + 0.017644468694925308, + 0.012665687128901482, + 0.014937542378902435, + 0.012981363572180271, + 0.010395195335149765, + 0.01448781043291092, + 0.013964680954813957, + 0.008980294689536095, + 0.011871918104588985, + 0.012727532535791397, + 0.010002408176660538, + 0.010030931793153286, + 0.013591954484581947, + 0.013261192478239536, + 0.01508938055485487, + 0.009675531648099422, + 0.008269545622169971, + 0.012051139958202839, + 0.011002139188349247, + 0.011208946816623211, + 0.010538240894675255, + 0.014649911783635616, + 0.010906409472227097, + 0.008403654210269451, + 0.008694235235452652, + 0.00894746370613575, + 0.01861514151096344, + 0.0340692363679409, + 0.020157620310783386, + 0.013210033066570759, + 0.0099875433370471, + 0.002081887098029256, + 0.0032774838618934155, + 0.0018050160724669695, + 0.003616133239120245, + 0.003494585631415248, + 0.004476999398320913, + 0.0029671734664589167, + 0.002108067274093628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.03391522541642189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048683732748031616 + ], + "opposite_logprob_tokens": [ + -3.0486836433410645 + ], + "logprob_reward": -0.048683732748031616, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048683732748031616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08250739425420761, + 0.16906560957431793, + 0.0359785296022892, + 0.11987507343292236, + 0.11196324229240417, + 0.12996947765350342, + 0.03759513050317764, + 0.22351112961769104, + 0.0648052915930748, + 0.024729127064347267 + ], + "credit_map": [ + 0.009511265903711319, + 0.010103370063006878, + 0.008814067579805851, + 0.012186005711555481, + 0.007130997721105814, + 0.01056526880711317, + 0.013018692843616009, + 0.011177724227309227, + 0.0175589919090271, + 0.036372918635606766, + 0.014442277140915394, + 0.01238140556961298, + 0.03241460770368576, + 0.01677171140909195, + 0.02021542750298977, + 0.01890825852751732, + 0.0038925164844840765, + 0.006811602506786585, + 0.00329603417776525, + 0.003398174187168479, + 0.006420648656785488, + 0.0034631004091352224, + 0.004550435114651918, + 0.004146020859479904, + 0.008139403536915779, + 0.010591739788651466, + 0.007329544518142939, + 0.01612985134124756, + 0.00888445321470499, + 0.009929690510034561, + 0.03845277801156044, + 0.02041761204600334, + 0.009936570189893246, + 0.01362589281052351, + 0.010589142329990864, + 0.013387767598032951, + 0.016702376306056976, + 0.01527150347828865, + 0.018586482852697372, + 0.013863510452210903, + 0.013606473803520203, + 0.0251082144677639, + 0.0075192684307694435, + 0.014729189686477184, + 0.012388469651341438, + 0.02019941620528698, + 0.017619123682379723, + 0.0187993124127388, + 0.004205729812383652, + 0.004520794842392206, + 0.004735138732939959, + 0.0058418032713234425, + 0.004602773115038872, + 0.004995244089514017, + 0.004873855039477348, + 0.0038197904359549284, + 0.02103099226951599, + 0.02407449670135975, + 0.01453845389187336, + 0.01827137917280197, + 0.06855519115924835, + 0.029398268088698387, + 0.02522362768650055, + 0.022418728098273277, + 0.00597828533500433, + 0.012010029517114162, + 0.0046535674482584, + 0.00938466191291809, + 0.008680816739797592, + 0.01137931365519762, + 0.007698139175772667, + 0.005020479671657085, + 0.00514441030099988, + 0.003068244783207774, + 0.002188082318753004, + 0.0022155686747282743, + 0.002806127769872546, + 0.004347467329353094, + 0.002441748743876815, + 0.0025174764450639486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.048683732748031616 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001181691884995 + ], + "opposite_logprob_tokens": [ + -2.695011854171753 + ], + "logprob_reward": -0.07001181691884995, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001181691884995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06880463659763336, + 0.03710693493485451, + 0.042479291558265686, + 0.1299085021018982, + 0.12272942066192627, + 0.2129283845424652, + 0.05092153325676918, + 0.06030960753560066, + 0.25561079382896423, + 0.019200976938009262 + ], + "credit_map": [ + 0.007172717247158289, + 0.007855277508497238, + 0.012773476541042328, + 0.01060197502374649, + 0.004054659511893988, + 0.005553775932639837, + 0.012693802826106548, + 0.008098945021629333, + 0.004216327797621489, + 0.006631951313465834, + 0.0032116014044731855, + 0.0031560237985104322, + 0.00486533110961318, + 0.0037063094787299633, + 0.006035319529473782, + 0.005284070037305355, + 0.0030548670329153538, + 0.0033755286131054163, + 0.0031885141506791115, + 0.004977052100002766, + 0.0039015612564980984, + 0.006398832891136408, + 0.01041119173169136, + 0.007171741221100092, + 0.011210205964744091, + 0.016932673752307892, + 0.010483679361641407, + 0.01288452185690403, + 0.01732240803539753, + 0.01754610799252987, + 0.024985425174236298, + 0.018543466925621033, + 0.01640218310058117, + 0.011922992765903473, + 0.011494942009449005, + 0.017607063055038452, + 0.01252098847180605, + 0.011413678526878357, + 0.021542081609368324, + 0.01982547901570797, + 0.017239343374967575, + 0.037138428539037704, + 0.02149456925690174, + 0.032074302434921265, + 0.022723093628883362, + 0.02449861541390419, + 0.02845550887286663, + 0.02930450439453125, + 0.004411844536662102, + 0.005494505632668734, + 0.005689876154065132, + 0.00746381189674139, + 0.006463748402893543, + 0.0069764144718647, + 0.005957006476819515, + 0.008464323356747627, + 0.006085796747356653, + 0.005367737263441086, + 0.007268744986504316, + 0.007386751472949982, + 0.010571018792688847, + 0.008705069310963154, + 0.008539947681128979, + 0.006384540349245071, + 0.018252884969115257, + 0.013912194408476353, + 0.017402036115527153, + 0.038690995424985886, + 0.04353395476937294, + 0.04025833308696747, + 0.059484388679265976, + 0.024075988680124283, + 0.0029122421983629465, + 0.002213340485468507, + 0.0016497125616297126, + 0.0025940502528101206, + 0.0017375652678310871, + 0.0021808838937431574, + 0.0036600506864488125, + 0.0022531296126544476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.07001181691884995 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251185968518257 + ], + "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": [ + 0.10752829164266586, + 0.1071510910987854, + 0.04495583474636078, + 0.09062658250331879, + 0.06986915320158005, + 0.2963200807571411, + 0.08067358285188675, + 0.047110460698604584, + 0.11961003392934799, + 0.036154936999082565 + ], + "credit_map": [ + 0.01098520029336214, + 0.012583580799400806, + 0.011591006070375443, + 0.01379260141402483, + 0.014383485540747643, + 0.009805861860513687, + 0.019812924787402153, + 0.014573614113032818, + 0.013355249539017677, + 0.013161005452275276, + 0.014449508860707283, + 0.00992863904684782, + 0.013628873974084854, + 0.0091858534142375, + 0.01696767285466194, + 0.016474274918437004, + 0.0039074840024113655, + 0.005838789511471987, + 0.004260548856109381, + 0.005024306010454893, + 0.008792674168944359, + 0.003538100980222225, + 0.008223812095820904, + 0.005370112136006355, + 0.005925802513957024, + 0.00903026107698679, + 0.007918124087154865, + 0.012283668853342533, + 0.010239042341709137, + 0.010354583151638508, + 0.020930759608745575, + 0.013944339007139206, + 0.005809249822050333, + 0.00645447988063097, + 0.006045087706297636, + 0.008757528848946095, + 0.013194694183766842, + 0.011174223385751247, + 0.010099324397742748, + 0.008334564045071602, + 0.044011663645505905, + 0.05763372778892517, + 0.02524728700518608, + 0.03075919859111309, + 0.02526373602449894, + 0.03464739769697189, + 0.03708348795771599, + 0.04167357459664345, + 0.011285229586064816, + 0.00870139617472887, + 0.011650629341602325, + 0.008895010687410831, + 0.009258655831217766, + 0.008521966636180878, + 0.010452373884618282, + 0.011908316053450108, + 0.004726688377559185, + 0.006014627870172262, + 0.004524341318756342, + 0.0050055659376084805, + 0.00706102279946208, + 0.006013411562889814, + 0.004995429888367653, + 0.008769373409450054, + 0.010872197337448597, + 0.010777910239994526, + 0.007264234591275454, + 0.012159482575953007, + 0.022740576416254044, + 0.013640761375427246, + 0.03208194300532341, + 0.01007291954010725, + 0.004297417122870684, + 0.0030777815263718367, + 0.00260311015881598, + 0.006029014941304922, + 0.0030500823631882668, + 0.006357243284583092, + 0.006080091930925846, + 0.004660192411392927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.005251185968518257 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043062712997198105 + ], + "opposite_logprob_tokens": [ + -3.168062686920166 + ], + "logprob_reward": -0.043062712997198105, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043062712997198105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12143600732088089, + 0.04687977209687233, + 0.022449785843491554, + 0.10401134192943573, + 0.12007803469896317, + 0.23939158022403717, + 0.05899777263402939, + 0.0709623247385025, + 0.16876964271068573, + 0.04702373594045639 + ], + "credit_map": [ + 0.019114945083856583, + 0.020768171176314354, + 0.017817290499806404, + 0.01409100741147995, + 0.011706819757819176, + 0.012747742235660553, + 0.010986213572323322, + 0.014203825034201145, + 0.0033751672599464655, + 0.007253927178680897, + 0.005665857810527086, + 0.004748431034386158, + 0.004582693334668875, + 0.006730556488037109, + 0.00847234483808279, + 0.006050795782357454, + 0.002236794214695692, + 0.002845336450263858, + 0.002557365456596017, + 0.0027773967012763023, + 0.002891157753765583, + 0.001989758340641856, + 0.004269768018275499, + 0.0028822103049606085, + 0.008289355784654617, + 0.011307724751532078, + 0.010136629454791546, + 0.012102809734642506, + 0.011585316620767117, + 0.00891342293471098, + 0.024633148685097694, + 0.017042936757206917, + 0.012507669627666473, + 0.01094616949558258, + 0.01049704197794199, + 0.01626078598201275, + 0.01637006178498268, + 0.01886434108018875, + 0.019181907176971436, + 0.015450059436261654, + 0.02761850133538246, + 0.05296070873737335, + 0.013352440670132637, + 0.025158200412988663, + 0.02342711202800274, + 0.03920392319560051, + 0.028069864958524704, + 0.029600821435451508, + 0.0055466038174927235, + 0.006442617159336805, + 0.005827953573316336, + 0.009592773392796516, + 0.005496310070157051, + 0.008159704506397247, + 0.008204412646591663, + 0.009727396070957184, + 0.006832662969827652, + 0.006537530571222305, + 0.00556025980040431, + 0.0066598267294466496, + 0.01958203874528408, + 0.01052562054246664, + 0.008107521571218967, + 0.007156864739954472, + 0.015533744357526302, + 0.022928450256586075, + 0.011484043672680855, + 0.018259579315781593, + 0.034719668328762054, + 0.02148887701332569, + 0.03219519555568695, + 0.01216007862240076, + 0.003637082874774933, + 0.003935355693101883, + 0.003340967698022723, + 0.005609416868537664, + 0.005030981730669737, + 0.005062188021838665, + 0.01363229751586914, + 0.006775442510843277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.043062712997198105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot begins with a high-angle, sweeping view, capturing the vastness of the valley before slowly and steadily panning forward. Under the soft, golden light of dusk, hundreds of deer converge into a vast torrent, a living, surging river, flowing along the valley floor, frost-stained ochre. This vast herd, their thick winter coats gradating from deep reddish-brown to light tan, moves in a synchronized, rhythmic cadence. The massive antlers of the males rise and fall within the herd, like dead logs floating on a river. As the camera zooms in, the vast collective scene breaks down into crisp individual details: some deer can be seen bounding gracefully over small mounds, while others graze on hay without breaking their rhythm. Their white breath condenses into wisps of mist in the cold, crisp air. The soft setting sun outlines a shimmering golden outline on the edges of the deer's fur and casts long, dynamic shadows behind them. The whole picture is shrouded in a tranquil, majestic and primitive atmosphere full of vitality, showing the awe-inspiring grandeur of nature.", + "seed": 28676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_3707_seed_28676_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": -0.06550067821517587, + "vlm_logprob": -0.06550067821517587, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06550067821517587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18891336023807526 + ], + "opposite_logprob_tokens": [ + -1.813913345336914 + ], + "logprob_reward": -0.18891336023807526, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18891336023807526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12895841896533966, + 0.04442688450217247, + 0.12850606441497803, + 0.1732703298330307, + 0.08812703937292099, + 0.07204260677099228, + 0.11511833965778351, + 0.14657622575759888, + 0.05114899203181267, + 0.05182503163814545 + ], + "credit_map": [ + 0.014610661193728447, + 0.013052345253527164, + 0.01737501658499241, + 0.015591729432344437, + 0.011637846007943153, + 0.0166306272149086, + 0.022869223728775978, + 0.01719096302986145, + 0.004017380066215992, + 0.005291706882417202, + 0.008607075549662113, + 0.005659302230924368, + 0.003831311594694853, + 0.005739755928516388, + 0.005730099510401487, + 0.005550255998969078, + 0.02333371713757515, + 0.03539428487420082, + 0.0075346119701862335, + 0.012292102910578251, + 0.011540901847183704, + 0.015273671597242355, + 0.009959958493709564, + 0.013176814652979374, + 0.006523082498461008, + 0.006642663851380348, + 0.012594547122716904, + 0.015442765317857265, + 0.00892709381878376, + 0.049746330827474594, + 0.024229422211647034, + 0.04916442930698395, + 0.004306611139327288, + 0.0040223863907158375, + 0.006188543513417244, + 0.009055591188371181, + 0.006770308595150709, + 0.021543197333812714, + 0.014770995825529099, + 0.02146940864622593, + 0.004467598628252745, + 0.010804281570017338, + 0.008611204102635384, + 0.008533181622624397, + 0.013675193302333355, + 0.007466143928468227, + 0.011106560938060284, + 0.007378444075584412, + 0.007097274996340275, + 0.010437073186039925, + 0.015586468391120434, + 0.014577637426555157, + 0.01779622584581375, + 0.011369487270712852, + 0.023466743528842926, + 0.014787429943680763, + 0.011742810718715191, + 0.02128622867166996, + 0.012450463138520718, + 0.01837908662855625, + 0.01853688433766365, + 0.016305766999721527, + 0.03231905773282051, + 0.01555592194199562, + 0.0034113957080990076, + 0.005112112499773502, + 0.0061226277612149715, + 0.0060278247110545635, + 0.007168620824813843, + 0.005115787032991648, + 0.00914324726909399, + 0.009047375060617924, + 0.0031042583286762238, + 0.005293813534080982, + 0.003932364750653505, + 0.005060495343059301, + 0.0072351256385445595, + 0.011188003234565258, + 0.009972985833883286, + 0.006037984974682331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.18891336023807526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005004584789276 + ], + "opposite_logprob_tokens": [ + -2.695050001144409 + ], + "logprob_reward": -0.07005004584789276, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005004584789276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1417021006345749, + 0.05255751684308052, + 0.24079357087612152, + 0.11872800439596176, + 0.06653614342212677, + 0.057509928941726685, + 0.08585033565759659, + 0.1028866171836853, + 0.04052344709634781, + 0.09291236847639084 + ], + "credit_map": [ + 0.010822883807122707, + 0.009769167751073837, + 0.023089716210961342, + 0.02755063772201538, + 0.010494602844119072, + 0.021779989823698997, + 0.022688327357172966, + 0.01550677977502346, + 0.0035837360192090273, + 0.004367918241769075, + 0.009699453599750996, + 0.00580727169290185, + 0.004220021888613701, + 0.011239463463425636, + 0.006900508888065815, + 0.006739145144820213, + 0.03922112286090851, + 0.0703253522515297, + 0.014660544693470001, + 0.02466145157814026, + 0.025953048840165138, + 0.015500801615417004, + 0.022394899278879166, + 0.028076350688934326, + 0.004205323755741119, + 0.004555618856102228, + 0.006029161158949137, + 0.006469636224210262, + 0.006098792422562838, + 0.04048429802060127, + 0.011928400956094265, + 0.03895677253603935, + 0.0034854812547564507, + 0.003262875135987997, + 0.00390872685238719, + 0.005682672373950481, + 0.004766751080751419, + 0.018812410533428192, + 0.009646781720221043, + 0.016970446333289146, + 0.003855274524539709, + 0.009261306375265121, + 0.005426034331321716, + 0.007307298947125673, + 0.012104044668376446, + 0.005809789057821035, + 0.007343063596636057, + 0.006403113249689341, + 0.0054910616017878056, + 0.008488531224429607, + 0.013085104525089264, + 0.013091303408145905, + 0.012409154325723648, + 0.007553334813565016, + 0.014951399527490139, + 0.010780444368720055, + 0.007767791859805584, + 0.011441798880696297, + 0.009315697476267815, + 0.010681457817554474, + 0.011117471382021904, + 0.013530518859624863, + 0.023515406996011734, + 0.015516473911702633, + 0.0041272337548434734, + 0.004121438134461641, + 0.005433786194771528, + 0.00528517784550786, + 0.005297569558024406, + 0.005731181241571903, + 0.0054141199216246605, + 0.005112937185913324, + 0.006820688955485821, + 0.006482528522610664, + 0.005294946953654289, + 0.0070141395553946495, + 0.014688797295093536, + 0.027693267911672592, + 0.015917446464300156, + 0.009000547230243683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07005004584789276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043055977672338486 + ], + "opposite_logprob_tokens": [ + -3.168056011199951 + ], + "logprob_reward": -0.043055977672338486, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043055977672338486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08344103395938873, + 0.03851901739835739, + 0.1960810124874115, + 0.07792560011148453, + 0.08036506921052933, + 0.05384514853358269, + 0.07435720413923264, + 0.124403215944767, + 0.2604471445083618, + 0.010615590028464794 + ], + "credit_map": [ + 0.009260711260139942, + 0.005758353974670172, + 0.013558932580053806, + 0.012509089894592762, + 0.00691953906789422, + 0.01205554697662592, + 0.014242993667721748, + 0.00913586188107729, + 0.0047836098819971085, + 0.004077714402228594, + 0.004561875481158495, + 0.004041070118546486, + 0.00443279929459095, + 0.006463989149779081, + 0.005454719997942448, + 0.004703239072114229, + 0.013110030442476273, + 0.017148733139038086, + 0.024466196075081825, + 0.03620845824480057, + 0.01655416190624237, + 0.039021726697683334, + 0.030546177178621292, + 0.019025517627596855, + 0.004053532611578703, + 0.004454908426851034, + 0.006472254171967506, + 0.006678330712020397, + 0.0071991547010838985, + 0.01966257393360138, + 0.014123642817139626, + 0.015281195752322674, + 0.004232044331729412, + 0.006395532749593258, + 0.004461606033146381, + 0.005575943272560835, + 0.007858735509216785, + 0.0208686925470829, + 0.00843447633087635, + 0.022538047283887863, + 0.003677801229059696, + 0.009118088521063328, + 0.0030893380753695965, + 0.005332238506525755, + 0.011045305989682674, + 0.00561343040317297, + 0.007441261783242226, + 0.008527682162821293, + 0.005219014827162027, + 0.007292502094060183, + 0.01059923879802227, + 0.010241195559501648, + 0.01194063201546669, + 0.008452915586531162, + 0.010834334418177605, + 0.009777369908988476, + 0.011147383600473404, + 0.021590150892734528, + 0.01314785610884428, + 0.018772536888718605, + 0.015399170108139515, + 0.011599740013480186, + 0.02052583545446396, + 0.01222054660320282, + 0.018255440518260002, + 0.030631117522716522, + 0.033180151134729385, + 0.03567086160182953, + 0.029042556881904602, + 0.05206381529569626, + 0.03803957998752594, + 0.023563606664538383, + 0.0010174651397392154, + 0.0015620817430317402, + 0.0008810756262391806, + 0.0010224307188764215, + 0.0016675000078976154, + 0.0011287740198895335, + 0.0020175736863166094, + 0.0013186892028898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.043055977672338486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021962826140224934 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021962826140224934, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021962826140224934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08283994346857071, + 0.05355524271726608, + 0.2955581247806549, + 0.0975908488035202, + 0.040036726742982864, + 0.09372754395008087, + 0.09667310863733292, + 0.09037534892559052, + 0.04570968821644783, + 0.10393344610929489 + ], + "credit_map": [ + 0.009757505729794502, + 0.008080671541392803, + 0.012208666652441025, + 0.012165737338364124, + 0.0065658739767968655, + 0.012901496142148972, + 0.010761234909296036, + 0.010398756712675095, + 0.003668173449113965, + 0.004782147239893675, + 0.010176250711083412, + 0.007168972864747047, + 0.004208632279187441, + 0.008016577921807766, + 0.00697512598708272, + 0.008559361100196838, + 0.04612627625465393, + 0.08821215480566025, + 0.020377319306135178, + 0.02997904270887375, + 0.029550600796937943, + 0.025950178503990173, + 0.027587639167904854, + 0.027774924412369728, + 0.00387727003544569, + 0.0038694264367222786, + 0.004791988059878349, + 0.00493447482585907, + 0.0042335218749940395, + 0.032020412385463715, + 0.010068949311971664, + 0.03379480540752411, + 0.0015288766007870436, + 0.0030004987493157387, + 0.0020033440086990595, + 0.003061545081436634, + 0.0060142632573843, + 0.009775379672646523, + 0.005477291531860828, + 0.009175528772175312, + 0.009912307374179363, + 0.021231530234217644, + 0.005311308428645134, + 0.008574817329645157, + 0.020584026351571083, + 0.008043327368795872, + 0.01077831070870161, + 0.009291911497712135, + 0.005601312033832073, + 0.010153536684811115, + 0.0057309153489768505, + 0.007383477408438921, + 0.027966341003775597, + 0.013067289255559444, + 0.012762002646923065, + 0.014008236117661, + 0.007671631872653961, + 0.012204881757497787, + 0.0065474966540932655, + 0.008863266557455063, + 0.010126134380698204, + 0.013467323035001755, + 0.019152875989675522, + 0.012341740541160107, + 0.00349369621835649, + 0.0054124947637319565, + 0.005419199354946613, + 0.006188901606947184, + 0.0055575892329216, + 0.008365943096578121, + 0.0062864418141543865, + 0.004985421895980835, + 0.006890703458338976, + 0.011335271410644054, + 0.0063009681180119514, + 0.00883941724896431, + 0.014909657649695873, + 0.02618645876646042, + 0.019631391391158104, + 0.00983957014977932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 3 + ], + "selected_reward": -0.0021962826140224934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328772470355034 + ], + "opposite_logprob_tokens": [ + -3.773287773132324 + ], + "logprob_reward": -0.02328772470355034, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328772470355034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10881160944700241, + 0.09643932431936264, + 0.05310651659965515, + 0.11635353416204453, + 0.03159287944436073, + 0.05609074980020523, + 0.13817238807678223, + 0.12244858592748642, + 0.044347986578941345, + 0.23263642191886902 + ], + "credit_map": [ + 0.009191486053168774, + 0.007517741061747074, + 0.01566767692565918, + 0.017135683447122574, + 0.009444418363273144, + 0.021739155054092407, + 0.014966650865972042, + 0.013148799538612366, + 0.008011355996131897, + 0.011152340099215508, + 0.015592849813401699, + 0.010501070879399776, + 0.008405559696257114, + 0.018840523436665535, + 0.00994141772389412, + 0.013994202949106693, + 0.004124682862311602, + 0.005519441328942776, + 0.006036924198269844, + 0.010524418205022812, + 0.00552471075206995, + 0.007447316776961088, + 0.007247732020914555, + 0.0066812895238399506, + 0.003027114551514387, + 0.0038191077765077353, + 0.00731303496286273, + 0.00828560907393694, + 0.005162274930626154, + 0.03966689482331276, + 0.0123861413449049, + 0.03669334575533867, + 0.0018694914178922772, + 0.003205677727237344, + 0.0037726876325905323, + 0.005069988779723644, + 0.0031331637874245644, + 0.0034613751340657473, + 0.006412178743630648, + 0.004668317269533873, + 0.004541246220469475, + 0.005494249053299427, + 0.004880120977759361, + 0.006545368116348982, + 0.010569404810667038, + 0.006743350997567177, + 0.008282732218503952, + 0.009034276939928532, + 0.010639192536473274, + 0.012848472222685814, + 0.013471928425133228, + 0.015698209404945374, + 0.0349915586411953, + 0.015453309752047062, + 0.019971126690506935, + 0.015098587609827518, + 0.0065993135794997215, + 0.00956515409052372, + 0.016827667132019997, + 0.020013360306620598, + 0.01407671719789505, + 0.02187410742044449, + 0.020963991060853004, + 0.012528277933597565, + 0.0034977493342012167, + 0.005722209811210632, + 0.004858305212110281, + 0.0054580168798565865, + 0.005461817607283592, + 0.00880393385887146, + 0.005693844053894281, + 0.0048521109856665134, + 0.00902424193918705, + 0.011697185225784779, + 0.009295405820012093, + 0.017366183921694756, + 0.03073427639901638, + 0.07144185155630112, + 0.06065651774406433, + 0.022420749068260193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.02328772470355034 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic drone shot begins with a high-angle, sweeping view, capturing the vastness of the valley before slowly and steadily panning forward. Under the soft, golden light of dusk, hundreds of deer converge into a vast torrent, a living, surging river, flowing along the valley floor, frost-stained ochre. This vast herd, their thick winter coats gradating from deep reddish-brown to light tan, moves in a synchronized, rhythmic cadence. The massive antlers of the males rise and fall within the herd, like dead logs floating on a river. As the camera zooms in, the vast collective scene breaks down into crisp individual details: some deer can be seen bounding gracefully over small mounds, while others graze on hay without breaking their rhythm. Their white breath condenses into wisps of mist in the cold, crisp air. The soft setting sun outlines a shimmering golden outline on the edges of the deer's fur and casts long, dynamic shadows behind them. The whole picture is shrouded in a tranquil, majestic and primitive atmosphere full of vitality, showing the awe-inspiring grandeur of nature.", + "seed": 28677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_8/index_3707_seed_28677_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": -0.10994398901239037, + "vlm_logprob": -0.10994398901239037, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.10994398901239037, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20854054391384125 + ], + "opposite_logprob_tokens": [ + -1.7085405588150024 + ], + "logprob_reward": -0.20854054391384125, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20854054391384125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3650527000427246, + 0.09172985702753067, + 0.2615560293197632, + 0.08140624314546585, + 0.016931194812059402, + 0.02761044166982174, + 0.017884353175759315, + 0.06837675720453262, + 0.03151540458202362, + 0.03793708235025406 + ], + "credit_map": [ + 0.04943718761205673, + 0.06473441421985626, + 0.028572384268045425, + 0.03429407253861427, + 0.046128712594509125, + 0.04374953731894493, + 0.03668728098273277, + 0.06144905835390091, + 0.009238064289093018, + 0.010239750146865845, + 0.011624462902545929, + 0.009652734734117985, + 0.013874955475330353, + 0.01168636605143547, + 0.012506657280027866, + 0.012906856834888458, + 0.019557956606149673, + 0.09402357041835785, + 0.01794322580099106, + 0.012179924175143242, + 0.01977316476404667, + 0.013205993920564651, + 0.025142163038253784, + 0.05972999706864357, + 0.007539716549217701, + 0.007786594796925783, + 0.010048861615359783, + 0.007560406345874071, + 0.009475329890847206, + 0.009555241093039513, + 0.01705404371023178, + 0.012386040762066841, + 0.002062857849523425, + 0.0035740972962230444, + 0.002845583949238062, + 0.0023087698500603437, + 0.0017706723883748055, + 0.0011884683044627309, + 0.0017157815163955092, + 0.0014649636577814817, + 0.003546385560184717, + 0.004338507074862719, + 0.003309420309960842, + 0.0027152211405336857, + 0.004101866856217384, + 0.0034152932930737734, + 0.0027451615314930677, + 0.0034385842736810446, + 0.0024570494424551725, + 0.0019861578475683928, + 0.002767295343801379, + 0.0031459995079785585, + 0.0012212262954562902, + 0.0011269471142441034, + 0.0027455859817564487, + 0.0024340893141925335, + 0.010009334422647953, + 0.008494115434587002, + 0.00655802758410573, + 0.00936438050121069, + 0.008948720060288906, + 0.006872582249343395, + 0.010117292404174805, + 0.008012301288545132, + 0.003982138819992542, + 0.005129146855324507, + 0.005467790178954601, + 0.005642710719257593, + 0.002433669986203313, + 0.0020426050759851933, + 0.003996286075562239, + 0.002821055008098483, + 0.006419174838811159, + 0.007098277099430561, + 0.004815821070224047, + 0.0037943802308291197, + 0.0036676269955933094, + 0.0029952621553093195, + 0.005184341222047806, + 0.003962201066315174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.20854054391384125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372552990913391 + ], + "opposite_logprob_tokens": [ + -3.4087255001068115 + ], + "logprob_reward": -0.03372552990913391, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372552990913391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4368191957473755, + 0.04004359617829323, + 0.2050212025642395, + 0.11004447937011719, + 0.019447972998023033, + 0.02439061366021633, + 0.021259281784296036, + 0.0691751167178154, + 0.04222249612212181, + 0.031576067209243774 + ], + "credit_map": [ + 0.06286537647247314, + 0.08933800458908081, + 0.03135445713996887, + 0.04173937439918518, + 0.0471295490860939, + 0.04200737923383713, + 0.03943876177072525, + 0.0829463079571724, + 0.004452763590961695, + 0.005654740147292614, + 0.004749024286866188, + 0.005504838190972805, + 0.006489485502243042, + 0.0028900003526359797, + 0.005570936482399702, + 0.004731808323413134, + 0.014899336732923985, + 0.07076356559991837, + 0.009795854799449444, + 0.008849571458995342, + 0.01462917122989893, + 0.01078786700963974, + 0.02252211794257164, + 0.05277373269200325, + 0.007064330857247114, + 0.007787788286805153, + 0.01039411872625351, + 0.00871838815510273, + 0.008369696326553822, + 0.007582338992506266, + 0.03516443818807602, + 0.02496338076889515, + 0.0021734710317105055, + 0.0020314091816544533, + 0.003331669606268406, + 0.0034217368811368942, + 0.0019857778679579496, + 0.0013868637615814805, + 0.0028889074455946684, + 0.002228138269856572, + 0.0029723611660301685, + 0.005388163961470127, + 0.002600966254249215, + 0.0021578033920377493, + 0.0029416289180517197, + 0.002913659205660224, + 0.002183799631893635, + 0.0032322295010089874, + 0.003318327246233821, + 0.0028787131886929274, + 0.0029133346397429705, + 0.0034666601568460464, + 0.002023556036874652, + 0.0019368751673027873, + 0.0025487099774181843, + 0.002173105487599969, + 0.011441493406891823, + 0.009719581343233585, + 0.007214718963950872, + 0.009680449962615967, + 0.006917235907167196, + 0.009788762778043747, + 0.00743837608024478, + 0.006974501069635153, + 0.005184545647352934, + 0.0062653254717588425, + 0.004387546330690384, + 0.005113380029797554, + 0.00428551621735096, + 0.006148120854049921, + 0.005926215089857578, + 0.004911850206553936, + 0.00627304520457983, + 0.006384107284247875, + 0.003237043274566531, + 0.0031681249383836985, + 0.0030829929746687412, + 0.0025251442566514015, + 0.0037433893885463476, + 0.0031622222159057856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03372552990913391 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252071738243103 + ], + "opposite_logprob_tokens": [ + -1.502071738243103 + ], + "logprob_reward": -0.252071738243103, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252071738243103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1842920035123825, + 0.0883202776312828, + 0.035227786749601364, + 0.17597243189811707, + 0.06445639580488205, + 0.13796454668045044, + 0.05012717470526695, + 0.06659871339797974, + 0.16601966321468353, + 0.0310209933668375 + ], + "credit_map": [ + 0.025000497698783875, + 0.03271343931555748, + 0.017732294276356697, + 0.021807661280035973, + 0.01690765842795372, + 0.012876087799668312, + 0.023089546710252762, + 0.03416481614112854, + 0.010380509309470654, + 0.010833358392119408, + 0.013071791268885136, + 0.0102413734421134, + 0.010157064534723759, + 0.01088366936892271, + 0.011807180009782314, + 0.010945331305265427, + 0.0032048330176621675, + 0.006480764597654343, + 0.004414604511111975, + 0.003998659551143646, + 0.0029806937091052532, + 0.002297201193869114, + 0.00751155149191618, + 0.004339477512985468, + 0.011906315572559834, + 0.013246733695268631, + 0.02088465541601181, + 0.01385434065014124, + 0.015225851908326149, + 0.010698553174734116, + 0.05309450626373291, + 0.037061482667922974, + 0.008515788242220879, + 0.01579166017472744, + 0.007348149083554745, + 0.007193104363977909, + 0.008396762423217297, + 0.006318294443190098, + 0.005871184170246124, + 0.005021451972424984, + 0.013148321770131588, + 0.029643937945365906, + 0.022703178226947784, + 0.015576181933283806, + 0.015387028455734253, + 0.01183431688696146, + 0.01428536418825388, + 0.015386213548481464, + 0.005659373011440039, + 0.005027554463595152, + 0.006873448379337788, + 0.008776340633630753, + 0.0057309470139443874, + 0.006296504754573107, + 0.0054979524575173855, + 0.0062650516629219055, + 0.013945742510259151, + 0.009558393619954586, + 0.007421554531902075, + 0.007154267746955156, + 0.008401486091315746, + 0.007150750141590834, + 0.0060211606323719025, + 0.006945358589291573, + 0.02153274416923523, + 0.02056882716715336, + 0.02387792430818081, + 0.026549573987722397, + 0.012843945063650608, + 0.017754580825567245, + 0.024886611849069595, + 0.018005449324846268, + 0.004564998205751181, + 0.00537170609459281, + 0.0027603295166045427, + 0.004805080592632294, + 0.0030609576497226954, + 0.0020872328896075487, + 0.0050509776920080185, + 0.0033197083976119757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.252071738243103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067339395172894 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067339395172894, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067339395172894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33514273166656494, + 0.08179143816232681, + 0.04441279172897339, + 0.03797400742769241, + 0.05125543102622032, + 0.039242908358573914, + 0.055637113749980927, + 0.2824857234954834, + 0.024928318336606026, + 0.04712948948144913 + ], + "credit_map": [ + 0.04346087574958801, + 0.054126035422086716, + 0.029744556173682213, + 0.041372090578079224, + 0.04051375389099121, + 0.04061662033200264, + 0.03859751671552658, + 0.04671128839254379, + 0.008165912702679634, + 0.012827142141759396, + 0.011158375069499016, + 0.008657692931592464, + 0.011134264059364796, + 0.010123224928975105, + 0.009517498314380646, + 0.010207335464656353, + 0.0031803660094738007, + 0.01611209288239479, + 0.0032993347849696875, + 0.003051784122362733, + 0.003963070455938578, + 0.002652819035574794, + 0.00357676949352026, + 0.00857655517756939, + 0.004086596891283989, + 0.0044667464680969715, + 0.006135179661214352, + 0.003707465948536992, + 0.006072203628718853, + 0.004721516277641058, + 0.005192024167627096, + 0.003592276480048895, + 0.006470774300396442, + 0.010706012137234211, + 0.010934354737401009, + 0.006382499821484089, + 0.004660317674279213, + 0.0046469662338495255, + 0.0037533792201429605, + 0.0037011278327554464, + 0.004324969835579395, + 0.009193225763738155, + 0.006064585875719786, + 0.004724529106169939, + 0.003922539763152599, + 0.004197962116450071, + 0.0033614409621804953, + 0.003453653072938323, + 0.008023707196116447, + 0.005907994695007801, + 0.007017467636615038, + 0.0072990404441952705, + 0.006456543691456318, + 0.005159210413694382, + 0.008314365521073341, + 0.007458782289177179, + 0.06280504912137985, + 0.052574485540390015, + 0.03121219389140606, + 0.027742063626646996, + 0.029833192005753517, + 0.02455442026257515, + 0.025641167536377907, + 0.028123172000050545, + 0.0025634500198066235, + 0.0030996622517704964, + 0.002167273545637727, + 0.003605472156777978, + 0.0021597486920654774, + 0.002616000361740589, + 0.005371683277189732, + 0.0033450275659561157, + 0.008330895565450191, + 0.00897266250103712, + 0.004734314978122711, + 0.005463025998324156, + 0.004931475035846233, + 0.00395447900518775, + 0.006165903992950916, + 0.004576732870191336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0067339395172894 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 valley with hundreds of deer moving together?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the deer depicted with thick winter coats ranging from deep reddish-brown to light tan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show the camera starting with a high-angle view and then zooming in?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lighting characterized by soft golden dusk light creating long shadows and a shimmering outline on the fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are visible details including deer bounding over mounds, grazing on hay, and white breath condensing into mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864819347858429 + ], + "opposite_logprob_tokens": [ + -3.0486481189727783 + ], + "logprob_reward": -0.04864819347858429, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864819347858429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2677210569381714, + 0.09342354536056519, + 0.2211633324623108, + 0.054571934044361115, + 0.028208903968334198, + 0.03806290775537491, + 0.06735187768936157, + 0.16811785101890564, + 0.04423007369041443, + 0.017148563638329506 + ], + "credit_map": [ + 0.045146744698286057, + 0.04073218256235123, + 0.023299407213926315, + 0.024129748344421387, + 0.036543622612953186, + 0.03788125514984131, + 0.025816824287176132, + 0.034171245992183685, + 0.015080754645168781, + 0.015625566244125366, + 0.012712448835372925, + 0.009986849501729012, + 0.010251777246594429, + 0.008150315843522549, + 0.011228526011109352, + 0.010387295857071877, + 0.016730695962905884, + 0.08077087253332138, + 0.012169470079243183, + 0.011774946935474873, + 0.01565385051071644, + 0.010428279638290405, + 0.021117132157087326, + 0.052518073469400406, + 0.003502584993839264, + 0.004750403109937906, + 0.006193940062075853, + 0.004996645729988813, + 0.006616997066885233, + 0.004954018164426088, + 0.014364969916641712, + 0.009192375466227531, + 0.0033317990601062775, + 0.005259878467768431, + 0.0032902078237384558, + 0.002471763640642166, + 0.003717459738254547, + 0.003873876528814435, + 0.0029945445712655783, + 0.0032693722750991583, + 0.003932066261768341, + 0.005622889846563339, + 0.004485775716602802, + 0.0034271185286343098, + 0.006352858617901802, + 0.003630329854786396, + 0.004131341353058815, + 0.006480521988123655, + 0.010360072366893291, + 0.008531025610864162, + 0.007034465204924345, + 0.00884078536182642, + 0.008339442312717438, + 0.007130549754947424, + 0.008281674236059189, + 0.008833861909806728, + 0.026396965608000755, + 0.02804582007229328, + 0.01593868061900139, + 0.01828896254301071, + 0.018763108178973198, + 0.017877884209156036, + 0.024574216455221176, + 0.018232202157378197, + 0.005172855220735073, + 0.005145531613379717, + 0.004760148469358683, + 0.006551676895469427, + 0.004492084961384535, + 0.004482646007090807, + 0.0084134042263031, + 0.0052117230370640755, + 0.0021240292116999626, + 0.0031067344825714827, + 0.002444145968183875, + 0.002850641729310155, + 0.001324760145507753, + 0.0010504461824893951, + 0.002435339381918311, + 0.0018124664202332497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.04864819347858429 + } + ], + "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/0000022/group_9/index_tensor([2808, 2808])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_tensor([2808, 2808])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f7d2c18ef3594c120512442ad228fde4ff77a9df --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_tensor([2808, 2808])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera slowly and smoothly rotates around a crystal sphere, keeping it centered within the frame. The sphere, flawless and highly polished to a luminous translucence, rests securely on a dark, almost indistinguishable base, set against a backdrop of pure, profound darkness. The sphere's interior, far from empty, houses a magnificent miniature universe, a landscape of deep indigo and violet, dotted with countless tiny stars that sparkle like diamonds and gently swirling nebulae of fiery pink and gold gas. The sphere itself rotates in a steady, hypnotic rhythm, causing the galaxies and stardust within it to shift, distort, and reshape themselves, as if enacting the evolution of the universe over eons. The only light source for the entire scene comes from the self-luminous universe inside the crystal ball. It emits a soft and mysterious light, casting a changing colorful halo on the smooth curved surface of the crystal ball and spreading into the surrounding darkness, creating a mysterious atmosphere that is deep, tranquil and full of infinite awe.", + "seed": 26164, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_2808_seed_26164_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": -0.015074790129438043, + "vlm_logprob": -0.015074790129438043, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.015074790129438043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007247466593980789 + ], + "opposite_logprob_tokens": [ + -5.632247447967529 + ], + "logprob_reward": -0.007247466593980789, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007247466593980789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07198169827461243, + 0.05812562257051468, + 0.06645219773054123, + 0.023720482364296913, + 0.02359398454427719, + 0.03348681330680847, + 0.013262979686260223, + 0.0416298508644104, + 0.05633086338639259, + 0.6114153861999512 + ], + "credit_map": [ + 0.010456274263560772, + 0.014233427122235298, + 0.006641861516982317, + 0.006334482692182064, + 0.010722250677645206, + 0.009425992146134377, + 0.00670842407271266, + 0.007458996959030628, + 0.006629351992160082, + 0.008288045413792133, + 0.016997408121824265, + 0.0033948796335607767, + 0.009955951012670994, + 0.004504821728914976, + 0.004895370453596115, + 0.0034597960766404867, + 0.0041060312651097775, + 0.028179442510008812, + 0.0021746258717030287, + 0.003041930729523301, + 0.02009575255215168, + 0.003272236557677388, + 0.003322095377370715, + 0.002260088687762618, + 0.0033555040135979652, + 0.0041904086247086525, + 0.002155301393941045, + 0.001584861078299582, + 0.004276413004845381, + 0.00287410500459373, + 0.003227885812520981, + 0.002056006109341979, + 0.0024756495840847492, + 0.005643225274980068, + 0.0019113935995846987, + 0.0019992957822978497, + 0.003807396162301302, + 0.0019885136280208826, + 0.002612432697787881, + 0.00315608037635684, + 0.004775451496243477, + 0.003763372078537941, + 0.0036342721432447433, + 0.00595328351482749, + 0.005102631635963917, + 0.003181697567924857, + 0.003998592030256987, + 0.0030775158666074276, + 0.002019181614741683, + 0.0021263384260237217, + 0.0011425127740949392, + 0.0015528828371316195, + 0.002451440552249551, + 0.0015517383581027389, + 0.0009587342501617968, + 0.0014601519796997309, + 0.006725255399942398, + 0.0061402576975524426, + 0.004349962808191776, + 0.006765496451407671, + 0.007302625570446253, + 0.003803394502028823, + 0.0031951237469911575, + 0.003347741672769189, + 0.008159180171787739, + 0.010027466341853142, + 0.005110376514494419, + 0.011432278901338577, + 0.008820476941764355, + 0.004514229018241167, + 0.004088268615305424, + 0.004178590606898069, + 0.025111623108386993, + 0.19075194001197815, + 0.01686456985771656, + 0.03829100355505943, + 0.2034430056810379, + 0.03759552910923958, + 0.03622158616781235, + 0.06313620507717133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.007247466593980789 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038116808980703354 + ], + "opposite_logprob_tokens": [ + -3.288116693496704 + ], + "logprob_reward": -0.038116808980703354, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038116808980703354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17964237928390503, + 0.08400000631809235, + 0.06394064426422119, + 0.0546414777636528, + 0.03536668047308922, + 0.039704322814941406, + 0.04662398248910904, + 0.04892231523990631, + 0.048175059258937836, + 0.3989832103252411 + ], + "credit_map": [ + 0.01835576817393303, + 0.03501738980412483, + 0.030314143747091293, + 0.00887040700763464, + 0.04432898759841919, + 0.017261534929275513, + 0.01435774751007557, + 0.01113639771938324, + 0.006927578244358301, + 0.01749470643699169, + 0.013625850901007652, + 0.0052450597286224365, + 0.022718841210007668, + 0.008229227736592293, + 0.005585927981883287, + 0.004172809887677431, + 0.01368025690317154, + 0.014335044659674168, + 0.006966653745621443, + 0.0037873247638344765, + 0.00965840369462967, + 0.004664041567593813, + 0.005232478026300669, + 0.005616434849798679, + 0.005078819114714861, + 0.009715468622744083, + 0.006865158677101135, + 0.004287625662982464, + 0.008393464609980583, + 0.009196801111102104, + 0.006320308428257704, + 0.00478382920846343, + 0.006231549195945263, + 0.007053337059915066, + 0.003285802435129881, + 0.002988094463944435, + 0.006452930625528097, + 0.004176092334091663, + 0.0027582256589084864, + 0.002420647069811821, + 0.0044255005195736885, + 0.00636032409965992, + 0.005546677857637405, + 0.004425568971782923, + 0.005740162450820208, + 0.004678443539887667, + 0.003989987540990114, + 0.004537653643637896, + 0.005657949950546026, + 0.00624888576567173, + 0.0029744410421699286, + 0.010735202580690384, + 0.008267899043858051, + 0.0034603066742420197, + 0.005280829034745693, + 0.003998461179435253, + 0.005758036393672228, + 0.006307902745902538, + 0.004418910015374422, + 0.012792066670954227, + 0.006530637852847576, + 0.003449884010478854, + 0.004823782481253147, + 0.0048410939052701, + 0.005323014687746763, + 0.008470755070447922, + 0.003369453828781843, + 0.011226276867091656, + 0.007916475646197796, + 0.003665143856778741, + 0.003501604311168194, + 0.004702328704297543, + 0.02847360447049141, + 0.140335813164711, + 0.013679389841854572, + 0.046580854803323746, + 0.07882494479417801, + 0.02639029547572136, + 0.016029000282287598, + 0.04866927117109299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.038116808980703354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096316389739513 + ], + "opposite_logprob_tokens": [ + -4.511096477508545 + ], + "logprob_reward": -0.011096316389739513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096316389739513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11233695596456528, + 0.07134370505809784, + 0.3894108533859253, + 0.07529040426015854, + 0.06521707773208618, + 0.02155325375497341, + 0.005789066199213266, + 0.028552087023854256, + 0.0521015003323555, + 0.17840509116649628 + ], + "credit_map": [ + 0.013983272016048431, + 0.017557982355356216, + 0.01099403202533722, + 0.008553438819944859, + 0.020598886534571648, + 0.0199848972260952, + 0.010565130971372128, + 0.010099314153194427, + 0.006500375922769308, + 0.019994955509901047, + 0.0065084402449429035, + 0.005516442470252514, + 0.013087623752653599, + 0.007474273908883333, + 0.007348271086812019, + 0.0049133216962218285, + 0.02503928169608116, + 0.16344495117664337, + 0.015059649012982845, + 0.01457350142300129, + 0.12977337837219238, + 0.01732996106147766, + 0.01307289395481348, + 0.011117199435830116, + 0.009707731194794178, + 0.014701947569847107, + 0.006161840632557869, + 0.0055578709580004215, + 0.0149436229839921, + 0.007874899543821812, + 0.011001828126609325, + 0.005340653471648693, + 0.003912812098860741, + 0.01920054852962494, + 0.004602478817105293, + 0.008176017552614212, + 0.011057053692638874, + 0.005912356544286013, + 0.008520657196640968, + 0.0038351574912667274, + 0.0028586050029844046, + 0.004108764231204987, + 0.002179007977247238, + 0.002481749514117837, + 0.003797460813075304, + 0.002333917422220111, + 0.001843038946390152, + 0.0019507113611325622, + 0.0007024394581094384, + 0.0008298594621010125, + 0.00044782826444134116, + 0.0012041715672239661, + 0.0009383729775436223, + 0.0005204752669669688, + 0.0006930949748493731, + 0.00045282405335456133, + 0.003468224313110113, + 0.004021008964627981, + 0.0024011777713894844, + 0.007465839851647615, + 0.0039460062980651855, + 0.0019002397311851382, + 0.002729726256802678, + 0.0026198632549494505, + 0.006769418716430664, + 0.00773124024271965, + 0.004356786608695984, + 0.012478131800889969, + 0.007986700162291527, + 0.0046892957761883736, + 0.0038436437025666237, + 0.004246282856911421, + 0.006115267518907785, + 0.06419415026903152, + 0.0044092778116464615, + 0.009903519414365292, + 0.06233493611216545, + 0.00930851511657238, + 0.008808613754808903, + 0.01333080418407917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.011096316389739513 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822702983394265 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.002822702983394265, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822702983394265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10937155783176422, + 0.08049668371677399, + 0.14207220077514648, + 0.04179101437330246, + 0.09860647469758987, + 0.03407707065343857, + 0.03669831529259682, + 0.03205447643995285, + 0.1396622359752655, + 0.28516995906829834 + ], + "credit_map": [ + 0.018094036728143692, + 0.023550719022750854, + 0.012596881948411465, + 0.006503711454570293, + 0.018503056839108467, + 0.012704412452876568, + 0.0069359648041427135, + 0.010482773184776306, + 0.006682886742055416, + 0.014955420978367329, + 0.01967945322394371, + 0.004841956309974194, + 0.013578051701188087, + 0.007427029311656952, + 0.0073089757934212685, + 0.006022911053150892, + 0.005790287628769875, + 0.05362294241786003, + 0.005985984578728676, + 0.005566938780248165, + 0.04921581223607063, + 0.007644164841622114, + 0.008826910518109798, + 0.005419153720140457, + 0.005614650435745716, + 0.006504939869046211, + 0.0034260174725204706, + 0.003144429996609688, + 0.007902606390416622, + 0.00646105594933033, + 0.005012813955545425, + 0.0037244975101202726, + 0.015579428523778915, + 0.01896713674068451, + 0.009205268695950508, + 0.010599026456475258, + 0.015876932069659233, + 0.009158702567219734, + 0.013385222293436527, + 0.005834753625094891, + 0.005926264449954033, + 0.004375408869236708, + 0.004188473802059889, + 0.005605307407677174, + 0.004061178304255009, + 0.003359237452968955, + 0.003304623067378998, + 0.0032565740402787924, + 0.004038461484014988, + 0.007878901436924934, + 0.0019236152293160558, + 0.00291048944927752, + 0.009446827694773674, + 0.003376225009560585, + 0.0031555420719087124, + 0.0039682528004050255, + 0.004208848346024752, + 0.004025048576295376, + 0.002925122855231166, + 0.00542270764708519, + 0.006118172779679298, + 0.002810792298987508, + 0.0033041846472769976, + 0.0032396011520177126, + 0.01899525336921215, + 0.0231325663626194, + 0.011753483675420284, + 0.028159549459815025, + 0.024236906319856644, + 0.012658142484724522, + 0.00927488412708044, + 0.01145144086331129, + 0.021036796271800995, + 0.09734790027141571, + 0.008776005357503891, + 0.03667866438627243, + 0.04898996651172638, + 0.01979132555425167, + 0.013180641457438469, + 0.039368659257888794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.002822702983394265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609065569937229 + ], + "opposite_logprob_tokens": [ + -4.1410908699035645 + ], + "logprob_reward": -0.01609065569937229, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609065569937229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08133711665868759, + 0.07925469428300858, + 0.08037230372428894, + 0.020957816392183304, + 0.030933205038309097, + 0.08498222380876541, + 0.2191847413778305, + 0.08136717975139618, + 0.17735092341899872, + 0.14425981044769287 + ], + "credit_map": [ + 0.012648986652493477, + 0.011734395287930965, + 0.011879864148795605, + 0.007923590950667858, + 0.012618434615433216, + 0.007655774243175983, + 0.010805327445268631, + 0.0060707442462444305, + 0.007552629336714745, + 0.014633497223258018, + 0.019134750589728355, + 0.0044012670405209064, + 0.016439886763691902, + 0.007192918099462986, + 0.0049515822902321815, + 0.0049481578171253204, + 0.004146052990108728, + 0.02741500362753868, + 0.00423736497759819, + 0.004094288684427738, + 0.02603224106132984, + 0.005041199736297131, + 0.0060830628499388695, + 0.0033230790868401527, + 0.0022528974805027246, + 0.0031604664400219917, + 0.0018243958475068212, + 0.002769793150946498, + 0.0034468090161681175, + 0.0023559725377708673, + 0.0035088772419840097, + 0.001638601766899228, + 0.003063248237594962, + 0.005843738093972206, + 0.002706078113988042, + 0.0027737764175981283, + 0.0065915631130337715, + 0.0035368045791983604, + 0.004191055428236723, + 0.002226938959211111, + 0.009410648606717587, + 0.012413828633725643, + 0.009492350742220879, + 0.01393035613000393, + 0.011632280424237251, + 0.007724331691861153, + 0.01060212217271328, + 0.009776300750672817, + 0.03107876144349575, + 0.02636181190609932, + 0.013300623744726181, + 0.0518505796790123, + 0.04082842171192169, + 0.012992793694138527, + 0.025647498667240143, + 0.017124248668551445, + 0.008193434216082096, + 0.013147757388651371, + 0.0056985425762832165, + 0.020407145842909813, + 0.01314239576458931, + 0.006158029660582542, + 0.010196052491664886, + 0.004423816222697496, + 0.02650410123169422, + 0.03237427771091461, + 0.015841880813241005, + 0.029148530215024948, + 0.031696002930402756, + 0.015925534069538116, + 0.011798840016126633, + 0.014061752706766129, + 0.00826776959002018, + 0.05712864175438881, + 0.0033637499436736107, + 0.01323653943836689, + 0.03823278471827507, + 0.007853957824409008, + 0.004866705741733313, + 0.011309650726616383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.01609065569937229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly and smoothly rotates around a crystal sphere, keeping it centered within the frame. The sphere, flawless and highly polished to a luminous translucence, rests securely on a dark, almost indistinguishable base, set against a backdrop of pure, profound darkness. The sphere's interior, far from empty, houses a magnificent miniature universe, a landscape of deep indigo and violet, dotted with countless tiny stars that sparkle like diamonds and gently swirling nebulae of fiery pink and gold gas. The sphere itself rotates in a steady, hypnotic rhythm, causing the galaxies and stardust within it to shift, distort, and reshape themselves, as if enacting the evolution of the universe over eons. The only light source for the entire scene comes from the self-luminous universe inside the crystal ball. It emits a soft and mysterious light, casting a changing colorful halo on the smooth curved surface of the crystal ball and spreading into the surrounding darkness, creating a mysterious atmosphere that is deep, tranquil and full of infinite awe.", + "seed": 26165, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_2808_seed_26165_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": -0.01731383812148124, + "vlm_logprob": -0.01731383812148124, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01731383812148124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008689330890774727 + ], + "opposite_logprob_tokens": [ + -6.2586894035339355 + ], + "logprob_reward": -0.008689330890774727, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.008689330890774727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05896260589361191, + 0.060110125690698624, + 0.1372125744819641, + 0.036011893302202225, + 0.06634628027677536, + 0.08817140758037567, + 0.05675198510289192, + 0.039366159588098526, + 0.4138360023498535, + 0.04323095828294754 + ], + "credit_map": [ + 0.01090402714908123, + 0.007495169993489981, + 0.005958431866019964, + 0.008700521662831306, + 0.005536341108381748, + 0.006352558732032776, + 0.0041977958753705025, + 0.009817760437726974, + 0.005653019063174725, + 0.012493366375565529, + 0.004308569245040417, + 0.0057383631356060505, + 0.008307038806378841, + 0.005294442176818848, + 0.005972840823233128, + 0.012342492118477821, + 0.041757240891456604, + 0.007633435074239969, + 0.012304316274821758, + 0.005794574040919542, + 0.005966991651803255, + 0.010525674559175968, + 0.007003325037658215, + 0.04622704163193703, + 0.006009994074702263, + 0.0035254177637398243, + 0.004567289259284735, + 0.004630066454410553, + 0.004366674460470676, + 0.004230510909110308, + 0.0044987197034060955, + 0.004183223005384207, + 0.01808467134833336, + 0.004678410943597555, + 0.005760353058576584, + 0.0047081043012440205, + 0.005492083262652159, + 0.007084260694682598, + 0.006517976988106966, + 0.014020423404872417, + 0.011611015535891056, + 0.012473381124436855, + 0.009242908097803593, + 0.014115996658802032, + 0.01528043020516634, + 0.007598226889967918, + 0.011267635971307755, + 0.00658181868493557, + 0.006831380072981119, + 0.007844564504921436, + 0.006068889517337084, + 0.007080198731273413, + 0.010201968252658844, + 0.004721719305962324, + 0.007775642443448305, + 0.00622762693092227, + 0.007012202870100737, + 0.002807836513966322, + 0.004459579475224018, + 0.0032394223380833864, + 0.0029217128176242113, + 0.007624299731105566, + 0.004916112404316664, + 0.0063849990256130695, + 0.12663979828357697, + 0.017924608662724495, + 0.026827681809663773, + 0.011449906043708324, + 0.017742225900292397, + 0.1285887509584427, + 0.02097819373011589, + 0.0636848509311676, + 0.005657996516674757, + 0.00375185813754797, + 0.0039151450619101524, + 0.0070752911269664764, + 0.0041147563606500626, + 0.005443683359771967, + 0.007721275556832552, + 0.005550954025238752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.008689330890774727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497291311621666 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.05497291311621666, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497291311621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08984649926424026, + 0.10563548654317856, + 0.2935733497142792, + 0.11227303743362427, + 0.02353886514902115, + 0.03287125006318092, + 0.03951632231473923, + 0.07086192816495895, + 0.17674897611141205, + 0.05513425171375275 + ], + "credit_map": [ + 0.020270798355340958, + 0.009058431722223759, + 0.011621464975178242, + 0.009238912723958492, + 0.007402077782899141, + 0.009035110473632812, + 0.008173026144504547, + 0.015046677552163601, + 0.009007632732391357, + 0.024839069694280624, + 0.008739144541323185, + 0.016686828806996346, + 0.014530014246702194, + 0.007421984802931547, + 0.013436106033623219, + 0.010974717326462269, + 0.08942386507987976, + 0.015786239877343178, + 0.02616087719798088, + 0.014983966946601868, + 0.013653875328600407, + 0.02231096290051937, + 0.014623846858739853, + 0.09662973880767822, + 0.01260213553905487, + 0.013268613256514072, + 0.01660829596221447, + 0.01587667316198349, + 0.01478454191237688, + 0.010229433886706829, + 0.018159663304686546, + 0.010743687860667706, + 0.003242108505219221, + 0.002110549947246909, + 0.0025769914500415325, + 0.0027881090063601732, + 0.003071520244702697, + 0.003419798333197832, + 0.0025533426087349653, + 0.0037764469161629677, + 0.004687812179327011, + 0.0028567551635205746, + 0.0032720717135816813, + 0.004054487682878971, + 0.004784510936588049, + 0.004411507863551378, + 0.004592611454427242, + 0.004211495630443096, + 0.004746153950691223, + 0.00498372595757246, + 0.004555780440568924, + 0.006442221812903881, + 0.0069465795531868935, + 0.0030045791063457727, + 0.004734694957733154, + 0.004102590959519148, + 0.011195627972483635, + 0.00460763368755579, + 0.007605907041579485, + 0.005547909997403622, + 0.005076882429420948, + 0.010581869632005692, + 0.008953864686191082, + 0.017292240634560585, + 0.06787242740392685, + 0.008046384900808334, + 0.015094121918082237, + 0.0048987530171871185, + 0.007176750805228949, + 0.054552242159843445, + 0.007555073592811823, + 0.011553242802619934, + 0.009194673970341682, + 0.00432813074439764, + 0.00608218414708972, + 0.005162786692380905, + 0.003549274057149887, + 0.0099746473133564, + 0.0076109603978693485, + 0.009231596253812313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.05497291311621666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067533585242927074 + ], + "opposite_logprob_tokens": [ + -5.006753444671631 + ], + "logprob_reward": -0.0067533585242927074, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067533585242927074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1488892287015915, + 0.0635487511754036, + 0.14595165848731995, + 0.03738224133849144, + 0.15185712277889252, + 0.06526309996843338, + 0.069735586643219, + 0.16139933466911316, + 0.0897747352719307, + 0.0661982074379921 + ], + "credit_map": [ + 0.01828739233314991, + 0.018736451864242554, + 0.01708907261490822, + 0.015316859818994999, + 0.014854967594146729, + 0.025373142212629318, + 0.019833208993077278, + 0.019398139789700508, + 0.006440062541514635, + 0.009620497934520245, + 0.006826081778854132, + 0.009383773431181908, + 0.00946797989308834, + 0.005642206873744726, + 0.006757901515811682, + 0.009410250931978226, + 0.04245877265930176, + 0.007813571952283382, + 0.01349369715899229, + 0.00939441379159689, + 0.007628888823091984, + 0.010498507879674435, + 0.010641909204423428, + 0.04402190446853638, + 0.007677115499973297, + 0.0028808414936065674, + 0.0048060971312224865, + 0.004021697212010622, + 0.0033479081466794014, + 0.00358534324914217, + 0.00395565340295434, + 0.007107585668563843, + 0.037268489599227905, + 0.009937062859535217, + 0.012651391327381134, + 0.01149569358676672, + 0.016143258661031723, + 0.015924740582704544, + 0.014658193103969097, + 0.03377830237150192, + 0.008398644626140594, + 0.009096581488847733, + 0.006340396590530872, + 0.009255613200366497, + 0.010180791839957237, + 0.008405538275837898, + 0.007197571452707052, + 0.0063879587687551975, + 0.007176249288022518, + 0.009276475757360458, + 0.005414260551333427, + 0.011317572556436062, + 0.016286876052618027, + 0.004857144318521023, + 0.010463224723935127, + 0.004943778272718191, + 0.05124754458665848, + 0.009603780694305897, + 0.01625676453113556, + 0.010932446457445621, + 0.011532636359333992, + 0.027657251805067062, + 0.014156944118440151, + 0.02001197636127472, + 0.015656208619475365, + 0.004079446662217379, + 0.007675382308661938, + 0.003909136634320021, + 0.004787032026797533, + 0.023277057334780693, + 0.007544770371168852, + 0.02284570224583149, + 0.005687433294951916, + 0.0053998795337975025, + 0.007351129315793514, + 0.0098369549959898, + 0.007325278129428625, + 0.00847356952726841, + 0.011526566930115223, + 0.010597393848001957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.0067533585242927074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942891743965447 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001942891743965447, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942891743965447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08045299351215363, + 0.11198493093252182, + 0.2824408710002899, + 0.060148343443870544, + 0.0576511025428772, + 0.07854001224040985, + 0.07837090641260147, + 0.16213099658489227, + 0.05744742229580879, + 0.03083237074315548 + ], + "credit_map": [ + 0.010660546831786633, + 0.008919316343963146, + 0.00931592658162117, + 0.009496888145804405, + 0.008033178746700287, + 0.013589022681117058, + 0.009289662353694439, + 0.011148451827466488, + 0.0094674713909626, + 0.018539873883128166, + 0.00883141066879034, + 0.024151217192411423, + 0.015665100887417793, + 0.00907424371689558, + 0.012946718372404575, + 0.013308897614479065, + 0.08747231215238571, + 0.014642549678683281, + 0.026332886889576912, + 0.013368361629545689, + 0.012171696871519089, + 0.021438121795654297, + 0.013316670432686806, + 0.09369824081659317, + 0.009278417564928532, + 0.005338363815099001, + 0.005988897290080786, + 0.009497473016381264, + 0.00697724986821413, + 0.005801893770694733, + 0.010523264296352863, + 0.006742784753441811, + 0.014978286810219288, + 0.004081731662154198, + 0.0063186814077198505, + 0.00563573045656085, + 0.004635033663362265, + 0.005817517172545195, + 0.006523552350699902, + 0.009660567156970501, + 0.010976184159517288, + 0.010131854563951492, + 0.007750051561743021, + 0.012145007960498333, + 0.014144711196422577, + 0.00689358776435256, + 0.011160743422806263, + 0.0053378683514893055, + 0.008768678642809391, + 0.012681095860898495, + 0.007124088704586029, + 0.008740595541894436, + 0.016925692558288574, + 0.006744188256561756, + 0.011633855290710926, + 0.005752714816480875, + 0.05484866350889206, + 0.007994680665433407, + 0.01712120696902275, + 0.009825268760323524, + 0.010762779042124748, + 0.02585885114967823, + 0.012692820280790329, + 0.0230267234146595, + 0.010852084495127201, + 0.002688516164198518, + 0.006224576383829117, + 0.00356326624751091, + 0.0036087241023778915, + 0.010300938971340656, + 0.006415461655706167, + 0.013793854974210262, + 0.00341268302872777, + 0.002905361820012331, + 0.003230065805837512, + 0.004302041605114937, + 0.0024648772086948156, + 0.004634076729416847, + 0.004657893907278776, + 0.005225370638072491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.001942891743965447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210696332156658 + ], + "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": [ + 0.12416662275791168, + 0.10651292651891708, + 0.2481190264225006, + 0.04656635969877243, + 0.08837377279996872, + 0.06496556103229523, + 0.085931695997715, + 0.13371920585632324, + 0.06415978819131851, + 0.0374850369989872 + ], + "credit_map": [ + 0.01982782781124115, + 0.013090434484183788, + 0.016544073820114136, + 0.013745942153036594, + 0.010101777501404285, + 0.016638444736599922, + 0.013364619575440884, + 0.02085351198911667, + 0.013142220675945282, + 0.024562766775488853, + 0.00840468518435955, + 0.011903532780706882, + 0.013850485906004906, + 0.006968360859900713, + 0.009896989911794662, + 0.01778390072286129, + 0.07635950297117233, + 0.01339046098291874, + 0.02197834849357605, + 0.012560529634356499, + 0.011017397977411747, + 0.01883566938340664, + 0.012354839593172073, + 0.08162228763103485, + 0.0074408333748579025, + 0.004936289973556995, + 0.006526775658130646, + 0.007044540252536535, + 0.005376227665692568, + 0.0038321991451084614, + 0.005244818981736898, + 0.0061646816320717335, + 0.024168910458683968, + 0.006322049535810947, + 0.008712979964911938, + 0.005946386139839888, + 0.007263895124197006, + 0.009273518808186054, + 0.008503302931785583, + 0.018182730302214622, + 0.008327258750796318, + 0.00633964641019702, + 0.006838529836386442, + 0.00806264765560627, + 0.010572971776127815, + 0.010074318386614323, + 0.00891477707773447, + 0.005835413932800293, + 0.008891086094081402, + 0.011737699620425701, + 0.0064558954909443855, + 0.010007557459175587, + 0.018826831132173538, + 0.010914784856140614, + 0.01190673653036356, + 0.00719110481441021, + 0.03885577619075775, + 0.007096226327121258, + 0.014043041504919529, + 0.009309687651693821, + 0.011358304880559444, + 0.02256343699991703, + 0.0130943413823843, + 0.017398394644260406, + 0.011879181489348412, + 0.0046879020519554615, + 0.007444359827786684, + 0.004493977874517441, + 0.00579771539196372, + 0.011084973812103271, + 0.005561014637351036, + 0.013210673816502094, + 0.004868599586188793, + 0.0030365525744855404, + 0.00446712039411068, + 0.004764334298670292, + 0.0034325351007282734, + 0.00602080300450325, + 0.006077408790588379, + 0.0048176865093410015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.014210696332156658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly and smoothly rotates around a crystal sphere, keeping it centered within the frame. The sphere, flawless and highly polished to a luminous translucence, rests securely on a dark, almost indistinguishable base, set against a backdrop of pure, profound darkness. The sphere's interior, far from empty, houses a magnificent miniature universe, a landscape of deep indigo and violet, dotted with countless tiny stars that sparkle like diamonds and gently swirling nebulae of fiery pink and gold gas. The sphere itself rotates in a steady, hypnotic rhythm, causing the galaxies and stardust within it to shift, distort, and reshape themselves, as if enacting the evolution of the universe over eons. The only light source for the entire scene comes from the self-luminous universe inside the crystal ball. It emits a soft and mysterious light, casting a changing colorful halo on the smooth curved surface of the crystal ball and spreading into the surrounding darkness, creating a mysterious atmosphere that is deep, tranquil and full of infinite awe.", + "seed": 26166, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_2808_seed_26166_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": -0.0230016706045717, + "vlm_logprob": -0.0230016706045717, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.0230016706045717, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006995944771915674 + ], + "opposite_logprob_tokens": [ + -6.881996154785156 + ], + "logprob_reward": -0.006995944771915674, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.006995944771915674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27021515369415283, + 0.1476837396621704, + 0.0678558349609375, + 0.079713374376297, + 0.058004576712846756, + 0.0981508120894432, + 0.06831879168748856, + 0.07525968551635742, + 0.09133602678775787, + 0.043461933732032776 + ], + "credit_map": [ + 0.02204236388206482, + 0.01852666772902012, + 0.07322245091199875, + 0.018782665953040123, + 0.03246530145406723, + 0.03785531967878342, + 0.024838337674736977, + 0.04248209297657013, + 0.019180061295628548, + 0.01775488629937172, + 0.021297655999660492, + 0.015460485592484474, + 0.013247760012745857, + 0.020871780812740326, + 0.017266584560275078, + 0.022604532539844513, + 0.01188623532652855, + 0.005329770036041737, + 0.007940832525491714, + 0.00825636275112629, + 0.0062526497058570385, + 0.012087099254131317, + 0.0075896079652011395, + 0.008513282984495163, + 0.014479747973382473, + 0.005901630502194166, + 0.007359158247709274, + 0.007951240986585617, + 0.007049292791634798, + 0.012098550796508789, + 0.008037821389734745, + 0.01683594100177288, + 0.00939514022320509, + 0.006707247346639633, + 0.006413255352526903, + 0.007695874199271202, + 0.006264910567551851, + 0.008372996933758259, + 0.005441352259367704, + 0.007713801693171263, + 0.010000194422900677, + 0.009940250776708126, + 0.014073311351239681, + 0.01616419292986393, + 0.01281645055860281, + 0.0155309634283185, + 0.01079573854804039, + 0.008829724043607712, + 0.006421930156648159, + 0.009148825891315937, + 0.006426692940294743, + 0.01343964971601963, + 0.006127375178039074, + 0.010175198316574097, + 0.009672506712377071, + 0.006906613707542419, + 0.010480068624019623, + 0.009805144742131233, + 0.005902788136154413, + 0.01175130158662796, + 0.008976848796010017, + 0.01158473826944828, + 0.010318633168935776, + 0.00644017243757844, + 0.009021253325045109, + 0.010420597158372402, + 0.006880574394017458, + 0.011558200232684612, + 0.013893994502723217, + 0.013770915567874908, + 0.011534975841641426, + 0.014255519956350327, + 0.006479077972471714, + 0.004268241114914417, + 0.004449491389095783, + 0.004815203137695789, + 0.00623273104429245, + 0.006194994784891605, + 0.005443110596388578, + 0.005579087417572737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.006995944771915674 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205679848790169 + ], + "opposite_logprob_tokens": [ + -2.81205677986145 + ], + "logprob_reward": -0.06205679848790169, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205679848790169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18748627603054047, + 0.1974581480026245, + 0.11450010538101196, + 0.07467633485794067, + 0.03578243777155876, + 0.0877227932214737, + 0.09037992358207703, + 0.06297706812620163, + 0.09558571875095367, + 0.05343114957213402 + ], + "credit_map": [ + 0.012737619690597057, + 0.026562916114926338, + 0.028247639536857605, + 0.022260505706071854, + 0.03258413448929787, + 0.022043975070118904, + 0.02355603501200676, + 0.01949344575405121, + 0.016762906685471535, + 0.02228153869509697, + 0.029706837609410286, + 0.019599612802267075, + 0.02738610841333866, + 0.024161366745829582, + 0.029039502143859863, + 0.028520269319415092, + 0.010027837008237839, + 0.015112858265638351, + 0.008672046475112438, + 0.018331125378608704, + 0.011129312217235565, + 0.018953174352645874, + 0.013993931002914906, + 0.018279828131198883, + 0.0060721589252352715, + 0.00840539950877428, + 0.0065569146536290646, + 0.014299462549388409, + 0.010871726088225842, + 0.011878008022904396, + 0.009881624951958656, + 0.006711040157824755, + 0.003986223600804806, + 0.004565719980746508, + 0.004222473129630089, + 0.005505766253918409, + 0.004394963383674622, + 0.005103715229779482, + 0.0037048019003123045, + 0.004298776388168335, + 0.014525366015732288, + 0.009266681037843227, + 0.011731648817658424, + 0.015425892546772957, + 0.006462553516030312, + 0.012939832173287868, + 0.008645783178508282, + 0.008725037798285484, + 0.012894419021904469, + 0.011238504201173782, + 0.010389273054897785, + 0.011143910698592663, + 0.009650628082454205, + 0.015106922015547752, + 0.011736116372048855, + 0.008220150135457516, + 0.006921934895217419, + 0.011342862620949745, + 0.0053443387150764465, + 0.009705410338938236, + 0.006369416136294603, + 0.009782777167856693, + 0.008352326229214668, + 0.0051580085419118404, + 0.009181725792586803, + 0.009808134287595749, + 0.007695462089031935, + 0.013108576647937298, + 0.0125010060146451, + 0.014999307692050934, + 0.018159259110689163, + 0.010132242925465107, + 0.005275927018374205, + 0.005576102063059807, + 0.004509546793997288, + 0.007457329425960779, + 0.008836132474243641, + 0.008308006450533867, + 0.007530366536229849, + 0.005937742535024881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06205679848790169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032043324317783117 + ], + "opposite_logprob_tokens": [ + -5.753204345703125 + ], + "logprob_reward": -0.0032043324317783117, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032043324317783117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2967151999473572, + 0.1701643317937851, + 0.07973790168762207, + 0.023459557443857193, + 0.05006301775574684, + 0.08004502207040787, + 0.028411321341991425, + 0.1593826711177826, + 0.08626924455165863, + 0.02575172670185566 + ], + "credit_map": [ + 0.02209550514817238, + 0.024413498118519783, + 0.08433984965085983, + 0.02623695135116577, + 0.029215555638074875, + 0.042951278388500214, + 0.026849590241909027, + 0.04061301797628403, + 0.033738430589437485, + 0.017207026481628418, + 0.025309352204203606, + 0.014482292346656322, + 0.013347133062779903, + 0.024845445528626442, + 0.015635639429092407, + 0.025599025189876556, + 0.006513505708426237, + 0.013302203267812729, + 0.009497684426605701, + 0.011882830411195755, + 0.007647230289876461, + 0.011748433113098145, + 0.008683447726070881, + 0.01046257745474577, + 0.002668392611667514, + 0.0033402154222130775, + 0.002076792297884822, + 0.0034244994167238474, + 0.002803205046802759, + 0.003625574056059122, + 0.0027348303701728582, + 0.0027860498521476984, + 0.004601362626999617, + 0.00567233981564641, + 0.004359090235084295, + 0.00821920670568943, + 0.008917070925235748, + 0.0073911044746637344, + 0.005602573975920677, + 0.005300272721797228, + 0.007707391399890184, + 0.006866326089948416, + 0.007159378379583359, + 0.020233556628227234, + 0.007557959761470556, + 0.013307429850101471, + 0.008258285000920296, + 0.00895470567047596, + 0.002875617239624262, + 0.004449863452464342, + 0.0032725685741752386, + 0.004001195076853037, + 0.0037129230331629515, + 0.004097587428987026, + 0.0037443137262016535, + 0.002257256768643856, + 0.018642984330654144, + 0.01745944283902645, + 0.011546325869858265, + 0.020333444699645042, + 0.017214732244610786, + 0.02654389850795269, + 0.03197256103157997, + 0.015669306740164757, + 0.007594348397105932, + 0.010296848602592945, + 0.007681667804718018, + 0.01372850313782692, + 0.01169657427817583, + 0.013139020651578903, + 0.010964247398078442, + 0.011168045923113823, + 0.0033407399896532297, + 0.003550371155142784, + 0.0024669670965522528, + 0.003927616868168116, + 0.0036586630158126354, + 0.0035847395192831755, + 0.0027856556698679924, + 0.0024369771126657724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0032043324317783117 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046377950347959995 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.0046377950347959995, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046377950347959995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2366284430027008, + 0.19919545948505402, + 0.11925055831670761, + 0.08243421465158463, + 0.03484134376049042, + 0.09937359392642975, + 0.05967063084244728, + 0.07214513421058655, + 0.0648222342133522, + 0.03163832053542137 + ], + "credit_map": [ + 0.017298873513936996, + 0.016375787556171417, + 0.07195371389389038, + 0.02019534632563591, + 0.019458314403891563, + 0.03766801208257675, + 0.02245495468378067, + 0.031223461031913757, + 0.022790106013417244, + 0.015781255438923836, + 0.023286551237106323, + 0.02356603369116783, + 0.017690161243081093, + 0.02726377360522747, + 0.027643859386444092, + 0.04117374122142792, + 0.012585444375872612, + 0.01590695045888424, + 0.008960413746535778, + 0.01929282210767269, + 0.011703995987772942, + 0.019996436312794685, + 0.012599760666489601, + 0.018204757943749428, + 0.009747057221829891, + 0.0070824227295815945, + 0.008166433311998844, + 0.017108701169490814, + 0.009672954678535461, + 0.013338341377675533, + 0.009829303249716759, + 0.0074890065006911755, + 0.0034975779708474874, + 0.005035515874624252, + 0.0028210245072841644, + 0.006268761586397886, + 0.005200117826461792, + 0.004734482616186142, + 0.0033057203982025385, + 0.003978148568421602, + 0.012286997400224209, + 0.012054439634084702, + 0.009723878465592861, + 0.013860241509974003, + 0.009109562262892723, + 0.01736871525645256, + 0.009964114055037498, + 0.015005651861429214, + 0.006912692915648222, + 0.008861765265464783, + 0.006455004680901766, + 0.011052954941987991, + 0.005857251584529877, + 0.009073976427316666, + 0.00737325893715024, + 0.004083727486431599, + 0.008886267431080341, + 0.00789437722414732, + 0.004471801687031984, + 0.009525500237941742, + 0.006939103361219168, + 0.011716580018401146, + 0.014271329157054424, + 0.00844018254429102, + 0.008757646195590496, + 0.006907080300152302, + 0.006102505140006542, + 0.00884275883436203, + 0.009223729372024536, + 0.009905853308737278, + 0.008174903690814972, + 0.006907758302986622, + 0.0038027786649763584, + 0.003975281026214361, + 0.002910569543018937, + 0.0049535976722836494, + 0.004686922766268253, + 0.004898342769593, + 0.0034775794483721256, + 0.0029332502745091915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0046377950347959995 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811348229646683 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.03811348229646683, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811348229646683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25226646661758423, + 0.14814531803131104, + 0.11472426354885101, + 0.07517743110656738, + 0.03631656616926193, + 0.058783356100320816, + 0.06582845002412796, + 0.1714380979537964, + 0.060501713305711746, + 0.016818325966596603 + ], + "credit_map": [ + 0.021525679156184196, + 0.01959863491356373, + 0.05877918750047684, + 0.026670679450035095, + 0.03135824576020241, + 0.0406578928232193, + 0.024856925010681152, + 0.02881922572851181, + 0.013005748391151428, + 0.014287340454757214, + 0.01905016601085663, + 0.017793431878089905, + 0.014020303264260292, + 0.02161688730120659, + 0.01896497793495655, + 0.029406467452645302, + 0.013833182863891125, + 0.014902576804161072, + 0.018187254667282104, + 0.013389467261731625, + 0.010232735425233841, + 0.01775740087032318, + 0.011780302971601486, + 0.014641355723142624, + 0.012870868667960167, + 0.006439294666051865, + 0.005918130278587341, + 0.00822718720883131, + 0.005387990269809961, + 0.012155980803072453, + 0.008177454583346844, + 0.016000522300601006, + 0.005316835828125477, + 0.003934876527637243, + 0.00401083379983902, + 0.005794019903987646, + 0.004205330274999142, + 0.005697907879948616, + 0.0037941746413707733, + 0.003562587546184659, + 0.007830915041267872, + 0.005800694692879915, + 0.009032047353684902, + 0.008764123544096947, + 0.007110282313078642, + 0.009371444582939148, + 0.00623588589951396, + 0.0046379645355045795, + 0.004469380248337984, + 0.009238941594958305, + 0.005376358050853014, + 0.016301967203617096, + 0.006055030506104231, + 0.010690063238143921, + 0.008236968889832497, + 0.005459737963974476, + 0.01770537905395031, + 0.02355758100748062, + 0.012811468914151192, + 0.028178200125694275, + 0.020299231633543968, + 0.02660716511309147, + 0.0268622487783432, + 0.015416828915476799, + 0.009482192806899548, + 0.005612403620034456, + 0.004989807028323412, + 0.006881935987621546, + 0.008793196640908718, + 0.009362085722386837, + 0.008914449252188206, + 0.006465643644332886, + 0.001498198020271957, + 0.0015985923819243908, + 0.001795029267668724, + 0.002061045030131936, + 0.002946678549051285, + 0.0024594177957624197, + 0.003203494008630514, + 0.0012558696325868368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03811348229646683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera slowly and smoothly rotates around a crystal sphere, keeping it centered within the frame. The sphere, flawless and highly polished to a luminous translucence, rests securely on a dark, almost indistinguishable base, set against a backdrop of pure, profound darkness. The sphere's interior, far from empty, houses a magnificent miniature universe, a landscape of deep indigo and violet, dotted with countless tiny stars that sparkle like diamonds and gently swirling nebulae of fiery pink and gold gas. The sphere itself rotates in a steady, hypnotic rhythm, causing the galaxies and stardust within it to shift, distort, and reshape themselves, as if enacting the evolution of the universe over eons. The only light source for the entire scene comes from the self-luminous universe inside the crystal ball. It emits a soft and mysterious light, casting a changing colorful halo on the smooth curved surface of the crystal ball and spreading into the surrounding darkness, creating a mysterious atmosphere that is deep, tranquil and full of infinite awe.", + "seed": 26167, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000022/group_9/index_2808_seed_26167_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": -0.018832973763346673, + "vlm_logprob": -0.018832973763346673, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.018832973763346673, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007465677335858345 + ], + "opposite_logprob_tokens": [ + -6.507465839385986 + ], + "logprob_reward": -0.007465677335858345, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.007465677335858345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.147299125790596, + 0.17148172855377197, + 0.03691462054848671, + 0.09791035950183868, + 0.04781646281480789, + 0.14425641298294067, + 0.10467834025621414, + 0.11441828310489655, + 0.08337921649217606, + 0.0518454872071743 + ], + "credit_map": [ + 0.022689664736390114, + 0.010260735638439655, + 0.018663661554455757, + 0.014340625144541264, + 0.011727848090231419, + 0.0366976335644722, + 0.015299110673367977, + 0.017619848251342773, + 0.02233000658452511, + 0.005762879271060228, + 0.013022207655012608, + 0.0076676104217767715, + 0.008029739372432232, + 0.06684325635433197, + 0.02958369255065918, + 0.018242331221699715, + 0.00565676111727953, + 0.0035281351301819086, + 0.0036765183322131634, + 0.005123339127749205, + 0.004208786878734827, + 0.006333677563816309, + 0.004232190549373627, + 0.004155208822339773, + 0.014139330945909023, + 0.010989642702043056, + 0.009560535661876202, + 0.01639287732541561, + 0.009454168379306793, + 0.017676467075943947, + 0.008678588084876537, + 0.011018751189112663, + 0.005819171667098999, + 0.0038402308709919453, + 0.00428999587893486, + 0.006847481708973646, + 0.004009672440588474, + 0.006424116902053356, + 0.009286312386393547, + 0.007299480028450489, + 0.016776567324995995, + 0.017207225784659386, + 0.016863396391272545, + 0.018113981932401657, + 0.01187462080270052, + 0.020344475284218788, + 0.021037010475993156, + 0.022039128467440605, + 0.011425988748669624, + 0.00925652589648962, + 0.006950723472982645, + 0.024703562259674072, + 0.008315082639455795, + 0.01101492065936327, + 0.02185697667300701, + 0.011154565960168839, + 0.01031026616692543, + 0.010207005776464939, + 0.013269338756799698, + 0.025334062054753304, + 0.00812963955104351, + 0.014096895232796669, + 0.01298025157302618, + 0.020090825855731964, + 0.010647697374224663, + 0.008017978630959988, + 0.0071936966851353645, + 0.01789690926671028, + 0.007608470972627401, + 0.00755565008148551, + 0.01208870392292738, + 0.01237010769546032, + 0.005423104390501976, + 0.003480806713923812, + 0.004426970146596432, + 0.007368359714746475, + 0.004981841892004013, + 0.005761109758168459, + 0.015691526234149933, + 0.0047117676585912704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.007465677335858345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371562063694 + ], + "opposite_logprob_tokens": [ + -3.4087157249450684 + ], + "logprob_reward": -0.03371562063694, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371562063694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16992813348770142, + 0.3219388425350189, + 0.0787048414349556, + 0.09297193586826324, + 0.03561070188879967, + 0.03915457800030708, + 0.06637148559093475, + 0.06692933291196823, + 0.04678329825401306, + 0.0816069170832634 + ], + "credit_map": [ + 0.023271284997463226, + 0.011702420189976692, + 0.028072288259863853, + 0.01973092183470726, + 0.010690640658140182, + 0.04563124477863312, + 0.014198766089975834, + 0.016630545258522034, + 0.05494050309062004, + 0.012859738431870937, + 0.022625064477324486, + 0.014764580875635147, + 0.015147000551223755, + 0.12906423211097717, + 0.05407876893877983, + 0.018458912149071693, + 0.008769004605710506, + 0.010460748337209225, + 0.010238166898488998, + 0.008080899715423584, + 0.006505624391138554, + 0.010900678113102913, + 0.007236719597131014, + 0.01651298813521862, + 0.011605433188378811, + 0.008608040399849415, + 0.00965691078454256, + 0.01940223202109337, + 0.008317935280501842, + 0.014260021038353443, + 0.010297042317688465, + 0.010824309661984444, + 0.004066067282110453, + 0.002835284685716033, + 0.005148554220795631, + 0.0049352808855473995, + 0.002883208217099309, + 0.004500624258071184, + 0.004879022482782602, + 0.006362654734402895, + 0.00501422630622983, + 0.004420637618750334, + 0.004328770563006401, + 0.0050973184406757355, + 0.0032782957423478365, + 0.00711958110332489, + 0.006345539819449186, + 0.0035502067767083645, + 0.007513342425227165, + 0.006250704172998667, + 0.005755698308348656, + 0.0129722710698843, + 0.005979178473353386, + 0.008893223479390144, + 0.012400513514876366, + 0.006606553681194782, + 0.0070892805233597755, + 0.006499936804175377, + 0.007316905073821545, + 0.0144302723929286, + 0.004305417183786631, + 0.009361212141811848, + 0.012210367247462273, + 0.005715939216315746, + 0.005024157930165529, + 0.004886311013251543, + 0.005241124890744686, + 0.007580338977277279, + 0.00479315547272563, + 0.004317489452660084, + 0.0056343055330216885, + 0.009306409396231174, + 0.009679408743977547, + 0.009293661452829838, + 0.012751982547342777, + 0.012940289452672005, + 0.00889122486114502, + 0.007717290427535772, + 0.011580736376345158, + 0.008752311579883099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.03371562063694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263637285679579 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263637285679579, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263637285679579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07005418837070465, + 0.44462308287620544, + 0.04510192573070526, + 0.10581259429454803, + 0.05593698471784592, + 0.1400027871131897, + 0.042449310421943665, + 0.027789296582341194, + 0.020403336733579636, + 0.047826554626226425 + ], + "credit_map": [ + 0.01019013300538063, + 0.004040455911308527, + 0.012047378346323967, + 0.005541451741009951, + 0.005195197183638811, + 0.02213628776371479, + 0.005067269783467054, + 0.00583601463586092, + 0.06348862498998642, + 0.010644199326634407, + 0.031633973121643066, + 0.016618715599179268, + 0.015471471473574638, + 0.18608678877353668, + 0.09438043087720871, + 0.02629890665411949, + 0.006050984375178814, + 0.004962109960615635, + 0.0036086533218622208, + 0.008691435679793358, + 0.005108131095767021, + 0.0075849140994250774, + 0.004422761034220457, + 0.00467293243855238, + 0.01373728085309267, + 0.009780769236385822, + 0.011165445670485497, + 0.026110853999853134, + 0.010501640848815441, + 0.012373999692499638, + 0.010746932588517666, + 0.011395672336220741, + 0.0060573783703148365, + 0.004750073421746492, + 0.006635329686105251, + 0.009683853946626186, + 0.005158486310392618, + 0.008430189453065395, + 0.008341532200574875, + 0.006880137603729963, + 0.015607834793627262, + 0.01603127270936966, + 0.015314471907913685, + 0.02416430041193962, + 0.012180939316749573, + 0.021698903292417526, + 0.022045651450753212, + 0.012959408573806286, + 0.004479282535612583, + 0.003854513168334961, + 0.006405757274478674, + 0.007093898486346006, + 0.00414102291688323, + 0.0045427544973790646, + 0.00612960709258914, + 0.0058024730533361435, + 0.0027024743612855673, + 0.0024407687596976757, + 0.0024389568716287613, + 0.006051770411431789, + 0.0018989546224474907, + 0.003401403781026602, + 0.004571287892758846, + 0.0042836773209273815, + 0.002002978464588523, + 0.002214630600064993, + 0.002292623044922948, + 0.003824578132480383, + 0.0020483394619077444, + 0.002572618657723069, + 0.002924276515841484, + 0.0025232932530343533, + 0.006193059496581554, + 0.004892944358289242, + 0.0051456959918141365, + 0.007442946080118418, + 0.0034582947846502066, + 0.006246546749025583, + 0.009498867206275463, + 0.004948201589286327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.005263637285679579 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12784594297409058, + 0.35318025946617126, + 0.06704507768154144, + 0.06772005558013916, + 0.08958762884140015, + 0.14618994295597076, + 0.022624991834163666, + 0.04128843545913696, + 0.028896236792206764, + 0.055621471256017685 + ], + "credit_map": [ + 0.017308227717876434, + 0.009331446141004562, + 0.018493808805942535, + 0.015239991247653961, + 0.008622528053820133, + 0.03283923491835594, + 0.015315038152039051, + 0.010695666074752808, + 0.050182778388261795, + 0.008634170517325401, + 0.0249429140239954, + 0.01366046816110611, + 0.012833599001169205, + 0.1454315334558487, + 0.07659438252449036, + 0.020900387316942215, + 0.00798244122415781, + 0.007510278839617968, + 0.005062237847596407, + 0.014319897629320621, + 0.010478364303708076, + 0.007923891767859459, + 0.007433264050632715, + 0.006334696896374226, + 0.007558469660580158, + 0.007166871335357428, + 0.007352902088314295, + 0.012827938422560692, + 0.005189895164221525, + 0.008265958167612553, + 0.007621325086802244, + 0.011736690066754818, + 0.010771761648356915, + 0.007551811169832945, + 0.007067629601806402, + 0.012354089878499508, + 0.008553226478397846, + 0.011211873963475227, + 0.01773975044488907, + 0.014337485656142235, + 0.014727359637618065, + 0.01599075458943844, + 0.016829336062073708, + 0.02175859548151493, + 0.013664060272276402, + 0.021258169785141945, + 0.02065981924533844, + 0.021301835775375366, + 0.0022404021583497524, + 0.001626639161258936, + 0.0013277141842991114, + 0.0035535262431949377, + 0.0029236178379505873, + 0.003988264594227076, + 0.00519876042380929, + 0.0017660671146586537, + 0.0035748854279518127, + 0.003867195453494787, + 0.0035800195764750242, + 0.010146383196115494, + 0.003666034433990717, + 0.00444659823551774, + 0.0077997734770178795, + 0.004207544028759003, + 0.0035857316106557846, + 0.002164003672078252, + 0.002249362412840128, + 0.005839673336595297, + 0.002488205675035715, + 0.003020665142685175, + 0.005522171966731548, + 0.00402641948312521, + 0.007505685556679964, + 0.005419196095317602, + 0.006030910182744265, + 0.008250613696873188, + 0.0037533848080784082, + 0.007040251977741718, + 0.011402344331145287, + 0.006219084374606609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a crystal sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and smoothly rotate around the sphere while keeping it centered?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sphere's interior filled with a miniature universe featuring indigo and violet landscapes, stars, and pink/gold nebulae?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the sphere rotate steadily while its internal galaxies and stardust shift, distort, and reshape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the only light source the self-luminous interior of the sphere, casting a colorful halo on the crystal surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308178275823593 + ], + "opposite_logprob_tokens": [ + -3.168081760406494 + ], + "logprob_reward": -0.04308178275823593, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308178275823593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06535899639129639, + 0.34897181391716003, + 0.03697327896952629, + 0.1294800490140915, + 0.16589350998401642, + 0.21268460154533386, + 0.0038586691953241825, + 0.012964999303221703, + 0.005041484721004963, + 0.01877262443304062 + ], + "credit_map": [ + 0.008318855427205563, + 0.006230518687516451, + 0.007908273488283157, + 0.010402793064713478, + 0.005882650148123503, + 0.01065498311072588, + 0.007740116212517023, + 0.008220810443162918, + 0.051314640790224075, + 0.00842934288084507, + 0.024707544595003128, + 0.01320359855890274, + 0.011951668187975883, + 0.14631493389606476, + 0.07183079421520233, + 0.02121928706765175, + 0.004125426523387432, + 0.004391909576952457, + 0.004330446012318134, + 0.005408554337918758, + 0.005101843737065792, + 0.004398334305733442, + 0.0041664824821054935, + 0.005050280597060919, + 0.01625756174325943, + 0.013641992583870888, + 0.013558275066316128, + 0.02363819256424904, + 0.01130436547100544, + 0.022625373676419258, + 0.011593111790716648, + 0.016861185431480408, + 0.02040369249880314, + 0.01256762258708477, + 0.01198760885745287, + 0.023112738505005836, + 0.014242730103433132, + 0.021440859884023666, + 0.03275294601917267, + 0.02938530407845974, + 0.021616507321596146, + 0.022892974317073822, + 0.02694084122776985, + 0.028921213001012802, + 0.018992479890584946, + 0.032758086919784546, + 0.030011512339115143, + 0.030550986528396606, + 0.00038527476135641336, + 0.00037052264087833464, + 0.0003182868822477758, + 0.0007579003577120602, + 0.0004056208417750895, + 0.00040555649320594966, + 0.0007467618561349809, + 0.00046874533290974796, + 0.00104371621273458, + 0.0010615314822643995, + 0.0009183135698549449, + 0.0027605777140706778, + 0.0008545512682758272, + 0.002790674567222595, + 0.002133134054020047, + 0.0014025005511939526, + 0.0005988971097394824, + 0.0005304847727529705, + 0.00040523192728869617, + 0.0010038560722023249, + 0.0006195864989422262, + 0.0005922835553064942, + 0.0007838608580641448, + 0.0005072838393971324, + 0.002061228035017848, + 0.0017140894196927547, + 0.0017373650334775448, + 0.0031652047764509916, + 0.002439627656713128, + 0.0027059572748839855, + 0.002647100016474724, + 0.002302051754668355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 4 + ], + "selected_reward": -0.04308178275823593 + } + ], + "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/0000055/group_0/index_tensor([4724, 4724])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_tensor([4724, 4724])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6d409977cab53065cf25fa53838fe44b746f9002 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_tensor([4724, 4724])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a fixed, reverent wide-angle shot, a breathtaking scene unfolds, rendered entirely in a delicate watercolor style, with soft, translucent colors gently bleeding into one another. Across the dramatically shifting canvas against a clearing sky, the last remnants of dark, moody storm clouds retreat, making way for an ever-expanding expanse of pale, luminous blue. The Gothic cathedral's majestic spires, constructed of rain-washed gray stone, dominate the lower frame; their intricate buttresses and pointed arches are softened by a painterly aesthetic, their details suggestive rather than sharply defined. A vibrant rainbow gracefully arches across the cathedral, slowly emerging from the humid air. Its unbroken, luminous arc is composed of soft-edged streaks of red, orange, yellow, green, and purple, seemingly painted using a wet-on-wet technique. Ethereal sunlight penetrates the clouds from a low angle, casting a warm golden glow that sweeps across the cathedral's facade, catching the damp stone and creating a shimmering, rain-washed atmosphere imbued with artistic grace and a profound sense of peace.", + "seed": 8573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_4724_seed_8573_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": -0.011334193544462324, + "vlm_logprob": -0.011334193544462324, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.011334193544462324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00292932684533298 + ], + "opposite_logprob_tokens": [ + -6.002929210662842 + ], + "logprob_reward": -0.00292932684533298, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00292932684533298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10063187777996063, + 0.025633178651332855, + 0.0172946285456419, + 0.020594170317053795, + 0.19339166581630707, + 0.06276897341012955, + 0.01834997534751892, + 0.05088656023144722, + 0.05653143301606178, + 0.45391759276390076 + ], + "credit_map": [ + 0.011183974333107471, + 0.013159609399735928, + 0.02449316345155239, + 0.01743326708674431, + 0.006748298183083534, + 0.00578291667625308, + 0.012906931340694427, + 0.008923700079321861, + 0.004336976446211338, + 0.003849874483421445, + 0.003931804560124874, + 0.003550007939338684, + 0.0027550505474209785, + 0.0016204542480409145, + 0.003266616491600871, + 0.0023223916068673134, + 0.0017549321055412292, + 0.002408173866569996, + 0.0031408697832375765, + 0.0021731872111558914, + 0.0029160380363464355, + 0.0015174302970990539, + 0.0018901699222624302, + 0.001493825577199459, + 0.001899127964861691, + 0.0028021358884871006, + 0.0032093569170683622, + 0.0025345641188323498, + 0.002702178666368127, + 0.0028909749817103148, + 0.0015695452457293868, + 0.00298628443852067, + 0.018923591822385788, + 0.0361657552421093, + 0.06343289464712143, + 0.02155241370201111, + 0.020873740315437317, + 0.011079209856688976, + 0.012045475654304028, + 0.009318570606410503, + 0.012224875390529633, + 0.00946404505521059, + 0.010241961106657982, + 0.0065850201062858105, + 0.006974987220019102, + 0.00634601479396224, + 0.00519208237528801, + 0.005739983171224594, + 0.002136858645826578, + 0.0023415025789290667, + 0.0029176243115216494, + 0.002294413046911359, + 0.002397407777607441, + 0.0017271505203098059, + 0.002625673310831189, + 0.0019093439914286137, + 0.00692326296120882, + 0.006601064465939999, + 0.009273766539990902, + 0.003707733005285263, + 0.006706549786031246, + 0.006245858035981655, + 0.005102421622723341, + 0.006325895432382822, + 0.0075268433429300785, + 0.008329362608492374, + 0.012601644732058048, + 0.006539466325193644, + 0.00558182829990983, + 0.007342256605625153, + 0.004008682444691658, + 0.004601346794515848, + 0.034701112657785416, + 0.16429942846298218, + 0.017083751037716866, + 0.015234487131237984, + 0.06915716081857681, + 0.11282989382743835, + 0.00838717631995678, + 0.032224513590335846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.00292932684533298 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023306358605623245 + ], + "opposite_logprob_tokens": [ + -3.773306369781494 + ], + "logprob_reward": -0.023306358605623245, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023306358605623245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1397133618593216, + 0.021322082728147507, + 0.06740051507949829, + 0.026683511212468147, + 0.10362612456083298, + 0.02997204288840294, + 0.020507682114839554, + 0.041318293660879135, + 0.01615196280181408, + 0.5333043932914734 + ], + "credit_map": [ + 0.01441227551549673, + 0.01997780054807663, + 0.021813873201608658, + 0.01859409175813198, + 0.012508835643529892, + 0.008353278040885925, + 0.027719486504793167, + 0.01633371412754059, + 0.0033332903403788805, + 0.0034375127870589495, + 0.00274012447334826, + 0.0028521916829049587, + 0.0020701922476291656, + 0.0021407296881079674, + 0.0024323281832039356, + 0.0023157137911766768, + 0.006859233137220144, + 0.008706759661436081, + 0.010171448811888695, + 0.005632070358842611, + 0.010500635951757431, + 0.009071693755686283, + 0.00813464354723692, + 0.008324029855430126, + 0.002240758389234543, + 0.003505326109007001, + 0.0026602703146636486, + 0.0017801151843741536, + 0.004524776246398687, + 0.004482551943510771, + 0.0022444177884608507, + 0.005245293024927378, + 0.009964392520487309, + 0.018976770341396332, + 0.03357066586613655, + 0.014530966989696026, + 0.009466007351875305, + 0.0063467142172157764, + 0.006032740231603384, + 0.004737870302051306, + 0.00561466533690691, + 0.004948291461914778, + 0.004313365556299686, + 0.004370968323200941, + 0.0032740796450525522, + 0.0028190715238451958, + 0.0023259310983121395, + 0.002305669477209449, + 0.003615452442318201, + 0.0028953549917787313, + 0.0029248239006847143, + 0.00219149817712605, + 0.00173739914316684, + 0.0027152278926223516, + 0.0022544872481375933, + 0.0021734375040978193, + 0.003953343722969294, + 0.005533680319786072, + 0.008837876841425896, + 0.006164387799799442, + 0.004776411224156618, + 0.0046718125231564045, + 0.004229054786264896, + 0.003151727607473731, + 0.0033574476838111877, + 0.00219493405893445, + 0.0026016829069703817, + 0.0015138948801904917, + 0.0018598389578983188, + 0.0017534400103613734, + 0.0016304249875247478, + 0.0012402990832924843, + 0.09037084877490997, + 0.20682783424854279, + 0.029084427282214165, + 0.021964745596051216, + 0.09047310799360275, + 0.05821824073791504, + 0.015258603729307652, + 0.021106619387865067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.023306358605623245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633444853127003 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007633444853127003, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633444853127003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04283076524734497, + 0.020448628813028336, + 0.01746329478919506, + 0.010631880722939968, + 0.023257657885551453, + 0.0497896745800972, + 0.014588465914130211, + 0.0293766800314188, + 0.02547278255224228, + 0.7661401033401489 + ], + "credit_map": [ + 0.004455494694411755, + 0.005241943057626486, + 0.007880118675529957, + 0.006583106704056263, + 0.004171629436314106, + 0.003269178792834282, + 0.006245633587241173, + 0.004983661230653524, + 0.002100984565913677, + 0.0025870895478874445, + 0.003763390937820077, + 0.0032162503339350224, + 0.0022582386154681444, + 0.0016735201934352517, + 0.0029125455766916275, + 0.00193660834338516, + 0.0023407256230711937, + 0.0024710253346711397, + 0.00384505745023489, + 0.0021700223442167044, + 0.0021384472493082285, + 0.0013247102033346891, + 0.0016544332029297948, + 0.0015188747784122825, + 0.001316202455200255, + 0.0014959449181333184, + 0.0022568099666386843, + 0.0017739171162247658, + 0.0012466521002352238, + 0.000835142214782536, + 0.0009256391786038876, + 0.0007815724820829928, + 0.002153136068955064, + 0.0038483606185764074, + 0.006155016366392374, + 0.003001732984557748, + 0.002643350977450609, + 0.0019801612943410873, + 0.0016496255993843079, + 0.0018262728117406368, + 0.006017182022333145, + 0.007025179918855429, + 0.009930086322128773, + 0.007438909728080034, + 0.005072784144431353, + 0.004039299674332142, + 0.005562521051615477, + 0.004703710786998272, + 0.0025126556865870953, + 0.0020658366847783327, + 0.0019074021838605404, + 0.0014965626178309321, + 0.0012336106738075614, + 0.002344754757359624, + 0.0016087351832538843, + 0.001418907311744988, + 0.002461438998579979, + 0.003760027000680566, + 0.0035612655337899923, + 0.0037937441375106573, + 0.005538563244044781, + 0.0030895371455699205, + 0.0045834919437766075, + 0.002588612027466297, + 0.005456377286463976, + 0.003278639167547226, + 0.003233660478144884, + 0.0022550232242792845, + 0.0031171843875199556, + 0.00266962219029665, + 0.0030979160219430923, + 0.002364360960200429, + 0.05490146949887276, + 0.2719462811946869, + 0.027451204136013985, + 0.026315517723560333, + 0.11475024372339249, + 0.19888189435005188, + 0.015209064818918705, + 0.056684426963329315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.007633444853127003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12439823895692825, + 0.041549202054739, + 0.05094136297702789, + 0.061748113483190536, + 0.2947053611278534, + 0.11966327577829361, + 0.06522807478904724, + 0.10968977957963943, + 0.06239393725991249, + 0.06968260556459427 + ], + "credit_map": [ + 0.015593918040394783, + 0.017697816714644432, + 0.024745728820562363, + 0.022172654047608376, + 0.011519343592226505, + 0.007436264306306839, + 0.01645479165017605, + 0.008777725510299206, + 0.004934174474328756, + 0.005426069721579552, + 0.00856619793921709, + 0.004976653028279543, + 0.004904451314359903, + 0.003272906644269824, + 0.005638474132865667, + 0.0038302759639918804, + 0.0057704816572368145, + 0.008709034882485867, + 0.005946233868598938, + 0.004327600821852684, + 0.012847821228206158, + 0.004690999165177345, + 0.004204506985843182, + 0.004444683436304331, + 0.005876639392226934, + 0.009899530559778214, + 0.006942969746887684, + 0.0045143188908696175, + 0.014767471700906754, + 0.005799034144729376, + 0.006517966277897358, + 0.0074301827698946, + 0.027571137994527817, + 0.05489107966423035, + 0.09263244271278381, + 0.030327914282679558, + 0.036445435136556625, + 0.01759718358516693, + 0.0193394236266613, + 0.015900755301117897, + 0.015818746760487556, + 0.016296876594424248, + 0.032017745077610016, + 0.010492052882909775, + 0.011233575642108917, + 0.01368679292500019, + 0.009638841263949871, + 0.010478646494448185, + 0.009743141941726208, + 0.008467643521726131, + 0.013631586916744709, + 0.009472288191318512, + 0.004622843582183123, + 0.005929327104240656, + 0.006444208323955536, + 0.006917037535458803, + 0.011349483393132687, + 0.01684528961777687, + 0.03338154777884483, + 0.012406802736222744, + 0.007285534869879484, + 0.007443779613822699, + 0.009046918712556362, + 0.011930425651371479, + 0.008225196972489357, + 0.008454508148133755, + 0.009016184136271477, + 0.010789974592626095, + 0.005722488276660442, + 0.008790119551122189, + 0.004752458073198795, + 0.0066430093720555305, + 0.0076166363433003426, + 0.009400419890880585, + 0.007738666143268347, + 0.009385033510625362, + 0.009778772480785847, + 0.011736877262592316, + 0.006404647137969732, + 0.007621556054800749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067339395172894 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067339395172894, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067339395172894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08504948019981384, + 0.015088880434632301, + 0.012553418055176735, + 0.031753506511449814, + 0.06199159473180771, + 0.03196699544787407, + 0.014812047593295574, + 0.04042071849107742, + 0.08280715346336365, + 0.6235561966896057 + ], + "credit_map": [ + 0.009702449664473534, + 0.011500591412186623, + 0.020777160301804543, + 0.013445974327623844, + 0.00542038818821311, + 0.004613456316292286, + 0.011429213918745518, + 0.008160239085555077, + 0.0022965013049542904, + 0.0026531179901212454, + 0.0019990536384284496, + 0.001814687391743064, + 0.002050918759778142, + 0.0011857195058837533, + 0.0017547374591231346, + 0.0013341453159227967, + 0.0010187735315412283, + 0.001522629987448454, + 0.0017244513146579266, + 0.001318011898547411, + 0.0017239749431610107, + 0.001993255689740181, + 0.0012609550030902028, + 0.0019913665018975735, + 0.0032294217962771654, + 0.005005738232284784, + 0.0037632030434906483, + 0.0021982339676469564, + 0.006858866196125746, + 0.003546847030520439, + 0.0030826725997030735, + 0.004068526905030012, + 0.005669794976711273, + 0.011537013575434685, + 0.018997129052877426, + 0.0050253416411578655, + 0.00873019639402628, + 0.003926443867385387, + 0.004414870869368315, + 0.0036908036563545465, + 0.006038399878889322, + 0.005282701924443245, + 0.0062560709193348885, + 0.003816544311121106, + 0.0030694014858454466, + 0.002789187477901578, + 0.0021617377642542124, + 0.002552950521931052, + 0.0021181411575526, + 0.0019251659978181124, + 0.00253051589243114, + 0.001770959235727787, + 0.0015177256427705288, + 0.0014717811718583107, + 0.002016072627156973, + 0.0014616857515648007, + 0.0064780255779623985, + 0.006095292512327433, + 0.00869793351739645, + 0.005141518544405699, + 0.004227785859256983, + 0.0035915409680455923, + 0.0035377012100070715, + 0.0026509177405387163, + 0.027319295331835747, + 0.01326526515185833, + 0.007578229531645775, + 0.004521260969340801, + 0.008264529518783092, + 0.009513480588793755, + 0.005916731897741556, + 0.006428366526961327, + 0.043682459741830826, + 0.20169556140899658, + 0.02255396731197834, + 0.020094014704227448, + 0.08132544904947281, + 0.19000650942325592, + 0.011021577753126621, + 0.0531766451895237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0067339395172894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, reverent wide-angle shot, a breathtaking scene unfolds, rendered entirely in a delicate watercolor style, with soft, translucent colors gently bleeding into one another. Across the dramatically shifting canvas against a clearing sky, the last remnants of dark, moody storm clouds retreat, making way for an ever-expanding expanse of pale, luminous blue. The Gothic cathedral's majestic spires, constructed of rain-washed gray stone, dominate the lower frame; their intricate buttresses and pointed arches are softened by a painterly aesthetic, their details suggestive rather than sharply defined. A vibrant rainbow gracefully arches across the cathedral, slowly emerging from the humid air. Its unbroken, luminous arc is composed of soft-edged streaks of red, orange, yellow, green, and purple, seemingly painted using a wet-on-wet technique. Ethereal sunlight penetrates the clouds from a low angle, casting a warm golden glow that sweeps across the cathedral's facade, catching the damp stone and creating a shimmering, rain-washed atmosphere imbued with artistic grace and a profound sense of peace.", + "seed": 8574, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_4724_seed_8574_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": -0.06140588521957398, + "vlm_logprob": -0.06140588521957398, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06140588521957398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2263566553592682 + ], + "opposite_logprob_tokens": [ + -1.6013566255569458 + ], + "logprob_reward": -0.2263566553592682, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2263566553592682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06186023727059364, + 0.12627385556697845, + 0.025784332305192947, + 0.02234242670238018, + 0.02180628664791584, + 0.01819675602018833, + 0.03329557925462723, + 0.038399677723646164, + 0.01863442361354828, + 0.633406400680542 + ], + "credit_map": [ + 0.008392405696213245, + 0.006782826967537403, + 0.008774259127676487, + 0.011065604165196419, + 0.007077081594616175, + 0.007962093688547611, + 0.0072682821191847324, + 0.004537682514637709, + 0.012901892885565758, + 0.04242175817489624, + 0.00469400966539979, + 0.004948454909026623, + 0.014071155339479446, + 0.035641878843307495, + 0.006820925511419773, + 0.004773794207721949, + 0.003004595870152116, + 0.003229194786399603, + 0.003737495979294181, + 0.005262322258204222, + 0.003099741181358695, + 0.0025473188143223524, + 0.0029436154291033745, + 0.001960047986358404, + 0.002327838446944952, + 0.0024697754997760057, + 0.002868613926693797, + 0.002123239217326045, + 0.0021697008050978184, + 0.001957467058673501, + 0.005211411044001579, + 0.0032143788412213326, + 0.0028577097691595554, + 0.003094454063102603, + 0.0028542629443109035, + 0.0023874551989138126, + 0.0025571133010089397, + 0.0032807548996061087, + 0.0029810976702719927, + 0.0017934403149411082, + 0.0018313024193048477, + 0.003336203284561634, + 0.001970227574929595, + 0.002129629720002413, + 0.003285919548943639, + 0.0023593557998538017, + 0.001792824245058, + 0.0014912921469658613, + 0.0032657815609127283, + 0.0038719538133591413, + 0.005862422753125429, + 0.004579001106321812, + 0.005531613249331713, + 0.00403785752132535, + 0.0033452240750193596, + 0.002801725175231695, + 0.0034693293273448944, + 0.0036380912642925978, + 0.00627101818099618, + 0.006074727047234774, + 0.0032622565049678087, + 0.0049537476152181625, + 0.006418366450816393, + 0.004312142264097929, + 0.0018629302503541112, + 0.0027273050509393215, + 0.0026141058187931776, + 0.0025735439267009497, + 0.001949432073161006, + 0.0026858451310545206, + 0.0023684795014560223, + 0.001852782559581101, + 0.06252826750278473, + 0.0930192843079567, + 0.07226013392210007, + 0.05554720386862755, + 0.07779970020055771, + 0.1750064343214035, + 0.05120313540101051, + 0.046042248606681824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.2263566553592682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006747319363057613 + ], + "opposite_logprob_tokens": [ + -5.006747245788574 + ], + "logprob_reward": -0.006747319363057613, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747319363057613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05912516638636589, + 0.20183099806308746, + 0.03404231369495392, + 0.010471119545400143, + 0.03304499015212059, + 0.016515374183654785, + 0.03025973215699196, + 0.02006625570356846, + 0.03269591182470322, + 0.5619481205940247 + ], + "credit_map": [ + 0.009922957979142666, + 0.009590838104486465, + 0.0052515314891934395, + 0.004402512684464455, + 0.013615373522043228, + 0.007028198800981045, + 0.005852405447512865, + 0.0034613539464771748, + 0.024590875953435898, + 0.04370701313018799, + 0.012213441543281078, + 0.01491811778396368, + 0.02620680257678032, + 0.05000196769833565, + 0.018221717327833176, + 0.011971085332334042, + 0.004724286030977964, + 0.005352400708943605, + 0.0034642787650227547, + 0.003635393688455224, + 0.005878156051039696, + 0.005009590182453394, + 0.003210823517292738, + 0.0027673873119056225, + 0.0012705775443464518, + 0.0013902237406000495, + 0.0012525945203378797, + 0.0010575652122497559, + 0.0015189105179160833, + 0.001559599069878459, + 0.001440579304471612, + 0.0009810705669224262, + 0.004005801398307085, + 0.007073611952364445, + 0.0030839848332107067, + 0.003026629565283656, + 0.003998137079179287, + 0.003952601924538612, + 0.0045192912220954895, + 0.003384936833754182, + 0.0014763068174943328, + 0.002535868901759386, + 0.0019300960702821612, + 0.00220078369602561, + 0.0035149483010172844, + 0.0018122667679563165, + 0.0017798342742025852, + 0.0012652694713324308, + 0.0032295126002281904, + 0.0029958058148622513, + 0.004980461206287146, + 0.0052093565464019775, + 0.0036190273240208626, + 0.004761886782944202, + 0.0033504674211144447, + 0.0021132188849151134, + 0.0025384367909282446, + 0.002407431835308671, + 0.002461931901052594, + 0.0027487033512443304, + 0.0021897528786212206, + 0.00378690124489367, + 0.0023072471376508474, + 0.0016258531250059605, + 0.004022514447569847, + 0.003980239853262901, + 0.004224997013807297, + 0.003942710813134909, + 0.003537075826898217, + 0.006653872784227133, + 0.0037668361328542233, + 0.0025676684454083443, + 0.0550280399620533, + 0.08228430896997452, + 0.0639246255159378, + 0.04919655621051788, + 0.06891307234764099, + 0.15672971308231354, + 0.04599441960453987, + 0.039877403527498245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.006747319363057613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15518032014369965, + 0.16888724267482758, + 0.1242513433098793, + 0.06542625278234482, + 0.02727523073554039, + 0.03085249848663807, + 0.025498690083622932, + 0.04362247884273529, + 0.041189175099134445, + 0.31781676411628723 + ], + "credit_map": [ + 0.016910146921873093, + 0.016933130100369453, + 0.025914236903190613, + 0.03216461092233658, + 0.01571609266102314, + 0.020798690617084503, + 0.016650494188070297, + 0.010092925280332565, + 0.01724044233560562, + 0.03250318020582199, + 0.0106017692014575, + 0.01241662260144949, + 0.028453627601265907, + 0.019890975207090378, + 0.031852174550294876, + 0.01592845655977726, + 0.01318802498281002, + 0.013268904760479927, + 0.020019788295030594, + 0.026806285604834557, + 0.01016016025096178, + 0.011641932651400566, + 0.018738195300102234, + 0.010428047738969326, + 0.0069364868104457855, + 0.006994974333792925, + 0.00727476179599762, + 0.006224425043910742, + 0.0057564666494727135, + 0.006488392595201731, + 0.01637493073940277, + 0.009375816211104393, + 0.0028675494249910116, + 0.005141527857631445, + 0.0026268428191542625, + 0.003279215656220913, + 0.0033589121885597706, + 0.0026981912087649107, + 0.003973059356212616, + 0.003329931292682886, + 0.003120690817013383, + 0.004207630176097155, + 0.004069129936397076, + 0.0049710022285580635, + 0.005818188656121492, + 0.003264413448050618, + 0.0027605362702161074, + 0.002640907885506749, + 0.002743961988016963, + 0.002795605920255184, + 0.004870061296969652, + 0.004186697769910097, + 0.003817663760855794, + 0.002668182598426938, + 0.002599106403067708, + 0.0018174081342294812, + 0.0047080256044864655, + 0.0048386091366410255, + 0.006178615614771843, + 0.007263424340635538, + 0.004059211816638708, + 0.005809069611132145, + 0.0069900318048894405, + 0.003775492310523987, + 0.004500627983361483, + 0.005081025417894125, + 0.005427350755780935, + 0.005247322842478752, + 0.004551015328615904, + 0.007630372419953346, + 0.00484702130779624, + 0.0039044402074068785, + 0.030311981216073036, + 0.04774411767721176, + 0.036504779011011124, + 0.029127366840839386, + 0.039224643260240555, + 0.08580904453992844, + 0.025747770443558693, + 0.023347072303295135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.014193889684975147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866306856274605 + ], + "opposite_logprob_tokens": [ + -3.0486631393432617 + ], + "logprob_reward": -0.04866306856274605, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866306856274605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04905787855386734, + 0.04232802614569664, + 0.016763275489211082, + 0.02640487067401409, + 0.00824756920337677, + 0.010005577467381954, + 0.024797599762678146, + 0.02433554269373417, + 0.04022497311234474, + 0.7578347325325012 + ], + "credit_map": [ + 0.007420590613037348, + 0.005946940742433071, + 0.008622778579592705, + 0.006148631684482098, + 0.006292934995144606, + 0.005893651396036148, + 0.005339952185750008, + 0.003392390441149473, + 0.00520141888409853, + 0.014214246533811092, + 0.001975258346647024, + 0.0022866185754537582, + 0.00647841626778245, + 0.0072750523686409, + 0.002976467600092292, + 0.0019205444259569049, + 0.002868823939934373, + 0.0022426603827625513, + 0.0016382737085223198, + 0.0015257487539201975, + 0.003174956887960434, + 0.002345380838960409, + 0.0016631638864055276, + 0.0013042647624388337, + 0.002874724566936493, + 0.003154104808345437, + 0.0027762516401708126, + 0.0023915241472423077, + 0.00509418873116374, + 0.0035386099480092525, + 0.0043590497225522995, + 0.002216413849964738, + 0.0009051894303411245, + 0.001237978576682508, + 0.0008703325875103474, + 0.0009375932277180254, + 0.0011211327509954572, + 0.0012646918185055256, + 0.0011627002386376262, + 0.0007479492924176157, + 0.0009072759421542287, + 0.001441231812350452, + 0.0011822737287729979, + 0.0009989072568714619, + 0.0019072307040914893, + 0.0011052739573642612, + 0.0012940330198034644, + 0.0011693509295582771, + 0.0024517299607396126, + 0.0025081189814954996, + 0.004599607549607754, + 0.005055822432041168, + 0.0029679567087441683, + 0.0019966992549598217, + 0.002643831307068467, + 0.0025738312397152185, + 0.0020699508022516966, + 0.0027257383335381746, + 0.0031341956928372383, + 0.0035170181654393673, + 0.002144839847460389, + 0.003829521592706442, + 0.004154775757342577, + 0.002759500639513135, + 0.005024475045502186, + 0.0057478961534798145, + 0.004446434322744608, + 0.004013256169855595, + 0.00468857167288661, + 0.00966580305248499, + 0.0037469863891601562, + 0.002891545882448554, + 0.07315076887607574, + 0.11177016794681549, + 0.08708794414997101, + 0.06823514401912689, + 0.09334662556648254, + 0.2074461728334427, + 0.061396580189466476, + 0.055401209741830826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04866306856274605 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068493127822876 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.011068493127822876, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068493127822876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0686895027756691, + 0.10897188633680344, + 0.058987732976675034, + 0.09482327848672867, + 0.03405608981847763, + 0.046958331018686295, + 0.047852300107479095, + 0.049034394323825836, + 0.0657128393650055, + 0.4249136745929718 + ], + "credit_map": [ + 0.007726900279521942, + 0.008920498192310333, + 0.00636716652661562, + 0.012491531670093536, + 0.010233629494905472, + 0.009193326346576214, + 0.009570538066327572, + 0.004185912199318409, + 0.015334313735365868, + 0.039966728538274765, + 0.004280213266611099, + 0.004652971401810646, + 0.01637362316250801, + 0.019148701801896095, + 0.0055573177523911, + 0.003658024361357093, + 0.006549788173288107, + 0.0057858251966536045, + 0.009393790736794472, + 0.011155483312904835, + 0.005773197393864393, + 0.006245694123208523, + 0.009212732315063477, + 0.004871224518865347, + 0.010695432312786579, + 0.008662765845656395, + 0.016681700944900513, + 0.010380856692790985, + 0.011564253829419613, + 0.009262328036129475, + 0.01719553768634796, + 0.0103803975507617, + 0.0041388762183487415, + 0.007955435663461685, + 0.002511441707611084, + 0.003659257898107171, + 0.004827827215194702, + 0.004551395773887634, + 0.00391458859667182, + 0.002497268607839942, + 0.005327822174876928, + 0.008612303994596004, + 0.005218925420194864, + 0.0063595036044716835, + 0.009418390691280365, + 0.004102345556020737, + 0.004816216416656971, + 0.0031028243247419596, + 0.0048918346874415874, + 0.004931081086397171, + 0.007480987813323736, + 0.008308014832437038, + 0.005697120912373066, + 0.008186504244804382, + 0.004764209035784006, + 0.0035925456322729588, + 0.00440002279356122, + 0.007038353011012077, + 0.00829531904309988, + 0.006947869900614023, + 0.004984128754585981, + 0.007822114042937756, + 0.006244043819606304, + 0.003302544355392456, + 0.007908094674348831, + 0.010101060383021832, + 0.00827036052942276, + 0.006347923539578915, + 0.0077268504537642, + 0.013120397925376892, + 0.0067862155847251415, + 0.0054519325494766235, + 0.04133767634630203, + 0.06381100416183472, + 0.048819635063409805, + 0.036124736070632935, + 0.05172352492809296, + 0.11563300341367722, + 0.03571077063679695, + 0.031753335148096085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.011068493127822876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, reverent wide-angle shot, a breathtaking scene unfolds, rendered entirely in a delicate watercolor style, with soft, translucent colors gently bleeding into one another. Across the dramatically shifting canvas against a clearing sky, the last remnants of dark, moody storm clouds retreat, making way for an ever-expanding expanse of pale, luminous blue. The Gothic cathedral's majestic spires, constructed of rain-washed gray stone, dominate the lower frame; their intricate buttresses and pointed arches are softened by a painterly aesthetic, their details suggestive rather than sharply defined. A vibrant rainbow gracefully arches across the cathedral, slowly emerging from the humid air. Its unbroken, luminous arc is composed of soft-edged streaks of red, orange, yellow, green, and purple, seemingly painted using a wet-on-wet technique. Ethereal sunlight penetrates the clouds from a low angle, casting a warm golden glow that sweeps across the cathedral's facade, catching the damp stone and creating a shimmering, rain-washed atmosphere imbued with artistic grace and a profound sense of peace.", + "seed": 8575, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_4724_seed_8575_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": -0.12483280580490827, + "vlm_logprob": -0.12483280580490827, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.12483280580490827, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3139651119709015 + ], + "opposite_logprob_tokens": [ + -1.313965082168579 + ], + "logprob_reward": -0.3139651119709015, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3139651119709015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14638440310955048, + 0.03647427260875702, + 0.05807659402489662, + 0.049154363572597504, + 0.08287175744771957, + 0.2688331604003906, + 0.10161624103784561, + 0.2056904286146164, + 0.027434172108769417, + 0.023464642465114594 + ], + "credit_map": [ + 0.012693013064563274, + 0.017039701342582703, + 0.020830389112234116, + 0.022665653377771378, + 0.029757268726825714, + 0.02032150886952877, + 0.009562484920024872, + 0.0135143818333745, + 0.003965646959841251, + 0.004486787598580122, + 0.005801744759082794, + 0.00497598759829998, + 0.006337231490761042, + 0.00424466747790575, + 0.0031458477023988962, + 0.0035163583233952522, + 0.006596717983484268, + 0.006432025693356991, + 0.008531689643859863, + 0.008143104612827301, + 0.010732606053352356, + 0.006795250345021486, + 0.005539615638554096, + 0.005305584520101547, + 0.005054315086454153, + 0.0053625646978616714, + 0.004572119098156691, + 0.00998381432145834, + 0.007043438032269478, + 0.005812307354062796, + 0.004479133524000645, + 0.006846671923995018, + 0.011916100047528744, + 0.016619747504591942, + 0.011145208962261677, + 0.009586463682353497, + 0.01106863934546709, + 0.007862158119678497, + 0.007791904732584953, + 0.006881527602672577, + 0.05682005360722542, + 0.061355944722890854, + 0.02712881751358509, + 0.031389251351356506, + 0.025720151141285896, + 0.02677873894572258, + 0.026529788970947266, + 0.01311041135340929, + 0.01438805740326643, + 0.015621274709701538, + 0.011977408081293106, + 0.016317252069711685, + 0.009298590011894703, + 0.008540671318769455, + 0.015251796692609787, + 0.010221191681921482, + 0.024051375687122345, + 0.01984311081469059, + 0.02830834686756134, + 0.038622453808784485, + 0.017342589795589447, + 0.017493443563580513, + 0.03374602645635605, + 0.02628307417035103, + 0.002655153628438711, + 0.002950067166239023, + 0.003551092464476824, + 0.004401168320327997, + 0.0038791063707321882, + 0.002490038052201271, + 0.003450962482020259, + 0.004056584555655718, + 0.002461726078763604, + 0.002214582171291113, + 0.002379884710535407, + 0.004804766271263361, + 0.0031999654602259398, + 0.002041493309661746, + 0.0039539942517876625, + 0.0024082320742309093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.3139651119709015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25215262174606323 + ], + "opposite_logprob_tokens": [ + -1.502152681350708 + ], + "logprob_reward": -0.25215262174606323, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25215262174606323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1627112329006195, + 0.08170963078737259, + 0.05906195938587189, + 0.057718515396118164, + 0.09718658775091171, + 0.15074416995048523, + 0.051675308495759964, + 0.2368863970041275, + 0.03652871400117874, + 0.06577751785516739 + ], + "credit_map": [ + 0.022939829155802727, + 0.018554572016000748, + 0.025113478302955627, + 0.0244954414665699, + 0.023049846291542053, + 0.019283119589090347, + 0.013319220393896103, + 0.015955720096826553, + 0.011248284950852394, + 0.012503915466368198, + 0.00888941902667284, + 0.01629834622144699, + 0.010438621044158936, + 0.009772226214408875, + 0.004811667371541262, + 0.007747138850390911, + 0.008876807056367397, + 0.0082229133695364, + 0.011365504935383797, + 0.007296922616660595, + 0.006817061919718981, + 0.005642469506710768, + 0.005008078180253506, + 0.005832194816321135, + 0.007560893893241882, + 0.0066196853294968605, + 0.01362881064414978, + 0.005971819628030062, + 0.009897341951727867, + 0.005717255175113678, + 0.004054767079651356, + 0.0042679389007389545, + 0.013487511314451694, + 0.019352680072188377, + 0.01109089981764555, + 0.011397646740078926, + 0.015000933781266212, + 0.010241392999887466, + 0.009167088195681572, + 0.007448425516486168, + 0.03170301020145416, + 0.03324463218450546, + 0.015612385235726833, + 0.020245498046278954, + 0.01375372614711523, + 0.014649368822574615, + 0.01356300339102745, + 0.007972534745931625, + 0.005656059365719557, + 0.005509987939149141, + 0.007338005118072033, + 0.010387683287262917, + 0.005589983891695738, + 0.004665177781134844, + 0.006083682179450989, + 0.006444723345339298, + 0.026388900354504585, + 0.02298380434513092, + 0.036398183554410934, + 0.03878973051905632, + 0.022054031491279602, + 0.01931525208055973, + 0.03931459039449692, + 0.0316418819129467, + 0.003472238313406706, + 0.004018251784145832, + 0.005037254653871059, + 0.00586981326341629, + 0.00436194334179163, + 0.0030813533812761307, + 0.006555493455380201, + 0.004132361616939306, + 0.00897842738777399, + 0.008139961399137974, + 0.017346397042274475, + 0.006338822655379772, + 0.00607301015406847, + 0.00414663041010499, + 0.010383215732872486, + 0.004371047019958496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.25215262174606323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601458847522736 + ], + "opposite_logprob_tokens": [ + -3.895601511001587 + ], + "logprob_reward": -0.020601458847522736, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601458847522736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48510244488716125, + 0.025661103427410126, + 0.08744019269943237, + 0.043332647532224655, + 0.028825776651501656, + 0.09333625435829163, + 0.08722461760044098, + 0.09454231709241867, + 0.03785375505685806, + 0.01668083481490612 + ], + "credit_map": [ + 0.08650965988636017, + 0.13543398678302765, + 0.07049734145402908, + 0.047204673290252686, + 0.055354904383420944, + 0.04523521661758423, + 0.023729214444756508, + 0.021137475967407227, + 0.002447302918881178, + 0.0021292350720614195, + 0.004772229585796595, + 0.00529592577368021, + 0.0029677506536245346, + 0.0027528884820640087, + 0.002325083827599883, + 0.0029706850182265043, + 0.007808535359799862, + 0.008608099073171616, + 0.012832818552851677, + 0.012900307774543762, + 0.02161364071071148, + 0.01129950676113367, + 0.0062166862189769745, + 0.00616059172898531, + 0.003953883890062571, + 0.005003019236028194, + 0.00432930514216423, + 0.005728389136493206, + 0.011233475059270859, + 0.006409152876585722, + 0.0038398494943976402, + 0.002835575258359313, + 0.0037215084303170443, + 0.0035795089788734913, + 0.004176089074462652, + 0.004464104305952787, + 0.003897316288203001, + 0.0030412222258746624, + 0.0033649138640612364, + 0.0025811127852648497, + 0.01046132855117321, + 0.013362000696361065, + 0.010437474586069584, + 0.01255445834249258, + 0.010751200839877129, + 0.008339311927556992, + 0.015895532444119453, + 0.011534951627254486, + 0.013487646356225014, + 0.014718937687575817, + 0.010429253801703453, + 0.013863095082342625, + 0.010572310537099838, + 0.007567614782601595, + 0.008830573409795761, + 0.007755185943096876, + 0.0119687644764781, + 0.010396216064691544, + 0.010386665351688862, + 0.01661105826497078, + 0.010644570924341679, + 0.008557736873626709, + 0.01169783715158701, + 0.014279473572969437, + 0.004011374432593584, + 0.0038595946971327066, + 0.005216423887759447, + 0.006345450412482023, + 0.004249333869665861, + 0.003240186721086502, + 0.006415280047804117, + 0.00451611028984189, + 0.0019556921906769276, + 0.0019459998002275825, + 0.002718404633924365, + 0.001958987209945917, + 0.002343710046261549, + 0.001783918123692274, + 0.002428806619718671, + 0.0015453153755515814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.020601458847522736 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029816951602697372 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.029816951602697372, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029816951602697372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.250254362821579, + 0.03332604467868805, + 0.06630812585353851, + 0.10271080583333969, + 0.07670523226261139, + 0.1264534741640091, + 0.1436702460050583, + 0.06519440561532974, + 0.058745212852954865, + 0.07663208991289139 + ], + "credit_map": [ + 0.035354066640138626, + 0.06257207691669464, + 0.03199255093932152, + 0.03350796177983284, + 0.034159623086452484, + 0.026569923385977745, + 0.012386525049805641, + 0.013711640611290932, + 0.003336819587275386, + 0.0044919634237885475, + 0.0030249101109802723, + 0.0035780000034719706, + 0.008608155883848667, + 0.00493332464247942, + 0.002626728732138872, + 0.002726141829043627, + 0.006557567045092583, + 0.009314355440437794, + 0.01019632164388895, + 0.010755756869912148, + 0.011989602819085121, + 0.007773872930556536, + 0.0048690009862184525, + 0.004851649980992079, + 0.00904912780970335, + 0.013183859176933765, + 0.008313635364174843, + 0.01328680757433176, + 0.028728656470775604, + 0.01607571542263031, + 0.007403056137263775, + 0.006669945083558559, + 0.00836670957505703, + 0.00869304221123457, + 0.011855464428663254, + 0.012762228958308697, + 0.010534921661019325, + 0.007696835789829493, + 0.008958016522228718, + 0.007838007062673569, + 0.024397676810622215, + 0.026010507717728615, + 0.011897770687937737, + 0.015520178712904453, + 0.013501943089067936, + 0.012449996545910835, + 0.014037887565791607, + 0.008637499064207077, + 0.02387278899550438, + 0.025225559249520302, + 0.016916563734412193, + 0.023344362154603004, + 0.014653251506388187, + 0.012079530395567417, + 0.014725148677825928, + 0.01285304594784975, + 0.007528626825660467, + 0.006153776776045561, + 0.009696275927126408, + 0.012016075663268566, + 0.005429835990071297, + 0.0050440337508916855, + 0.011664526537060738, + 0.007661257404834032, + 0.005460296757519245, + 0.005437916144728661, + 0.007760745007544756, + 0.010704398155212402, + 0.006539993919432163, + 0.004738033749163151, + 0.010109491646289825, + 0.007994337007403374, + 0.006908172741532326, + 0.005997523665428162, + 0.007368803024291992, + 0.014089982025325298, + 0.010821962729096413, + 0.006357411853969097, + 0.017009515315294266, + 0.008078716695308685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.029816951602697372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627884857356548 + ], + "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": [ + 0.38639944791793823, + 0.15131503343582153, + 0.10212179273366928, + 0.04607441648840904, + 0.04624522104859352, + 0.09962619096040726, + 0.030091267079114914, + 0.07237899303436279, + 0.03388448432087898, + 0.03186311572790146 + ], + "credit_map": [ + 0.05816330015659332, + 0.10372627526521683, + 0.05483899638056755, + 0.04724540561437607, + 0.04699292778968811, + 0.03889889270067215, + 0.017797978594899178, + 0.018735678866505623, + 0.01712956465780735, + 0.02818603627383709, + 0.01987408846616745, + 0.01360983308404684, + 0.03499801829457283, + 0.01974821649491787, + 0.008922836743295193, + 0.008846437558531761, + 0.016084933653473854, + 0.011769319884479046, + 0.031082026660442352, + 0.008586965501308441, + 0.012498031370341778, + 0.00935439858585596, + 0.005740154068917036, + 0.007005964405834675, + 0.00504651153460145, + 0.005798795260488987, + 0.006835957523435354, + 0.0071480716578662395, + 0.006430353969335556, + 0.005092125851660967, + 0.0045878398232162, + 0.005134761333465576, + 0.006886262446641922, + 0.008320986293256283, + 0.006181955337524414, + 0.006047191098332405, + 0.0059686158783733845, + 0.004449599422514439, + 0.00500944908708334, + 0.0033811628818511963, + 0.021007314324378967, + 0.02329026535153389, + 0.011196011677384377, + 0.0103177884593606, + 0.009207431226968765, + 0.009855760261416435, + 0.009747082367539406, + 0.005004541482776403, + 0.003609792795032263, + 0.0029819239862263203, + 0.00355291529558599, + 0.004026574548333883, + 0.002172237727791071, + 0.0028707135934382677, + 0.005241068545728922, + 0.005636040586978197, + 0.007966557517647743, + 0.007242843043059111, + 0.00997085589915514, + 0.013078710995614529, + 0.006649447605013847, + 0.006127036176621914, + 0.01264810748398304, + 0.008695428259670734, + 0.002699529519304633, + 0.002528774319216609, + 0.003840215038508177, + 0.005754805635660887, + 0.0039020373951643705, + 0.002928330795839429, + 0.0070517705753445625, + 0.00517902011051774, + 0.003450196236371994, + 0.003322714241221547, + 0.004200848285108805, + 0.004843478556722403, + 0.004649442154914141, + 0.0031610906589776278, + 0.004782132338732481, + 0.003453214652836323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.007627884857356548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a fixed, reverent wide-angle shot, a breathtaking scene unfolds, rendered entirely in a delicate watercolor style, with soft, translucent colors gently bleeding into one another. Across the dramatically shifting canvas against a clearing sky, the last remnants of dark, moody storm clouds retreat, making way for an ever-expanding expanse of pale, luminous blue. The Gothic cathedral's majestic spires, constructed of rain-washed gray stone, dominate the lower frame; their intricate buttresses and pointed arches are softened by a painterly aesthetic, their details suggestive rather than sharply defined. A vibrant rainbow gracefully arches across the cathedral, slowly emerging from the humid air. Its unbroken, luminous arc is composed of soft-edged streaks of red, orange, yellow, green, and purple, seemingly painted using a wet-on-wet technique. Ethereal sunlight penetrates the clouds from a low angle, casting a warm golden glow that sweeps across the cathedral's facade, catching the damp stone and creating a shimmering, rain-washed atmosphere imbued with artistic grace and a profound sense of peace.", + "seed": 8576, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_0/index_4724_seed_8576_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": -0.05393229676410556, + "vlm_logprob": -0.05393229676410556, + "vlm_yes_no_margin": 3.399999976158142, + "vlm_token_credit": -0.05393229676410556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14329297840595245 + ], + "opposite_logprob_tokens": [ + -2.0182929039001465 + ], + "logprob_reward": -0.14329297840595245, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14329297840595245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2044297307729721, + 0.037891000509262085, + 0.051779501140117645, + 0.058502454310655594, + 0.03221612796187401, + 0.08955195546150208, + 0.1348806917667389, + 0.2025410383939743, + 0.06581753492355347, + 0.12238996475934982 + ], + "credit_map": [ + 0.019110284745693207, + 0.020456790924072266, + 0.036404676735401154, + 0.023553254082798958, + 0.027039356529712677, + 0.02298206463456154, + 0.020646680146455765, + 0.03423662111163139, + 0.00493013858795166, + 0.004083483945578337, + 0.007569636218249798, + 0.003878709627315402, + 0.004784847609698772, + 0.004887603223323822, + 0.0033373215701431036, + 0.004419262055307627, + 0.008377610705792904, + 0.007803839631378651, + 0.008503975346684456, + 0.005373065825551748, + 0.006725026294589043, + 0.005028443410992622, + 0.005272780545055866, + 0.004694757983088493, + 0.014872661791741848, + 0.007737390697002411, + 0.010006751865148544, + 0.005781480111181736, + 0.005464279092848301, + 0.006933385040611029, + 0.0038240852300077677, + 0.0038824202492833138, + 0.005079085938632488, + 0.004739020485430956, + 0.003774910466745496, + 0.003468150272965431, + 0.004557396750897169, + 0.004185399040579796, + 0.002693077316507697, + 0.003719090251252055, + 0.015729419887065887, + 0.015234106220304966, + 0.010471212677657604, + 0.009440952911973, + 0.013462006114423275, + 0.011653143912553787, + 0.005856179632246494, + 0.007704938296228647, + 0.02113565430045128, + 0.02108074724674225, + 0.017325688153505325, + 0.01586494781076908, + 0.024310603737831116, + 0.01396927610039711, + 0.010251285508275032, + 0.010942494496703148, + 0.029047297313809395, + 0.027098026126623154, + 0.02534872107207775, + 0.027053669095039368, + 0.031693585216999054, + 0.026272457093000412, + 0.017563000321388245, + 0.018464282155036926, + 0.014828464016318321, + 0.010644540190696716, + 0.009112445637583733, + 0.006483857054263353, + 0.00945010781288147, + 0.006817564833909273, + 0.004092568997293711, + 0.00438799150288105, + 0.03391589969396591, + 0.01853398233652115, + 0.01726529560983181, + 0.013586832210421562, + 0.012687001377344131, + 0.012024926953017712, + 0.007007790729403496, + 0.007368237245827913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.14329297840595245 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003860175609589 + ], + "opposite_logprob_tokens": [ + -2.6950385570526123 + ], + "logprob_reward": -0.07003860175609589, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003860175609589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16231442987918854, + 0.11156728863716125, + 0.15971702337265015, + 0.0739409402012825, + 0.07011851668357849, + 0.052182603627443314, + 0.06069640442728996, + 0.054810378700494766, + 0.10989771038293839, + 0.14475470781326294 + ], + "credit_map": [ + 0.01679723896086216, + 0.02042069472372532, + 0.032668955624103546, + 0.016392294317483902, + 0.011266998015344143, + 0.016235586255788803, + 0.01641104370355606, + 0.03212161734700203, + 0.013583964668214321, + 0.012476377189159393, + 0.022963015362620354, + 0.010441992431879044, + 0.01124387327581644, + 0.010928003117442131, + 0.008093534037470818, + 0.021836524829268456, + 0.024404557421803474, + 0.04630642756819725, + 0.026998162269592285, + 0.01080403197556734, + 0.01427795086055994, + 0.016643613576889038, + 0.009110847488045692, + 0.011171442456543446, + 0.01652022823691368, + 0.010710111819207668, + 0.010197456926107407, + 0.0068807220086455345, + 0.008270198479294777, + 0.008874522522091866, + 0.005618347320705652, + 0.006869350094348192, + 0.008195656351745129, + 0.01376750972121954, + 0.00712550338357687, + 0.006740042474120855, + 0.007526725064963102, + 0.009761858731508255, + 0.006770663894712925, + 0.01023055985569954, + 0.010372812859714031, + 0.005988821852952242, + 0.007002119440585375, + 0.0060687558725476265, + 0.005692190956324339, + 0.004678328987210989, + 0.005308549385517836, + 0.007071023341268301, + 0.008774037472903728, + 0.009965711273252964, + 0.00773889385163784, + 0.0067610302940011024, + 0.009900986216962337, + 0.005986558739095926, + 0.005644142162054777, + 0.005925043486058712, + 0.009888784028589725, + 0.009062228724360466, + 0.006846768781542778, + 0.006084978114813566, + 0.006133164279162884, + 0.006825611926615238, + 0.005405668634921312, + 0.004563176538795233, + 0.020379146561026573, + 0.0155238788574934, + 0.014087766408920288, + 0.01142493449151516, + 0.017020242288708687, + 0.014928688295185566, + 0.008282308466732502, + 0.008250747807323933, + 0.040694281458854675, + 0.02633928693830967, + 0.020726436749100685, + 0.014137713238596916, + 0.015476716682314873, + 0.012828086502850056, + 0.007284102961421013, + 0.007268084213137627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.07003860175609589 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583124831318855 + ], + "opposite_logprob_tokens": [ + -3.895583152770996 + ], + "logprob_reward": -0.020583124831318855, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583124831318855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23829025030136108, + 0.11933810263872147, + 0.11061319708824158, + 0.09143935143947601, + 0.12342719733715057, + 0.13300497829914093, + 0.06604712456464767, + 0.03485576808452606, + 0.03208842873573303, + 0.05089566484093666 + ], + "credit_map": [ + 0.017022069543600082, + 0.027070458978414536, + 0.027361011132597923, + 0.02551460452377796, + 0.025780219584703445, + 0.022131597623229027, + 0.03920462355017662, + 0.05420565605163574, + 0.008223269134759903, + 0.014415706507861614, + 0.009561405517160892, + 0.014936125837266445, + 0.010154576040804386, + 0.009625785052776337, + 0.02071341685950756, + 0.03170781210064888, + 0.013253583572804928, + 0.012049593962728977, + 0.018965687602758408, + 0.013666174374520779, + 0.0158909410238266, + 0.007334921509027481, + 0.0200013667345047, + 0.009450915269553661, + 0.012525269761681557, + 0.012601812370121479, + 0.018467867746949196, + 0.009595670737326145, + 0.00978020578622818, + 0.00853344239294529, + 0.01233762502670288, + 0.007597445975989103, + 0.014362744055688381, + 0.019209573045372963, + 0.011171798221766949, + 0.01138636190444231, + 0.021063312888145447, + 0.014177709817886353, + 0.016231132671236992, + 0.015824561938643456, + 0.028841795399785042, + 0.01846984028816223, + 0.01577971689403057, + 0.01249835267663002, + 0.018123699352145195, + 0.010678541846573353, + 0.014611436054110527, + 0.014001586474478245, + 0.008488772436976433, + 0.008094173856079578, + 0.007993502542376518, + 0.00854396726936102, + 0.01228959672152996, + 0.005847596097737551, + 0.008334114216268063, + 0.006455401889979839, + 0.004662604071199894, + 0.004215158522129059, + 0.004187897313386202, + 0.004414171911776066, + 0.004808630328625441, + 0.003953411243855953, + 0.004301613196730614, + 0.004312279634177685, + 0.007477287668734789, + 0.004843014292418957, + 0.0041610547341406345, + 0.003072042716667056, + 0.004068085923790932, + 0.0033682326320558786, + 0.0032376213930547237, + 0.0018610861152410507, + 0.014889520592987537, + 0.008212858811020851, + 0.00693281926214695, + 0.005231963470578194, + 0.00504338787868619, + 0.0044207340106368065, + 0.0030701514333486557, + 0.00309422193095088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.020583124831318855 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980329655110836 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.02980329655110836, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980329655110836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11082801222801208, + 0.04876425489783287, + 0.03774026408791542, + 0.09550353139638901, + 0.12361672520637512, + 0.12789000570774078, + 0.09385029971599579, + 0.21919812262058258, + 0.03158048540353775, + 0.1110282838344574 + ], + "credit_map": [ + 0.010932235047221184, + 0.011251135729253292, + 0.020922953262925148, + 0.009373054839670658, + 0.014745654538273811, + 0.012594352476298809, + 0.011050824075937271, + 0.01995779573917389, + 0.005640299990773201, + 0.00616320688277483, + 0.00828314945101738, + 0.0043227458372712135, + 0.0036564876791089773, + 0.004467316437512636, + 0.005132573191076517, + 0.011098473332822323, + 0.005095365457236767, + 0.00532618397846818, + 0.006008016411215067, + 0.0035371712874621153, + 0.005273868329823017, + 0.004314500372856855, + 0.00510419113561511, + 0.0030809680465608835, + 0.016422100365161896, + 0.01696069911122322, + 0.013103929348289967, + 0.006918830331414938, + 0.010593267157673836, + 0.011648895218968391, + 0.007164133246988058, + 0.012691672891378403, + 0.029770510271191597, + 0.017347237095236778, + 0.014086208306252956, + 0.011377299204468727, + 0.017015991732478142, + 0.015713561326265335, + 0.007612050045281649, + 0.01069386675953865, + 0.016746504232287407, + 0.022391509264707565, + 0.014583349227905273, + 0.016369596123695374, + 0.01931164599955082, + 0.017933806404471397, + 0.00947743933647871, + 0.01107616163790226, + 0.018501801416277885, + 0.015357209369540215, + 0.012739522382616997, + 0.010605145245790482, + 0.012932712212204933, + 0.0077916523441672325, + 0.008422606624662876, + 0.0074996487237513065, + 0.026228010654449463, + 0.025982562452554703, + 0.02885456196963787, + 0.031659357249736786, + 0.04075101390480995, + 0.024715155363082886, + 0.02118723839521408, + 0.019820228219032288, + 0.005009116139262915, + 0.005139555782079697, + 0.004219192545861006, + 0.0039428602904081345, + 0.004155866801738739, + 0.003463116707280278, + 0.002868188312277198, + 0.002782587194815278, + 0.02523631788790226, + 0.022329075261950493, + 0.01608257368206978, + 0.010529684834182262, + 0.012881513684988022, + 0.011249013245105743, + 0.00646947743371129, + 0.006250630132853985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.02980329655110836 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a delicate watercolor style with soft, translucent colors bleeding into one another?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, moody storm clouds visible retreating across a clearing sky to reveal pale, luminous blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a Gothic cathedral with rain-washed gray stone spires, buttresses, and pointed arches visible in the lower frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a vibrant rainbow with soft-edged streaks of red, orange, yellow, green, and purple arching across the cathedral?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a fixed, reverent wide-angle view with ethereal sunlight casting a warm golden glow from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059434822760522366 + ], + "opposite_logprob_tokens": [ + -5.130943298339844 + ], + "logprob_reward": -0.0059434822760522366, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059434822760522366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12249136716127396, + 0.03264159336686134, + 0.1033729687333107, + 0.05999092012643814, + 0.0519925020635128, + 0.06438259780406952, + 0.06031594052910805, + 0.15543648600578308, + 0.15964335203170776, + 0.18973220884799957 + ], + "credit_map": [ + 0.013537121005356312, + 0.012272820807993412, + 0.014948242343962193, + 0.010224011726677418, + 0.011811048723757267, + 0.012850326485931873, + 0.010314338840544224, + 0.03653346374630928, + 0.0034122569486498833, + 0.004536812659353018, + 0.004047326743602753, + 0.003590018954128027, + 0.003042286727577448, + 0.003252406371757388, + 0.004734331741929054, + 0.0060261571779847145, + 0.013483594171702862, + 0.029866496101021767, + 0.022728361189365387, + 0.007256696466356516, + 0.009108644910156727, + 0.00893168430775404, + 0.006598412524908781, + 0.005399088840931654, + 0.013418270274996758, + 0.00739221042022109, + 0.008267886936664581, + 0.006125983782112598, + 0.005255230702459812, + 0.0067914691753685474, + 0.00559623260051012, + 0.007143640425056219, + 0.009892861358821392, + 0.006703887600451708, + 0.006114274263381958, + 0.005076180212199688, + 0.007558501325547695, + 0.007039247080683708, + 0.0043936967849731445, + 0.005213860422372818, + 0.011769005097448826, + 0.008529742248356342, + 0.00807475671172142, + 0.00728711998090148, + 0.007569713983684778, + 0.007371692452579737, + 0.006657059770077467, + 0.0071235219947993755, + 0.008300308138132095, + 0.009548231028020382, + 0.00782687496393919, + 0.007458745036274195, + 0.00908742193132639, + 0.00546914990991354, + 0.005491527263075113, + 0.007133689243346453, + 0.019997594878077507, + 0.01958681270480156, + 0.019646717235445976, + 0.023963885381817818, + 0.023273414000868797, + 0.01475608628243208, + 0.017589343711733818, + 0.016622649505734444, + 0.038619980216026306, + 0.025728445500135422, + 0.02208893746137619, + 0.015019867569208145, + 0.020394228398799896, + 0.013414070010185242, + 0.011212019249796867, + 0.013165803626179695, + 0.039963316172361374, + 0.04053613543510437, + 0.02645585872232914, + 0.020658649504184723, + 0.02435300126671791, + 0.01584045961499214, + 0.011523190885782242, + 0.010401614010334015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0059434822760522366 + } + ], + "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/0000055/group_1/index_tensor([3886, 3886])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_tensor([3886, 3886])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ac5a674ace0ec191c69140859bcc0a7f63fbae85 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_tensor([3886, 3886])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium close-up depicts a young man with rich olive skin and intense, dark eyes, set against a dimly lit, out-of-focus interior. He is dramatically illuminated by a single, harsh key light from the side, sculpting his features with high contrast and casting deep shadows that heighten the sense of tension. He suddenly and violently recoils, as if struck by an unseen force. His head jerks back, his shoulders recoil sharply, and his upper body instinctively twists away from the camera. His previously focused eyes are now wide with pure, unadulterated disbelief, his brow furrowed in confusion and shock. His lips part slightly in a silent, shocked gasp, his entire posture frozen in instinctive reaction, creating an atmosphere of intense tension and drama of a sudden, devastating revelation.", + "seed": 31941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31941_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": -0.4166620999574661, + "vlm_logprob": -0.4166620999574661, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.4166620999574661, + "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 medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021864555776119232 + ], + "opposite_logprob_tokens": [ + -4.896864414215088 + ], + "logprob_reward": -0.021864555776119232, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.021864555776119232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12437276542186737, + 0.08712384104728699, + 0.0876106396317482, + 0.05940529331564903, + 0.08640933036804199, + 0.07181650400161743, + 0.06158545985817909, + 0.24060842394828796, + 0.05818486958742142, + 0.12288283556699753 + ], + "credit_map": [ + 0.011621102690696716, + 0.01407604943960905, + 0.019078535959124565, + 0.02369273081421852, + 0.016728941351175308, + 0.011900177225470543, + 0.015113593079149723, + 0.01216164045035839, + 0.006374324671924114, + 0.008977184072136879, + 0.01304231584072113, + 0.013633673079311848, + 0.0077207377180457115, + 0.011240974068641663, + 0.015215755440294743, + 0.010918883606791496, + 0.011722582392394543, + 0.009737466461956501, + 0.009595971554517746, + 0.008049353957176208, + 0.007669790182262659, + 0.009236285462975502, + 0.009038110263645649, + 0.022561082616448402, + 0.0063700503669679165, + 0.005288258660584688, + 0.006934719160199165, + 0.00934121198952198, + 0.0075712609104812145, + 0.006904765963554382, + 0.008330268785357475, + 0.00866476446390152, + 0.007610278204083443, + 0.012505907565355301, + 0.014874241314828396, + 0.009336264804005623, + 0.012134169228374958, + 0.014954134821891785, + 0.007964638993144035, + 0.0070297010242938995, + 0.006177349481731653, + 0.007323083467781544, + 0.005743698682636023, + 0.01047512423247099, + 0.009416407905519009, + 0.00876869261264801, + 0.016013894230127335, + 0.007898259907960892, + 0.006957858335226774, + 0.006651108153164387, + 0.008236407302320004, + 0.008029325865209103, + 0.007203841116279364, + 0.007263142615556717, + 0.006841261405497789, + 0.010402525775134563, + 0.017299562692642212, + 0.012006223201751709, + 0.03048642724752426, + 0.017510496079921722, + 0.011352800764143467, + 0.08450651168823242, + 0.00878711137920618, + 0.058659303933382034, + 0.006640344858169556, + 0.008301041088998318, + 0.007489732000976801, + 0.01018610130995512, + 0.005389229394495487, + 0.006005620118230581, + 0.005102889612317085, + 0.009069913998246193, + 0.012825311161577702, + 0.015637461096048355, + 0.02870112471282482, + 0.02188641019165516, + 0.009995457716286182, + 0.011626731604337692, + 0.008526483550667763, + 0.013683862052857876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.021864555776119232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023315560072660446 + ], + "opposite_logprob_tokens": [ + -3.773315668106079 + ], + "logprob_reward": -0.023315560072660446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023315560072660446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10322166979312897, + 0.10432572662830353, + 0.0847531333565712, + 0.16872040927410126, + 0.18431630730628967, + 0.0334927923977375, + 0.03671005368232727, + 0.12497478723526001, + 0.0838063508272171, + 0.0756787434220314 + ], + "credit_map": [ + 0.01172672025859356, + 0.012854031287133694, + 0.014930718578398228, + 0.01711057499051094, + 0.013865464366972446, + 0.010736552067101002, + 0.010864019393920898, + 0.011133597232401371, + 0.008074040524661541, + 0.012925969436764717, + 0.012954345904290676, + 0.013691219501197338, + 0.00959410984069109, + 0.010868692770600319, + 0.01607060432434082, + 0.020146753638982773, + 0.009207670576870441, + 0.01079933438450098, + 0.014019410125911236, + 0.008844524621963501, + 0.007920837961137295, + 0.008063527755439281, + 0.008826002478599548, + 0.017071835696697235, + 0.017329230904579163, + 0.02069810777902603, + 0.024564730003476143, + 0.03969616815447807, + 0.017787596210837364, + 0.017751891165971756, + 0.016652323305606842, + 0.014240378513932228, + 0.014766260050237179, + 0.026783693581819534, + 0.03301815316081047, + 0.022477200254797935, + 0.026434287428855896, + 0.025444388389587402, + 0.01831391453742981, + 0.017078405246138573, + 0.0037632144521921873, + 0.003236746182665229, + 0.004207920283079147, + 0.004034203477203846, + 0.0038859350606799126, + 0.004888087045401335, + 0.005202890373766422, + 0.004273796919733286, + 0.004908409435302019, + 0.002803493058308959, + 0.005507863126695156, + 0.004965418484061956, + 0.003388708457350731, + 0.004375939257442951, + 0.004798988811671734, + 0.005961231887340546, + 0.005721186753362417, + 0.005294562317430973, + 0.009687702171504498, + 0.0065325964242219925, + 0.005358362104743719, + 0.04834466800093651, + 0.0042502377182245255, + 0.039785485714673996, + 0.008166216313838959, + 0.013223174959421158, + 0.010225190781056881, + 0.021873347461223602, + 0.006442870479077101, + 0.006710105109959841, + 0.007491478696465492, + 0.009673976339399815, + 0.005935803987085819, + 0.00619953079149127, + 0.02116980589926243, + 0.010379860177636147, + 0.005097843706607819, + 0.014049281366169453, + 0.005831233225762844, + 0.007015384268015623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.023315560072660446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.910355567932129 + ], + "opposite_logprob_tokens": [ + -0.1603556126356125 + ], + "logprob_reward": -1.910355567932129, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.910355567932129, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06845103204250336, + 0.11796405166387558, + 0.10255745053291321, + 0.25881338119506836, + 0.07593754678964615, + 0.04509676620364189, + 0.038121555000543594, + 0.15823210775852203, + 0.10121192783117294, + 0.033614251762628555 + ], + "credit_map": [ + 0.008039040490984917, + 0.008871723897755146, + 0.00818382017314434, + 0.010734844952821732, + 0.010303249582648277, + 0.0071686687879264355, + 0.008747657760977745, + 0.006402012892067432, + 0.010019281879067421, + 0.014773019589483738, + 0.016834497451782227, + 0.019120046868920326, + 0.012979450635612011, + 0.013276474550366402, + 0.01694253273308277, + 0.014018728397786617, + 0.013288768008351326, + 0.01964801549911499, + 0.016460848972201347, + 0.013449988327920437, + 0.012255096808075905, + 0.008809897117316723, + 0.008338636718690395, + 0.010306183248758316, + 0.022859513759613037, + 0.030192220583558083, + 0.030820634216070175, + 0.04303845390677452, + 0.03704717010259628, + 0.030827144160866737, + 0.039260946214199066, + 0.024767255410552025, + 0.003924067132174969, + 0.008114914409816265, + 0.009436991065740585, + 0.010443762876093388, + 0.012098484672605991, + 0.011782549321651459, + 0.009387499652802944, + 0.010749259032309055, + 0.004057054873555899, + 0.004591070581227541, + 0.005207782611250877, + 0.005988205783069134, + 0.006104593630880117, + 0.007141341920942068, + 0.006376530975103378, + 0.005630176514387131, + 0.006460549309849739, + 0.003866673680022359, + 0.004258368629962206, + 0.005850233603268862, + 0.003633785294368863, + 0.004377121105790138, + 0.004112962633371353, + 0.005561852362006903, + 0.012446660548448563, + 0.01498119905591011, + 0.014816087670624256, + 0.030766066163778305, + 0.012206423096358776, + 0.02846328541636467, + 0.015683192759752274, + 0.028869161382317543, + 0.00834811944514513, + 0.011130286380648613, + 0.013011486269533634, + 0.01339139323681593, + 0.012889813631772995, + 0.010420986451208591, + 0.014139690436422825, + 0.01788012869656086, + 0.0038809701800346375, + 0.0033863079734146595, + 0.008843746036291122, + 0.00595821812748909, + 0.0027616280131042004, + 0.003009829670190811, + 0.0033080605790019035, + 0.0024654832668602467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -1.910355567932129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906996458768845 + ], + "opposite_logprob_tokens": [ + -2.5790698528289795 + ], + "logprob_reward": -0.07906996458768845, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906996458768845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08002054691314697, + 0.09092123806476593, + 0.0860535204410553, + 0.2146243005990982, + 0.03930412232875824, + 0.04396922513842583, + 0.04178384691476822, + 0.2689071297645569, + 0.07455933094024658, + 0.059856731444597244 + ], + "credit_map": [ + 0.009126865305006504, + 0.010434566996991634, + 0.011603305116295815, + 0.016361888498067856, + 0.011226462200284004, + 0.0066427746787667274, + 0.008315014652907848, + 0.006309662014245987, + 0.00785009190440178, + 0.010084243491292, + 0.014266664162278175, + 0.011830756440758705, + 0.009107505902647972, + 0.010270397178828716, + 0.015583949163556099, + 0.011927634477615356, + 0.009400464594364166, + 0.015324472449719906, + 0.016244426369667053, + 0.011803247965872288, + 0.008631731383502483, + 0.007648486644029617, + 0.008943098597228527, + 0.00805759709328413, + 0.018515536561608315, + 0.019400792196393013, + 0.029471006244421005, + 0.030523136258125305, + 0.03604978695511818, + 0.027716873213648796, + 0.029918236657977104, + 0.023028917610645294, + 0.0036249689292162657, + 0.004953350406140089, + 0.004534847103059292, + 0.004508635960519314, + 0.006024476606398821, + 0.0056355795823037624, + 0.005018515046685934, + 0.0050037456676363945, + 0.003709956770762801, + 0.004487748257815838, + 0.004778106231242418, + 0.006450997199863195, + 0.00637784181162715, + 0.005281148012727499, + 0.006993986200541258, + 0.005889439955353737, + 0.007221532985568047, + 0.004307137336581945, + 0.0047655836679041386, + 0.006043478846549988, + 0.004228977486491203, + 0.004719187971204519, + 0.004196775145828724, + 0.006301173474639654, + 0.016830774024128914, + 0.013761168345808983, + 0.020083360373973846, + 0.01839384250342846, + 0.010086838155984879, + 0.09475598484277725, + 0.012171955779194832, + 0.08282322436571121, + 0.0060631767846643925, + 0.007498159538954496, + 0.009350973181426525, + 0.009207873605191708, + 0.00974718201905489, + 0.008293230086565018, + 0.010977148078382015, + 0.013421591371297836, + 0.003828286426141858, + 0.00493336608633399, + 0.011102374643087387, + 0.011555973440408707, + 0.00548721756786108, + 0.01122069638222456, + 0.005363305099308491, + 0.006365509703755379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.07906996458768845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487048514187336 + ], + "opposite_logprob_tokens": [ + -3.0487048625946045 + ], + "logprob_reward": -0.0487048514187336, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487048514187336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06457884609699249, + 0.08898834139108658, + 0.08454825729131699, + 0.20732620358467102, + 0.083722323179245, + 0.05041972175240517, + 0.036186523735523224, + 0.24307581782341003, + 0.08298511058092117, + 0.05816880986094475 + ], + "credit_map": [ + 0.01080532930791378, + 0.008153234608471394, + 0.010005245916545391, + 0.009371007792651653, + 0.006783176679164171, + 0.007340445648878813, + 0.0061736335046589375, + 0.005946773570030928, + 0.007649808190762997, + 0.008009742014110088, + 0.015566970221698284, + 0.011407813988626003, + 0.007166580762714148, + 0.012643909081816673, + 0.015353898517787457, + 0.01118962187319994, + 0.009198269806802273, + 0.014705372974276543, + 0.010711390525102615, + 0.009033533744513988, + 0.007901335135102272, + 0.00913825910538435, + 0.00800536759197712, + 0.01585472747683525, + 0.018782716244459152, + 0.020178088918328285, + 0.025133227929472923, + 0.028836799785494804, + 0.035187236964702606, + 0.026279279962182045, + 0.03211648017168045, + 0.02081238105893135, + 0.005253627896308899, + 0.007710444275289774, + 0.012405987828969955, + 0.009614312089979649, + 0.015458246693015099, + 0.01685844361782074, + 0.008595450781285763, + 0.007825815118849277, + 0.0043524703942239285, + 0.00504628662019968, + 0.0037667532451450825, + 0.0056124720722436905, + 0.007734501268714666, + 0.008237848058342934, + 0.009313703514635563, + 0.006355690769851208, + 0.003575756913051009, + 0.0040227132849395275, + 0.005314321257174015, + 0.004931387957185507, + 0.005545718129724264, + 0.004147747531533241, + 0.004326417110860348, + 0.004322463180869818, + 0.011799253523349762, + 0.01146429032087326, + 0.02070293016731739, + 0.012195009738206863, + 0.010143713094294071, + 0.0908622145652771, + 0.01063394732773304, + 0.07527446001768112, + 0.008649581111967564, + 0.010937687009572983, + 0.010981923900544643, + 0.012621797621250153, + 0.010886677540838718, + 0.008470856584608555, + 0.007741301320493221, + 0.012695281766355038, + 0.008703590370714664, + 0.005871382541954517, + 0.014469224028289318, + 0.007962275296449661, + 0.0036890599876642227, + 0.005085664801299572, + 0.004761572927236557, + 0.007626040372997522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0487048514187336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a young man with rich olive skin and intense, dark eyes, set against a dimly lit, out-of-focus interior. He is dramatically illuminated by a single, harsh key light from the side, sculpting his features with high contrast and casting deep shadows that heighten the sense of tension. He suddenly and violently recoils, as if struck by an unseen force. His head jerks back, his shoulders recoil sharply, and his upper body instinctively twists away from the camera. His previously focused eyes are now wide with pure, unadulterated disbelief, his brow furrowed in confusion and shock. His lips part slightly in a silent, shocked gasp, his entire posture frozen in instinctive reaction, creating an atmosphere of intense tension and drama of a sudden, devastating revelation.", + "seed": 31942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31942_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": -0.3270591001957655, + "vlm_logprob": -0.3270591001957655, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.3270591001957655, + "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 medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028287403285503387 + ], + "opposite_logprob_tokens": [ + -4.903287410736084 + ], + "logprob_reward": -0.028287403285503387, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.028287403285503387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30585402250289917, + 0.2130419760942459, + 0.09021865576505661, + 0.06615845113992691, + 0.08406874537467957, + 0.032273177057504654, + 0.07022427767515182, + 0.04600661247968674, + 0.033725298941135406, + 0.05842879042029381 + ], + "credit_map": [ + 0.02545638009905815, + 0.02117353305220604, + 0.03748098388314247, + 0.02086528390645981, + 0.08577007800340652, + 0.06778640300035477, + 0.022879144176840782, + 0.024442214518785477, + 0.017692549154162407, + 0.012848231010138988, + 0.04637070745229721, + 0.015282323583960533, + 0.017054887488484383, + 0.0710282027721405, + 0.017327776178717613, + 0.015437307767570019, + 0.007841451093554497, + 0.008667895570397377, + 0.013871319591999054, + 0.010627328418195248, + 0.011703683994710445, + 0.009782346896827221, + 0.01924636960029602, + 0.008478258736431599, + 0.005485160741955042, + 0.008992359042167664, + 0.011354482732713223, + 0.007940319366753101, + 0.01011546142399311, + 0.007643774151802063, + 0.008978043682873249, + 0.0056488486006855965, + 0.006046113092452288, + 0.011744299903512001, + 0.010245108976960182, + 0.013289525173604488, + 0.0069426544941961765, + 0.012613242492079735, + 0.013347569853067398, + 0.009840228594839573, + 0.002724048448726535, + 0.004199657589197159, + 0.0050484552048146725, + 0.005048815626651049, + 0.0035544082056730986, + 0.0038431466091424227, + 0.004900088533759117, + 0.00295455614104867, + 0.004638493526726961, + 0.012666795402765274, + 0.005231295246630907, + 0.019538288936018944, + 0.00536657078191638, + 0.011731156148016453, + 0.006109988782554865, + 0.004941691644489765, + 0.0026240390725433826, + 0.0035571958869695663, + 0.0025322737637907267, + 0.002877394203096628, + 0.009801426902413368, + 0.015357453376054764, + 0.0035443780943751335, + 0.005712450016289949, + 0.0029874672181904316, + 0.0031749664340168238, + 0.0033616628497838974, + 0.005493352189660072, + 0.0043771578930318356, + 0.004933328367769718, + 0.003670893842354417, + 0.005726472940295935, + 0.004748516716063023, + 0.010359849780797958, + 0.004656778182834387, + 0.00848589837551117, + 0.008040158078074455, + 0.012378419749438763, + 0.004351112060248852, + 0.005408058874309063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.028287403285503387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026375919580459595 + ], + "opposite_logprob_tokens": [ + -3.6513760089874268 + ], + "logprob_reward": -0.026375919580459595, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026375919580459595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16527770459651947, + 0.10836338251829147, + 0.09890202432870865, + 0.0662606880068779, + 0.3114922046661377, + 0.09302220493555069, + 0.022780457511544228, + 0.01701621152460575, + 0.049300964921712875, + 0.06758414953947067 + ], + "credit_map": [ + 0.01451453473418951, + 0.022265546023845673, + 0.023694908246397972, + 0.017513707280158997, + 0.025675630196928978, + 0.025903943926095963, + 0.016929734498262405, + 0.018779689446091652, + 0.014553497545421124, + 0.010503524914383888, + 0.01824377104640007, + 0.011172189377248287, + 0.011742036789655685, + 0.019822504371404648, + 0.011767113581299782, + 0.01055873092263937, + 0.009010066278278828, + 0.009925305843353271, + 0.01631571166217327, + 0.008998000994324684, + 0.010137936100363731, + 0.01760292612016201, + 0.011216942220926285, + 0.01569513790309429, + 0.006843095179647207, + 0.006706638727337122, + 0.00839559081941843, + 0.006010774523019791, + 0.0094580864533782, + 0.00684369308874011, + 0.011074457317590714, + 0.010928357020020485, + 0.026333730667829514, + 0.03252303972840309, + 0.03913957625627518, + 0.04857897385954857, + 0.02449660561978817, + 0.04685555398464203, + 0.04517202079296112, + 0.04839271306991577, + 0.010349585674703121, + 0.005352713167667389, + 0.03359643369913101, + 0.006774670910090208, + 0.008735733106732368, + 0.015417586080729961, + 0.005910275503993034, + 0.006885210983455181, + 0.0018591622356325388, + 0.003706448944285512, + 0.0020957060623914003, + 0.005642857402563095, + 0.001982898684218526, + 0.003157448722049594, + 0.002477878239005804, + 0.0018580571049824357, + 0.0014825583202764392, + 0.002568500116467476, + 0.0013813136611133814, + 0.002182785188779235, + 0.0029597736429423094, + 0.0026811053976416588, + 0.002097370568662882, + 0.0016628039302304387, + 0.005193339195102453, + 0.009348076768219471, + 0.004515809938311577, + 0.009347469545900822, + 0.006582216825336218, + 0.006685767322778702, + 0.003510570852085948, + 0.004117710515856743, + 0.00543469563126564, + 0.0073422328568995, + 0.005104275420308113, + 0.006514993961900473, + 0.007179471664130688, + 0.012584222480654716, + 0.010478700511157513, + 0.012945561669766903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.026375919580459595 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.5021064281463623 + ], + "opposite_logprob_tokens": [ + -0.25210636854171753 + ], + "logprob_reward": -1.5021064281463623, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5021064281463623, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07956834882497787, + 0.24360214173793793, + 0.13972482085227966, + 0.1082254946231842, + 0.056887175887823105, + 0.18983328342437744, + 0.05236891657114029, + 0.039682645350694656, + 0.07157965749502182, + 0.018527522683143616 + ], + "credit_map": [ + 0.008547666482627392, + 0.009042797610163689, + 0.01272708922624588, + 0.008258934132754803, + 0.010423933155834675, + 0.016762185841798782, + 0.006003504153341055, + 0.007802237756550312, + 0.021127285435795784, + 0.016112729907035828, + 0.04928704723715782, + 0.01783018745481968, + 0.02209245227277279, + 0.078385129570961, + 0.019996877759695053, + 0.01877042092382908, + 0.011781161651015282, + 0.010054913349449635, + 0.03142376244068146, + 0.013405593112111092, + 0.01733616553246975, + 0.01787339523434639, + 0.025270448997616768, + 0.012579379603266716, + 0.009441651403903961, + 0.00994256604462862, + 0.025628013536334038, + 0.007498478051275015, + 0.013068453408777714, + 0.013521337881684303, + 0.016304586082696915, + 0.012820412404835224, + 0.004957106895744801, + 0.005992807447910309, + 0.011213657446205616, + 0.008117862977087498, + 0.0048867021687328815, + 0.0082640266045928, + 0.007006559055298567, + 0.0064484537579119205, + 0.01761787384748459, + 0.013935072347521782, + 0.06604580581188202, + 0.01581789180636406, + 0.01979042775928974, + 0.028529418632388115, + 0.016697602346539497, + 0.011399200186133385, + 0.003519010730087757, + 0.008504433557391167, + 0.003937837667763233, + 0.013313010334968567, + 0.004293397068977356, + 0.008047507144510746, + 0.005206082947552204, + 0.005547637585550547, + 0.002704755635932088, + 0.0034533014986664057, + 0.0030023069120943546, + 0.004757789429277182, + 0.006902223918586969, + 0.0098516084253788, + 0.004934210330247879, + 0.004076451528817415, + 0.006669125519692898, + 0.013167992234230042, + 0.008009123615920544, + 0.009509816765785217, + 0.01029362715780735, + 0.011107832193374634, + 0.006468404084444046, + 0.00635373592376709, + 0.0018727848073467612, + 0.002485833829268813, + 0.0016312012448906898, + 0.0031586617697030306, + 0.002271824050694704, + 0.002544415881857276, + 0.002147764666005969, + 0.00241503631696105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -1.5021064281463623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869803786277771 + ], + "opposite_logprob_tokens": [ + -3.0486979484558105 + ], + "logprob_reward": -0.04869803786277771, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869803786277771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12416127324104309, + 0.12327879667282104, + 0.1649491935968399, + 0.1844160556793213, + 0.05369916558265686, + 0.1363433599472046, + 0.04770229011774063, + 0.04948003962635994, + 0.08853647112846375, + 0.02743338793516159 + ], + "credit_map": [ + 0.011419420130550861, + 0.018337901681661606, + 0.018858689814805984, + 0.013253791257739067, + 0.016441186890006065, + 0.02454645372927189, + 0.010943279601633549, + 0.010360543616116047, + 0.008765374310314655, + 0.014814471825957298, + 0.02285647578537464, + 0.017069747671484947, + 0.012377812527120113, + 0.016568560153245926, + 0.02110041119158268, + 0.009725932031869888, + 0.01320026908069849, + 0.01475985161960125, + 0.02851170301437378, + 0.015224067494273186, + 0.022634845227003098, + 0.020850997418165207, + 0.0319862999022007, + 0.01778116077184677, + 0.015851814299821854, + 0.014338740147650242, + 0.042755160480737686, + 0.01451820507645607, + 0.02345653809607029, + 0.022643310949206352, + 0.02750319056212902, + 0.023349083960056305, + 0.004200534895062447, + 0.005554441828280687, + 0.011263885535299778, + 0.008367754518985748, + 0.004757293500006199, + 0.006655976176261902, + 0.007169337477535009, + 0.0057299379259347916, + 0.014233501628041267, + 0.00752918841317296, + 0.05184544250369072, + 0.009305289015173912, + 0.013599633239209652, + 0.02259511500597, + 0.008525311946868896, + 0.008709871210157871, + 0.0029811153654009104, + 0.007155299186706543, + 0.0037860681768506765, + 0.013078466057777405, + 0.0038862319197505713, + 0.007344634272158146, + 0.004760344047099352, + 0.004710127599537373, + 0.0033274113666266203, + 0.004853595048189163, + 0.003656918415799737, + 0.006310129072517157, + 0.008555016480386257, + 0.01267332211136818, + 0.005991898477077484, + 0.004111741669476032, + 0.008008988574147224, + 0.0157521590590477, + 0.009444093331694603, + 0.011379474774003029, + 0.013049706816673279, + 0.014577697962522507, + 0.008640093728899956, + 0.007684257812798023, + 0.0023877767380326986, + 0.0037922735791653395, + 0.0022677958477288485, + 0.005359529983252287, + 0.0039520710706710815, + 0.004635764751583338, + 0.0028748693875968456, + 0.0021633044816553593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.04869803786277771 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982771210372448 + ], + "opposite_logprob_tokens": [ + -3.52982759475708 + ], + "logprob_reward": -0.02982771210372448, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982771210372448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16028453409671783, + 0.360871821641922, + 0.06389233469963074, + 0.1450049728155136, + 0.03930475562810898, + 0.04246805980801582, + 0.03474550321698189, + 0.06497719883918762, + 0.07062014937400818, + 0.01783061772584915 + ], + "credit_map": [ + 0.023792706429958344, + 0.019178543239831924, + 0.01750032790005207, + 0.013845881447196007, + 0.04240655153989792, + 0.018520362675189972, + 0.014267477206885815, + 0.01077270321547985, + 0.0317554734647274, + 0.02202666364610195, + 0.07323137670755386, + 0.028025785461068153, + 0.029633773490786552, + 0.11772051453590393, + 0.030901603400707245, + 0.02757667936384678, + 0.006235224194824696, + 0.006165407598018646, + 0.010100772604346275, + 0.006907887756824493, + 0.007782990112900734, + 0.008255883120000362, + 0.011121831834316254, + 0.007322340738028288, + 0.01241406798362732, + 0.011430589482188225, + 0.030988330021500587, + 0.011633696965873241, + 0.019046472385525703, + 0.01761205680668354, + 0.021993553265929222, + 0.01988622546195984, + 0.0029034181497991085, + 0.0048822504468262196, + 0.006031584460288286, + 0.004434491507709026, + 0.0034406445920467377, + 0.00705587025731802, + 0.005177841987460852, + 0.005378656089305878, + 0.003042925149202347, + 0.005332224071025848, + 0.008044308051466942, + 0.005790550261735916, + 0.004665814805775881, + 0.00546857388690114, + 0.006764228921383619, + 0.0033594341948628426, + 0.002659095684066415, + 0.005402929149568081, + 0.0034086781088262796, + 0.008476858027279377, + 0.0029282814357429743, + 0.004914296790957451, + 0.00357401673682034, + 0.0033813503105193377, + 0.004651641007512808, + 0.003947395831346512, + 0.003241786267608404, + 0.004925362300127745, + 0.013564257882535458, + 0.021295791491866112, + 0.004777830094099045, + 0.008573140017688274, + 0.006222511176019907, + 0.012525530532002449, + 0.007736795581877232, + 0.009075635112822056, + 0.01000128872692585, + 0.011406471952795982, + 0.007468148600310087, + 0.006183771416544914, + 0.0016097880434244871, + 0.0032929009757936, + 0.001999866683036089, + 0.0032900834921747446, + 0.0020894985646009445, + 0.0023889674339443445, + 0.001799015561118722, + 0.0013604984851554036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.02982771210372448 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a young man with rich olive skin and intense, dark eyes, set against a dimly lit, out-of-focus interior. He is dramatically illuminated by a single, harsh key light from the side, sculpting his features with high contrast and casting deep shadows that heighten the sense of tension. He suddenly and violently recoils, as if struck by an unseen force. His head jerks back, his shoulders recoil sharply, and his upper body instinctively twists away from the camera. His previously focused eyes are now wide with pure, unadulterated disbelief, his brow furrowed in confusion and shock. His lips part slightly in a silent, shocked gasp, his entire posture frozen in instinctive reaction, creating an atmosphere of intense tension and drama of a sudden, devastating revelation.", + "seed": 31943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31943_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": -0.3165836252272129, + "vlm_logprob": -0.3165836252272129, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.3165836252272129, + "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 medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017452945932745934 + ], + "opposite_logprob_tokens": [ + -4.892452716827393 + ], + "logprob_reward": -0.017452945932745934, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017452945932745934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6696455478668213, + 0.031917162239551544, + 0.034890398383140564, + 0.061940666288137436, + 0.01985686458647251, + 0.036445312201976776, + 0.019070493057370186, + 0.0958615094423294, + 0.013758032582700253, + 0.016614094376564026 + ], + "credit_map": [ + 0.029574938118457794, + 0.04244528338313103, + 0.02013898268342018, + 0.031074313446879387, + 0.02218717336654663, + 0.28162655234336853, + 0.02198309451341629, + 0.22061516344547272, + 0.0025290236808359623, + 0.003918878734111786, + 0.0029949883464723825, + 0.0048786550760269165, + 0.003965695854276419, + 0.002643544226884842, + 0.004996113944798708, + 0.005990259815007448, + 0.003646966302767396, + 0.004526976030319929, + 0.00909331999719143, + 0.002924871165305376, + 0.004372002091258764, + 0.0031033915001899004, + 0.004827654920518398, + 0.0023952133487910032, + 0.005817175377160311, + 0.008300943300127983, + 0.008581535890698433, + 0.007366925943642855, + 0.012221772223711014, + 0.00607334217056632, + 0.008077518083155155, + 0.005501449108123779, + 0.0017303398344665766, + 0.0029396822210401297, + 0.0020658313296735287, + 0.002327605616301298, + 0.00282882247120142, + 0.002385497558861971, + 0.004084436688572168, + 0.0014946460723876953, + 0.0032341538462787867, + 0.004212929867208004, + 0.0031978441402316093, + 0.005639012902975082, + 0.0039403922855854034, + 0.0058088586665689945, + 0.005336221307516098, + 0.00507589615881443, + 0.0014995456440374255, + 0.0020746635273098946, + 0.0025602017994970083, + 0.0025189172010868788, + 0.0025040837936103344, + 0.0017524866852909327, + 0.002742813667282462, + 0.0034177799243479967, + 0.00793289951980114, + 0.005876452196389437, + 0.027664655819535255, + 0.0046918983571231365, + 0.006125219631940126, + 0.03510995954275131, + 0.003871220862492919, + 0.004589192569255829, + 0.0017193149542436004, + 0.0009920414304360747, + 0.0025830166414380074, + 0.0014823259552940726, + 0.0013922065263614058, + 0.0014858539216220379, + 0.002201308263465762, + 0.001901964540593326, + 0.0018289742292836308, + 0.002196416025981307, + 0.00214434415102005, + 0.0023697633296251297, + 0.003074427368119359, + 0.002148454077541828, + 0.00167393300216645, + 0.0011777810286730528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.017452945932745934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.4634787440299988, + 0.030562661588191986, + 0.04393356293439865, + 0.0723739042878151, + 0.032743167132139206, + 0.02227192185819149, + 0.017847757786512375, + 0.14249223470687866, + 0.12107552587985992, + 0.0532204769551754 + ], + "credit_map": [ + 0.019285986199975014, + 0.029300255700945854, + 0.013709070160984993, + 0.022849442437291145, + 0.016385845839977264, + 0.19607862830162048, + 0.013529244810342789, + 0.15234032273292542, + 0.004079270176589489, + 0.003865990787744522, + 0.003423161106184125, + 0.005962477531284094, + 0.0028629344888031483, + 0.002740089548751712, + 0.003932274878025055, + 0.003696463769301772, + 0.0045030429027974606, + 0.008153007365763187, + 0.009686265140771866, + 0.005318152718245983, + 0.005444306414574385, + 0.0032272604294121265, + 0.0042518023401498795, + 0.0033497249241918325, + 0.007221630308777094, + 0.005720349494367838, + 0.010609693825244904, + 0.004412192385643721, + 0.018820758908987045, + 0.009896748699247837, + 0.00966858770698309, + 0.006023942027240992, + 0.0026148564647883177, + 0.003521444508805871, + 0.003363466588780284, + 0.004144888836890459, + 0.006232594605535269, + 0.003959054592996836, + 0.005191271658986807, + 0.003715590573847294, + 0.0016995485639199615, + 0.0020459007937461138, + 0.001555285300128162, + 0.002109359484165907, + 0.002153713023290038, + 0.0053774211555719376, + 0.0023351721465587616, + 0.004995520226657391, + 0.001183372805826366, + 0.0021653245203197002, + 0.00369128305464983, + 0.002535091945901513, + 0.0013414326822385192, + 0.0026580241974443197, + 0.002085508080199361, + 0.0021877214312553406, + 0.012323111295700073, + 0.006989504676312208, + 0.04454553499817848, + 0.006387380417436361, + 0.007926986552774906, + 0.05212762579321861, + 0.005663279909640551, + 0.00652880547568202, + 0.016291093081235886, + 0.009054383262991905, + 0.03342907875776291, + 0.007605528924614191, + 0.012862255796790123, + 0.017637653276324272, + 0.009980856440961361, + 0.014214678667485714, + 0.005370717961341143, + 0.008838478475809097, + 0.006116754841059446, + 0.006184027995914221, + 0.008194922469556332, + 0.008370057679712772, + 0.005887595936655998, + 0.00425792345777154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "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 shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.4063336849212646 + ], + "opposite_logprob_tokens": [ + -0.28133368492126465 + ], + "logprob_reward": -1.4063336849212646, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063336849212646, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.7246878147125244, + 0.034597303718328476, + 0.03278474509716034, + 0.08796494454145432, + 0.013143589720129967, + 0.06909885257482529, + 0.014295516535639763, + 0.006394614465534687, + 0.011540166102349758, + 0.005492565222084522 + ], + "credit_map": [ + 0.030055467039346695, + 0.04514218866825104, + 0.020941555500030518, + 0.03429479897022247, + 0.024538395926356316, + 0.30949103832244873, + 0.02103688381612301, + 0.23918741941452026, + 0.0024861234705895185, + 0.0040456079877913, + 0.0028046616353094578, + 0.00400324072688818, + 0.004571596160531044, + 0.0032615289092063904, + 0.007771049160510302, + 0.005653491709381342, + 0.002611197531223297, + 0.002956218784675002, + 0.006143093109130859, + 0.0040272134356200695, + 0.005067794118076563, + 0.00404237350448966, + 0.004940863698720932, + 0.0029959885869175196, + 0.009585628286004066, + 0.00880920048803091, + 0.00816901121288538, + 0.007990493439137936, + 0.0180729478597641, + 0.014708105474710464, + 0.011921430937945843, + 0.008708123117685318, + 0.0010291836224496365, + 0.0021308616269379854, + 0.001337702269665897, + 0.0014984911540523171, + 0.0013445049989968538, + 0.0020411957520991564, + 0.001778514706529677, + 0.00198313407599926, + 0.00465366942808032, + 0.0045893192291259766, + 0.005336258094757795, + 0.004357026889920235, + 0.005140199791640043, + 0.02044004015624523, + 0.004667163360863924, + 0.01991516724228859, + 0.0012666404945775867, + 0.0017058219527825713, + 0.001599530573002994, + 0.0024505052715539932, + 0.0016214834759011865, + 0.0014889013255015016, + 0.002270198892802, + 0.0018924339674413204, + 0.0005024938145652413, + 0.00037406967021524906, + 0.0016950418939813972, + 0.00042794147157110274, + 0.0008372887969017029, + 0.0017390488646924496, + 0.00046875388943590224, + 0.00034997554030269384, + 0.0015068664215505123, + 0.0009058996802195907, + 0.0030757016502320766, + 0.000663577753584832, + 0.0012976644793525338, + 0.001554657588712871, + 0.001234441646374762, + 0.0013013556599617004, + 0.0004924170207232237, + 0.0007295040413737297, + 0.0006068964139558375, + 0.0007045104866847396, + 0.0008176927804015577, + 0.0006015382241457701, + 0.000981131219305098, + 0.000558874337002635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -1.4063336849212646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10039852559566498 + ], + "opposite_logprob_tokens": [ + -2.350398540496826 + ], + "logprob_reward": -0.10039852559566498, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10039852559566498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7128922343254089, + 0.037542589008808136, + 0.039046380668878555, + 0.09953296184539795, + 0.012407909147441387, + 0.03379914164543152, + 0.013963575474917889, + 0.009555957280099392, + 0.02938118763267994, + 0.011878005228936672 + ], + "credit_map": [ + 0.029594026505947113, + 0.043999072164297104, + 0.020845012739300728, + 0.033755671232938766, + 0.024415114894509315, + 0.30467289686203003, + 0.02102546952664852, + 0.23458503186702728, + 0.00259734527207911, + 0.003950060345232487, + 0.0029012809973210096, + 0.004756327718496323, + 0.005495743360370398, + 0.003294770373031497, + 0.008678406476974487, + 0.005868657026439905, + 0.003506668144837022, + 0.00347910332493484, + 0.008341713808476925, + 0.003742293454706669, + 0.00611584447324276, + 0.004601228982210159, + 0.006116197444498539, + 0.003143338020890951, + 0.011510574258863926, + 0.008661775849759579, + 0.011426126584410667, + 0.007575727999210358, + 0.023784354329109192, + 0.014697548933327198, + 0.013370241969823837, + 0.008506625890731812, + 0.000953570066485554, + 0.0020725808572024107, + 0.001132529927417636, + 0.001439520507119596, + 0.0016428041271865368, + 0.0020279972814023495, + 0.001665759482420981, + 0.0014731497503817081, + 0.0027570209931582212, + 0.003285340964794159, + 0.0028050465043634176, + 0.004249533638358116, + 0.003231574548408389, + 0.006452395115047693, + 0.0039440118707716465, + 0.0070742228999733925, + 0.0011544226435944438, + 0.0017561602871865034, + 0.0023192435037344694, + 0.0025307496543973684, + 0.0014391880249604583, + 0.0011809717398136854, + 0.0021619482431560755, + 0.0014208935899659991, + 0.00091885233996436, + 0.0011222694301977754, + 0.001882640179246664, + 0.0009888423373922706, + 0.001564205507747829, + 0.0007336479611694813, + 0.001565473503433168, + 0.0007800281164236367, + 0.00422390503808856, + 0.002463966142386198, + 0.008899440988898277, + 0.0017523050773888826, + 0.0032689645886421204, + 0.0038297364953905344, + 0.0020720832981169224, + 0.002870791358873248, + 0.0007556686759926379, + 0.0021805651485919952, + 0.0010024997172877192, + 0.00157355482224375, + 0.0022195226047188044, + 0.0011645297054201365, + 0.0018447930924594402, + 0.0011368724517524242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.10039852559566498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813069313764572 + ], + "opposite_logprob_tokens": [ + -3.288130760192871 + ], + "logprob_reward": -0.03813069313764572, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813069313764572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7959801554679871, + 0.022086989134550095, + 0.02071010135114193, + 0.05254742503166199, + 0.01764945313334465, + 0.036558568477630615, + 0.012492578476667404, + 0.012643187306821346, + 0.022342925891280174, + 0.006988642271608114 + ], + "credit_map": [ + 0.03329228237271309, + 0.04919475317001343, + 0.023120807483792305, + 0.03756965324282646, + 0.027029402554035187, + 0.3393664062023163, + 0.024140188470482826, + 0.262266606092453, + 0.0017676532734185457, + 0.002783284755423665, + 0.002077691024169326, + 0.00204166816547513, + 0.0030778327491134405, + 0.0019801349844783545, + 0.004636950325220823, + 0.0037217738572508097, + 0.001878943876363337, + 0.0015712124295532703, + 0.004518396686762571, + 0.002156817354261875, + 0.0033726212568581104, + 0.0017950645415112376, + 0.0033950230572372675, + 0.002022020285949111, + 0.005349081940948963, + 0.0059700701385736465, + 0.004866673611104488, + 0.005355744622647762, + 0.009469614364206791, + 0.009248582646250725, + 0.006766245234757662, + 0.0055214106105268, + 0.00100144452881068, + 0.002931977156549692, + 0.00120782561134547, + 0.0021593011915683746, + 0.003164722118526697, + 0.0023242889437824488, + 0.003043393837288022, + 0.001816498814150691, + 0.0021923521999269724, + 0.0023756742011755705, + 0.0025779646821320057, + 0.0027468909975141287, + 0.002804734278470278, + 0.010909445583820343, + 0.002653589239344001, + 0.01029791310429573, + 0.001328779966570437, + 0.0013076053000986576, + 0.0014709807001054287, + 0.002299841260537505, + 0.0014471921604126692, + 0.0009238719940185547, + 0.0020244414918124676, + 0.001689863740466535, + 0.0009910905500873923, + 0.0009308615699410439, + 0.003065661992877722, + 0.0009885394247248769, + 0.0017301500774919987, + 0.002639468526467681, + 0.0014229881344363093, + 0.000874425983056426, + 0.0034021586179733276, + 0.0017214330146089196, + 0.0071619837544858456, + 0.0011777362087741494, + 0.0024236522149294615, + 0.0023247746285051107, + 0.00210399879142642, + 0.0020271881949156523, + 0.00046553509309887886, + 0.001037353533320129, + 0.0006820623530074954, + 0.000971406523603946, + 0.0014636781997978687, + 0.0007240732666105032, + 0.0011064462596550584, + 0.0005380870425142348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.03813069313764572 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium close-up depicts a young man with rich olive skin and intense, dark eyes, set against a dimly lit, out-of-focus interior. He is dramatically illuminated by a single, harsh key light from the side, sculpting his features with high contrast and casting deep shadows that heighten the sense of tension. He suddenly and violently recoils, as if struck by an unseen force. His head jerks back, his shoulders recoil sharply, and his upper body instinctively twists away from the camera. His previously focused eyes are now wide with pure, unadulterated disbelief, his brow furrowed in confusion and shock. His lips part slightly in a silent, shocked gasp, his entire posture frozen in instinctive reaction, creating an atmosphere of intense tension and drama of a sudden, devastating revelation.", + "seed": 31944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_1/index_3886_seed_31944_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": -0.4450620137155056, + "vlm_logprob": -0.4450620137155056, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.4450620137155056, + "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 medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01933828555047512 + ], + "opposite_logprob_tokens": [ + -5.019338130950928 + ], + "logprob_reward": -0.01933828555047512, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.01933828555047512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11808184534311295, + 0.11893092840909958, + 0.0661502406001091, + 0.0451846644282341, + 0.06294524669647217, + 0.18528518080711365, + 0.12046636641025543, + 0.19620707631111145, + 0.031754955649375916, + 0.05499345809221268 + ], + "credit_map": [ + 0.011775866150856018, + 0.012978475540876389, + 0.012350290082395077, + 0.02002905309200287, + 0.010883383452892303, + 0.010815119370818138, + 0.01903349533677101, + 0.020216165110468864, + 0.0168224535882473, + 0.012322482652962208, + 0.03542429208755493, + 0.01013390813022852, + 0.010582582093775272, + 0.012038975022733212, + 0.010832972824573517, + 0.010773262940347195, + 0.005346735008060932, + 0.007773519493639469, + 0.010077092796564102, + 0.012254099361598492, + 0.009056639857590199, + 0.00487883435562253, + 0.010263357311487198, + 0.006499968469142914, + 0.003884833073243499, + 0.005337736569344997, + 0.005595460534095764, + 0.008993351832032204, + 0.006944290362298489, + 0.0036150729283690453, + 0.006578752771019936, + 0.004235164728015661, + 0.009711701422929764, + 0.007408729754388332, + 0.011131110601127148, + 0.00480116019025445, + 0.008760053664445877, + 0.006021855864673853, + 0.008887389674782753, + 0.006223240401595831, + 0.024586770683526993, + 0.01521551888436079, + 0.0201819259673357, + 0.0072892215102910995, + 0.025466931983828545, + 0.06283323466777802, + 0.010717536322772503, + 0.018994033336639404, + 0.011721311137080193, + 0.01523231528699398, + 0.020527292042970657, + 0.01267306786030531, + 0.014018050394952297, + 0.015535152517259121, + 0.016344184055924416, + 0.014414988458156586, + 0.029221201315522194, + 0.00931807141751051, + 0.08130686730146408, + 0.010450846515595913, + 0.017256220802664757, + 0.025493059307336807, + 0.01460188813507557, + 0.008558942005038261, + 0.0030216118320822716, + 0.0022425688803195953, + 0.004211814608424902, + 0.002458869479596615, + 0.0031859937589615583, + 0.006511275190860033, + 0.00289400527253747, + 0.0072288173250854015, + 0.005717530846595764, + 0.00530955009162426, + 0.008611995726823807, + 0.00468720868229866, + 0.007878362201154232, + 0.0076690479181706905, + 0.007565096952021122, + 0.007554664276540279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.01933828555047512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026379751041531563 + ], + "opposite_logprob_tokens": [ + -3.6513798236846924 + ], + "logprob_reward": -0.026379751041531563, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026379751041531563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20024122297763824, + 0.23360027372837067, + 0.07653868943452835, + 0.0615851990878582, + 0.0276056956499815, + 0.12222803384065628, + 0.06252187490463257, + 0.05736999213695526, + 0.022790957242250443, + 0.13551808893680573 + ], + "credit_map": [ + 0.018005207180976868, + 0.01626458764076233, + 0.01844879798591137, + 0.023957831785082817, + 0.01475625578314066, + 0.022300411015748978, + 0.028161056339740753, + 0.05834706872701645, + 0.03070181794464588, + 0.02055148221552372, + 0.04534232243895531, + 0.030692996457219124, + 0.017113080248236656, + 0.025325622409582138, + 0.02224467508494854, + 0.04162827506661415, + 0.00684313802048564, + 0.007929248735308647, + 0.015261821448802948, + 0.010465346276760101, + 0.010719608515501022, + 0.008811517618596554, + 0.011784113943576813, + 0.004723900929093361, + 0.004897132050246, + 0.010332928039133549, + 0.008834250271320343, + 0.009000427089631557, + 0.010106435976922512, + 0.0047496915794909, + 0.009874792769551277, + 0.0037895427085459232, + 0.002666864078491926, + 0.002978083910420537, + 0.004432321060448885, + 0.0020603653974831104, + 0.003941006027162075, + 0.0022464441135525703, + 0.005475890357047319, + 0.003804719541221857, + 0.013219621032476425, + 0.005692066624760628, + 0.017859628424048424, + 0.00509980134665966, + 0.015592187643051147, + 0.04214821010828018, + 0.006994070950895548, + 0.015622450970113277, + 0.00646138796582818, + 0.006904854439198971, + 0.011700109578669071, + 0.007159266620874405, + 0.007340373005717993, + 0.008279663510620594, + 0.007674111519008875, + 0.007002110127359629, + 0.0063252802938222885, + 0.004622036591172218, + 0.014824852347373962, + 0.004318087361752987, + 0.006598388310521841, + 0.007606067229062319, + 0.005616767797619104, + 0.0074585131369531155, + 0.003174288896843791, + 0.0026969199534505606, + 0.0028453716076910496, + 0.0023787845857441425, + 0.002984427148476243, + 0.003439327236264944, + 0.0029933482874184847, + 0.0022784871980547905, + 0.014530972577631474, + 0.012367977760732174, + 0.01898597553372383, + 0.012326684780418873, + 0.015865713357925415, + 0.02121652662754059, + 0.011600605212152004, + 0.028623634949326515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.026379751041531563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -2.017812728881836 + ], + "opposite_logprob_tokens": [ + -0.14281275868415833 + ], + "logprob_reward": -2.017812728881836, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.017812728881836, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06859887391328812, + 0.030708985403180122, + 0.052309758961200714, + 0.09064177423715591, + 0.11056998372077942, + 0.23402740061283112, + 0.0708598792552948, + 0.29679858684539795, + 0.022845525294542313, + 0.02263925038278103 + ], + "credit_map": [ + 0.009851458482444286, + 0.010258708149194717, + 0.006217302288860083, + 0.012270456179976463, + 0.007402782794088125, + 0.008909649215638638, + 0.007477844133973122, + 0.006210679188370705, + 0.0029923466499894857, + 0.004996494390070438, + 0.003349834121763706, + 0.0046509625390172005, + 0.004567301366478205, + 0.0031113841105252504, + 0.0034907469525933266, + 0.0035499120131134987, + 0.0038401337806135416, + 0.006777205038815737, + 0.005082727875560522, + 0.006251966115087271, + 0.010972831398248672, + 0.004021741449832916, + 0.010890890844166279, + 0.00447226082906127, + 0.0074853613041341305, + 0.008507689461112022, + 0.015660090371966362, + 0.010247808881103992, + 0.01843499392271042, + 0.008158590644598007, + 0.015435934998095036, + 0.006711299996823072, + 0.007438547443598509, + 0.007555690594017506, + 0.03507651016116142, + 0.006352428812533617, + 0.00799305085092783, + 0.0328756719827652, + 0.005831367801874876, + 0.007446715608239174, + 0.029389724135398865, + 0.013390901498496532, + 0.030126553028821945, + 0.008760579861700535, + 0.0320504829287529, + 0.08367954194545746, + 0.012297404929995537, + 0.024332212284207344, + 0.0074897943995893, + 0.00598907982930541, + 0.014341252855956554, + 0.008148041553795338, + 0.00700917886570096, + 0.010803237557411194, + 0.008709409274160862, + 0.008369878865778446, + 0.04419833421707153, + 0.0137461107224226, + 0.12538772821426392, + 0.01387907937169075, + 0.02623424492776394, + 0.03901464119553566, + 0.02181696891784668, + 0.012521459721028805, + 0.0032367906533181667, + 0.002423984697088599, + 0.003500318620353937, + 0.0027449000626802444, + 0.0026566467713564634, + 0.002615047385916114, + 0.002348960842937231, + 0.003318874863907695, + 0.002405295381322503, + 0.0020759643521159887, + 0.003093879669904709, + 0.0016184686683118343, + 0.002459392650052905, + 0.0038584480062127113, + 0.0023391058202832937, + 0.004788695368915796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -2.017812728881836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11306820809841156 + ], + "opposite_logprob_tokens": [ + -2.238068103790283 + ], + "logprob_reward": -0.11306820809841156, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11306820809841156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05310484394431114, + 0.04780074581503868, + 0.07209809124469757, + 0.07451983541250229, + 0.0772588849067688, + 0.40128791332244873, + 0.08178019523620605, + 0.09252599626779556, + 0.044899094849824905, + 0.054724376648664474 + ], + "credit_map": [ + 0.005256763193756342, + 0.0076283239759504795, + 0.004337675869464874, + 0.011150750331580639, + 0.0052947476506233215, + 0.004392039496451616, + 0.007319174706935883, + 0.0077253710478544235, + 0.004121705889701843, + 0.00591290881857276, + 0.006222131196409464, + 0.007379923481494188, + 0.005632389802485704, + 0.00484894635155797, + 0.005305816885083914, + 0.008376926183700562, + 0.005104715004563332, + 0.007364090997725725, + 0.0076864976435899734, + 0.009414643980562687, + 0.015260733664035797, + 0.005498514976352453, + 0.015202431008219719, + 0.0065664686262607574, + 0.0059050400741398335, + 0.008806283585727215, + 0.009555417113006115, + 0.008396508172154427, + 0.015013111755251884, + 0.007160198409110308, + 0.012390446849167347, + 0.0072928303852677345, + 0.00655102776363492, + 0.011601068079471588, + 0.015906337648630142, + 0.006106649525463581, + 0.010346751660108566, + 0.01018030010163784, + 0.010478080250322819, + 0.0060886722058057785, + 0.049610842019319534, + 0.02251983806490898, + 0.052499350160360336, + 0.015241818502545357, + 0.055318254977464676, + 0.14386366307735443, + 0.02107667364180088, + 0.04115748405456543, + 0.009320969693362713, + 0.007493708748370409, + 0.016349157318472862, + 0.00851823017001152, + 0.008115766569972038, + 0.01268092729151249, + 0.009492134675383568, + 0.009809310548007488, + 0.013631084002554417, + 0.006228865589946508, + 0.03524148091673851, + 0.006204456090927124, + 0.007979369722306728, + 0.01057513989508152, + 0.006615377962589264, + 0.006050231400877237, + 0.006852410733699799, + 0.0046089752577245235, + 0.008794073015451431, + 0.00404207082465291, + 0.0049370951019227505, + 0.005776616744697094, + 0.005011274013668299, + 0.004876584280282259, + 0.005137904081493616, + 0.007513045333325863, + 0.006424478255212307, + 0.0051930914632976055, + 0.005795464385300875, + 0.008460693061351776, + 0.005780015140771866, + 0.010419690981507301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.11306820809841156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up of a young man with rich olive skin and dark eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dimly lit, out-of-focus interior illuminated by a single harsh key light from the side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character suddenly and violently recoil with his head jerking back and shoulders recoiling sharply?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the character's upper body twist away from the camera while his eyes widen in disbelief and his brow furrows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the character's lips part slightly in a silent gasp while his posture remains frozen in reaction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871109500527382 + ], + "opposite_logprob_tokens": [ + -3.048711061477661 + ], + "logprob_reward": -0.04871109500527382, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871109500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05190175026655197, + 0.09880242496728897, + 0.0661287009716034, + 0.07498275488615036, + 0.12228535115718842, + 0.21699388325214386, + 0.04659254476428032, + 0.24825234711170197, + 0.046559788286685944, + 0.02750047668814659 + ], + "credit_map": [ + 0.00633491575717926, + 0.0076854657381772995, + 0.004676518961787224, + 0.009475192055106163, + 0.005065270233899355, + 0.005760586354881525, + 0.0058115096762776375, + 0.007092293817549944, + 0.010681083425879478, + 0.009067189879715443, + 0.012708673253655434, + 0.010984782129526138, + 0.008835500106215477, + 0.01224244199693203, + 0.00938560999929905, + 0.024897141382098198, + 0.0051106177270412445, + 0.006461073644459248, + 0.007530013099312782, + 0.009484672918915749, + 0.01176612637937069, + 0.005348545033484697, + 0.014454802498221397, + 0.005972843151539564, + 0.006152597721666098, + 0.006486828438937664, + 0.012990829534828663, + 0.008463110774755478, + 0.015792490914463997, + 0.006898877210915089, + 0.012931025587022305, + 0.005266991909593344, + 0.0076660276390612125, + 0.006113376468420029, + 0.0353546068072319, + 0.007057116366922855, + 0.009600413963198662, + 0.03911404311656952, + 0.00828547403216362, + 0.009094297885894775, + 0.02686353214085102, + 0.01324805524200201, + 0.027981512248516083, + 0.008548850193619728, + 0.030132228508591652, + 0.0765027329325676, + 0.012011156417429447, + 0.021705834195017815, + 0.004464829806238413, + 0.0067627206444740295, + 0.007228899281471968, + 0.005823270417749882, + 0.005386282689869404, + 0.005042741075158119, + 0.007522806525230408, + 0.0043609910644590855, + 0.03873541206121445, + 0.011249342933297157, + 0.104320228099823, + 0.012109394185245037, + 0.021310707554221153, + 0.030427951365709305, + 0.018368171527981758, + 0.011731145903468132, + 0.006861184723675251, + 0.004317933227866888, + 0.008940824307501316, + 0.003291639732196927, + 0.0046865507028996944, + 0.005488267634063959, + 0.005371583625674248, + 0.0076018040999770164, + 0.002663982566446066, + 0.0036653829738497734, + 0.0033720172941684723, + 0.003118805354461074, + 0.0034410040825605392, + 0.004195541609078646, + 0.0035229939967393875, + 0.003520747646689415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.04871109500527382 + } + ], + "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/0000055/group_10/index_tensor([4092, 4092])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_tensor([4092, 4092])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dbe166b32b42f89c013134d3c618312382749cad --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_tensor([4092, 4092])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Told from a low-crouching first-person perspective, the camera advances in a shaky, handheld motion, cautiously navigating a scene of apocalyptic disaster. The ground is a chaotic mess, littered with powdered cement, twisted rebar, and the charred husks of military equipment. In the immediate foreground, a pair of hands clad in fingerless black tactical gloves can be seen, one pressed against the jagged concrete, the other gripping the butt of a futuristic assault rifle. In the middle distance, a burning vehicle billows thick, oily black smoke, while a gritty, gray dust kicked up by the chaos hangs heavily in the air, obscuring the skeletal remains of bombed-out high-rises against an oppressive, brassy sky. The entire frame periodically shakes with the deep, concussive vibrations of distant impacts, causing the ground to tremble violently and sending small pebbles and debris skittering across its surface. With each jolt, the camera lurches, momentarily losing focus, before steadily reorienting and continuing its slow, deliberate crawl forward. Harsh sunlight struggles to penetrate the thick haze, casting long, distorted shadows and creating sharp, gleaming highlights on shards of glass and discarded bullet casings, creating a desaturated and austere palette that heightens the atmosphere of extreme tension and visceral disorientation.", + "seed": 11012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11012_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": -0.01792518878355622, + "vlm_logprob": -0.01792518878355622, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.01792518878355622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03742186352610588 + ], + "opposite_logprob_tokens": [ + -3.662421941757202 + ], + "logprob_reward": -0.03742186352610588, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03742186352610588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.243142232298851, + 0.07033735513687134, + 0.04502160847187042, + 0.10512436181306839, + 0.0666477158665657, + 0.06576226651668549, + 0.08489016443490982, + 0.07181397825479507, + 0.09378213435411453, + 0.15347814559936523 + ], + "credit_map": [ + 0.036131229251623154, + 0.021940743550658226, + 0.06310174614191055, + 0.024307534098625183, + 0.023486079648137093, + 0.02708660438656807, + 0.021739136427640915, + 0.02534916251897812, + 0.006257664877921343, + 0.011653393507003784, + 0.007229435723274946, + 0.009991972707211971, + 0.010225770063698292, + 0.008663571439683437, + 0.00911371223628521, + 0.007201844826340675, + 0.003954204730689526, + 0.006186440587043762, + 0.0063956924714148045, + 0.005825659725815058, + 0.005709763616323471, + 0.004762860480695963, + 0.006330130621790886, + 0.005856858566403389, + 0.008792298845946789, + 0.013376714661717415, + 0.016333403065800667, + 0.014274408109486103, + 0.0187124814838171, + 0.010170605033636093, + 0.013054337352514267, + 0.010410112328827381, + 0.006747766863554716, + 0.010191556066274643, + 0.010619851760566235, + 0.01085362397134304, + 0.006869580131024122, + 0.010250222869217396, + 0.004742123652249575, + 0.006372991483658552, + 0.007867520675063133, + 0.008026628755033016, + 0.009494436904788017, + 0.006677613593637943, + 0.010455229319632053, + 0.010861314833164215, + 0.0044527375139296055, + 0.007926784455776215, + 0.008883383125066757, + 0.008177066221833229, + 0.009637963958084583, + 0.009803305380046368, + 0.00803467445075512, + 0.01111582387238741, + 0.018696341663599014, + 0.01054160576313734, + 0.005695418454706669, + 0.009522157721221447, + 0.006263241171836853, + 0.007737953215837479, + 0.011261739768087864, + 0.008341298438608646, + 0.01419062353670597, + 0.00880154687911272, + 0.00526223611086607, + 0.010100478306412697, + 0.011853398755192757, + 0.011274171993136406, + 0.017579244449734688, + 0.011204080656170845, + 0.01546651590615511, + 0.011042012833058834, + 0.012621482834219933, + 0.018489396199584007, + 0.019237006083130836, + 0.019538311287760735, + 0.014766771346330643, + 0.023230381309986115, + 0.027239929884672165, + 0.018354864791035652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.03742186352610588 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023315908387303352 + ], + "opposite_logprob_tokens": [ + -3.773315906524658 + ], + "logprob_reward": -0.023315908387303352, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023315908387303352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.245652437210083, + 0.0772828534245491, + 0.03879356011748314, + 0.042118534445762634, + 0.08056337386369705, + 0.04288964346051216, + 0.02932910993695259, + 0.05969046801328659, + 0.07343761622905731, + 0.3102423846721649 + ], + "credit_map": [ + 0.035564035177230835, + 0.0863901674747467, + 0.042247362434864044, + 0.022765949368476868, + 0.023144058883190155, + 0.012740621343255043, + 0.012453760951757431, + 0.010346478782594204, + 0.006877095904201269, + 0.009482610039412975, + 0.013107135891914368, + 0.013026643544435501, + 0.009677882306277752, + 0.0067494651302695274, + 0.011236315593123436, + 0.007125701755285263, + 0.003697506384924054, + 0.009089257568120956, + 0.006921149790287018, + 0.004499825648963451, + 0.004292102064937353, + 0.00381236569955945, + 0.0033066975884139538, + 0.0031746579334139824, + 0.0036290795542299747, + 0.005498149897903204, + 0.0060178604908287525, + 0.005743714980781078, + 0.004971583839505911, + 0.0036177770234644413, + 0.007926658727228642, + 0.004713710397481918, + 0.006088797934353352, + 0.009606012143194675, + 0.00799042172729969, + 0.008838185109198093, + 0.009931742213666439, + 0.013934335671365261, + 0.01169921737164259, + 0.012474664486944675, + 0.005176324862986803, + 0.004934288561344147, + 0.005677283741533756, + 0.007879065349698067, + 0.006200216710567474, + 0.006524043157696724, + 0.003080220427364111, + 0.0034181992523372173, + 0.0029014581814408302, + 0.004391944967210293, + 0.0035241208970546722, + 0.004072688054293394, + 0.0028755918610841036, + 0.004516359884291887, + 0.0043581025674939156, + 0.002688844921067357, + 0.0032580189872533083, + 0.009192230179905891, + 0.00568876089528203, + 0.006021907087415457, + 0.011229915544390678, + 0.007808180060237646, + 0.010303442366421223, + 0.006188009865581989, + 0.004250530153512955, + 0.005132449325174093, + 0.011381866410374641, + 0.009657674469053745, + 0.007886768318712711, + 0.009738212451338768, + 0.01669779047369957, + 0.008692324161529541, + 0.030148649588227272, + 0.0408104807138443, + 0.03574527055025101, + 0.042702510952949524, + 0.03061199001967907, + 0.0421878956258297, + 0.05619348585605621, + 0.031842101365327835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.023315908387303352 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550304643809795 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550304643809795, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550304643809795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1979711651802063, + 0.14062727987766266, + 0.04220953211188316, + 0.13273286819458008, + 0.058787450194358826, + 0.05475451052188873, + 0.12630930542945862, + 0.06817752122879028, + 0.059696391224861145, + 0.11873403936624527 + ], + "credit_map": [ + 0.026735560968518257, + 0.02047801949083805, + 0.04042733833193779, + 0.030046315863728523, + 0.020258072763681412, + 0.018284855410456657, + 0.019195733591914177, + 0.022545238956809044, + 0.01178266853094101, + 0.023648470640182495, + 0.016523147001862526, + 0.021131834015250206, + 0.022565806284546852, + 0.016370777040719986, + 0.018713215366005898, + 0.009891360066831112, + 0.004197109956294298, + 0.004633067641407251, + 0.005610163789242506, + 0.005812210496515036, + 0.005665350239723921, + 0.0037206520792096853, + 0.005611245986074209, + 0.006959725636988878, + 0.009053343906998634, + 0.017664024606347084, + 0.01641698367893696, + 0.012172743678092957, + 0.01864255964756012, + 0.016578475013375282, + 0.022392088547348976, + 0.019812636077404022, + 0.006579399108886719, + 0.009402362629771233, + 0.008567904122173786, + 0.006619932129979134, + 0.00652447110041976, + 0.010640494525432587, + 0.0047537642531096935, + 0.00569911440834403, + 0.005361296236515045, + 0.006122322287410498, + 0.006345738656818867, + 0.012233770452439785, + 0.005862032528966665, + 0.008009248413145542, + 0.004798051901161671, + 0.006022046320140362, + 0.013087310828268528, + 0.014867009595036507, + 0.014975929632782936, + 0.016853054985404015, + 0.010532164014875889, + 0.019566725939512253, + 0.024599067866802216, + 0.011828015558421612, + 0.005364345386624336, + 0.008311848156154156, + 0.006996018812060356, + 0.0065445806831121445, + 0.01041689794510603, + 0.007323587778955698, + 0.013119475916028023, + 0.010100760497152805, + 0.0027475515380501747, + 0.007917012088000774, + 0.007524795830249786, + 0.004322424530982971, + 0.01765701174736023, + 0.0071973539888858795, + 0.00657625962048769, + 0.0057539730332791805, + 0.010490325279533863, + 0.01934356987476349, + 0.01424394827336073, + 0.01257874071598053, + 0.011810108087956905, + 0.018247682601213455, + 0.0162703525274992, + 0.015749307349324226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.012550304643809795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23812267184257507, + 0.09244806319475174, + 0.07723784446716309, + 0.053960852324962616, + 0.07165813446044922, + 0.12683703005313873, + 0.058743689209222794, + 0.04335203766822815, + 0.15729190409183502, + 0.08034773170948029 + ], + "credit_map": [ + 0.023276222869753838, + 0.04484330862760544, + 0.08852816373109818, + 0.024596020579338074, + 0.015094593167304993, + 0.016675753518939018, + 0.012892396189272404, + 0.012216228060424328, + 0.010004833340644836, + 0.01901709847152233, + 0.01017734408378601, + 0.01253405399620533, + 0.016098517924547195, + 0.009887315332889557, + 0.008080323226749897, + 0.006648587062954903, + 0.007338383700698614, + 0.02014700323343277, + 0.015074051916599274, + 0.007461066357791424, + 0.008976275101304054, + 0.007584352977573872, + 0.006233809515833855, + 0.004422914236783981, + 0.0031696998048573732, + 0.007274842355400324, + 0.004754757508635521, + 0.00591230858117342, + 0.006217246409505606, + 0.007414708845317364, + 0.01112731546163559, + 0.008089975453913212, + 0.005751129239797592, + 0.009132020175457, + 0.008088644593954086, + 0.007177336607128382, + 0.008105659857392311, + 0.012465844862163067, + 0.009657950140535831, + 0.011279555968940258, + 0.013275252655148506, + 0.018290942534804344, + 0.017441188916563988, + 0.020399050787091255, + 0.022841060534119606, + 0.012891949154436588, + 0.011294442228972912, + 0.010403157211840153, + 0.005818414501845837, + 0.004886039532721043, + 0.007418585941195488, + 0.00789173785597086, + 0.004709616303443909, + 0.013335051015019417, + 0.00947999581694603, + 0.005204253364354372, + 0.0031372627709060907, + 0.006090765818953514, + 0.004149541724473238, + 0.00374322896823287, + 0.008506822399795055, + 0.006030454766005278, + 0.005732350517064333, + 0.005961614660918713, + 0.00962817668914795, + 0.010717547498643398, + 0.016455532982945442, + 0.020148539915680885, + 0.017191082239151, + 0.015193684957921505, + 0.0490725040435791, + 0.01888485997915268, + 0.005949589889496565, + 0.014102720655500889, + 0.00959780253469944, + 0.010478679090738297, + 0.0083389887586236, + 0.012554613873362541, + 0.012710143812000751, + 0.0066152033396065235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.5768004655838013, + 0.033551279455423355, + 0.022929010912775993, + 0.023271415382623672, + 0.13304558396339417, + 0.02614288032054901, + 0.025395287200808525, + 0.04523134604096413, + 0.0464400053024292, + 0.06719278544187546 + ], + "credit_map": [ + 0.043983034789562225, + 0.16219709813594818, + 0.15864452719688416, + 0.04827424883842468, + 0.05582420527935028, + 0.060961976647377014, + 0.02273116633296013, + 0.024184176698327065, + 0.003725592978298664, + 0.00370441609993577, + 0.005249024834483862, + 0.006410363595932722, + 0.005082744639366865, + 0.0035547688603401184, + 0.003222069004550576, + 0.002602299675345421, + 0.0016404244815930724, + 0.004562947433441877, + 0.0036021375562995672, + 0.0029451227746903896, + 0.0028910476248711348, + 0.0023024417459964752, + 0.002127612940967083, + 0.0028572764713317156, + 0.002143976278603077, + 0.004129711072891951, + 0.0036508459597826004, + 0.002689565997570753, + 0.003112112171947956, + 0.0018495036056265235, + 0.003439567517489195, + 0.002256134059280157, + 0.011002752929925919, + 0.018422935158014297, + 0.015182162635028362, + 0.012822899036109447, + 0.01568399742245674, + 0.02331402339041233, + 0.017027053982019424, + 0.019589753821492195, + 0.002724290359765291, + 0.0019522585207596421, + 0.004984214901924133, + 0.004061516374349594, + 0.0028574932366609573, + 0.004217187408357859, + 0.0017626184271648526, + 0.003583300858736038, + 0.0027473943773657084, + 0.0036276623141020536, + 0.003075527725741267, + 0.0045027416199445724, + 0.002347712405025959, + 0.003990577068179846, + 0.0032924239058047533, + 0.0018112483667209744, + 0.004080724902451038, + 0.00548944715410471, + 0.003887286875396967, + 0.005271402187645435, + 0.004953840747475624, + 0.006064014974981546, + 0.00961859431117773, + 0.005866033956408501, + 0.002873872872442007, + 0.00287387752905488, + 0.0057884640991687775, + 0.006103967782109976, + 0.005079985596239567, + 0.0053053246811032295, + 0.013025745749473572, + 0.005388767924159765, + 0.008866045624017715, + 0.010840637609362602, + 0.008701616898179054, + 0.008003821596503258, + 0.006911627948284149, + 0.010374083183705807, + 0.00668593542650342, + 0.006809020880609751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.00525687774643302 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a low-crouching first-person perspective, the camera advances in a shaky, handheld motion, cautiously navigating a scene of apocalyptic disaster. The ground is a chaotic mess, littered with powdered cement, twisted rebar, and the charred husks of military equipment. In the immediate foreground, a pair of hands clad in fingerless black tactical gloves can be seen, one pressed against the jagged concrete, the other gripping the butt of a futuristic assault rifle. In the middle distance, a burning vehicle billows thick, oily black smoke, while a gritty, gray dust kicked up by the chaos hangs heavily in the air, obscuring the skeletal remains of bombed-out high-rises against an oppressive, brassy sky. The entire frame periodically shakes with the deep, concussive vibrations of distant impacts, causing the ground to tremble violently and sending small pebbles and debris skittering across its surface. With each jolt, the camera lurches, momentarily losing focus, before steadily reorienting and continuing its slow, deliberate crawl forward. Harsh sunlight struggles to penetrate the thick haze, casting long, distorted shadows and creating sharp, gleaming highlights on shards of glass and discarded bullet casings, creating a desaturated and austere palette that heightens the atmosphere of extreme tension and visceral disorientation.", + "seed": 11013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11013_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": -0.018815412186086178, + "vlm_logprob": -0.018815412186086178, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.018815412186086178, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032137081027030945 + ], + "opposite_logprob_tokens": [ + -4.157136917114258 + ], + "logprob_reward": -0.032137081027030945, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.032137081027030945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09669185429811478, + 0.06852979958057404, + 0.05666081979870796, + 0.03654918447136879, + 0.1339658796787262, + 0.06303485482931137, + 0.05641988664865494, + 0.0675739198923111, + 0.35086140036582947, + 0.06971236318349838 + ], + "credit_map": [ + 0.013529852032661438, + 0.008803587406873703, + 0.015092626214027405, + 0.01724035106599331, + 0.009027697145938873, + 0.00842911098152399, + 0.011957916431128979, + 0.012610714882612228, + 0.011336724273860455, + 0.012901701964437962, + 0.010492002591490746, + 0.005402795970439911, + 0.009419247508049011, + 0.005906000267714262, + 0.0065616704523563385, + 0.006509655155241489, + 0.004346158821135759, + 0.004360030870884657, + 0.007638722192496061, + 0.006219796370714903, + 0.005834226030856371, + 0.00615057535469532, + 0.01566125825047493, + 0.00645005377009511, + 0.0030133479740470648, + 0.005860032979398966, + 0.004128831904381514, + 0.003896470181643963, + 0.004263320472091436, + 0.005757685285061598, + 0.005598914809525013, + 0.004030582029372454, + 0.021288732066750526, + 0.010845501907169819, + 0.028216077014803886, + 0.015247473493218422, + 0.013957303948700428, + 0.018247351050376892, + 0.010380563326179981, + 0.01578286476433277, + 0.012483177706599236, + 0.006499074399471283, + 0.012545419856905937, + 0.0071028899401426315, + 0.005182079970836639, + 0.006285388953983784, + 0.007685021962970495, + 0.005251798313111067, + 0.008722621947526932, + 0.007902796380221844, + 0.005961711052805185, + 0.004737100563943386, + 0.0127352150157094, + 0.006977238226681948, + 0.005754677578806877, + 0.0036285247188061476, + 0.011352959088981152, + 0.009618810378015041, + 0.008575605228543282, + 0.006595965940505266, + 0.009836984798312187, + 0.010178236290812492, + 0.006692827213555574, + 0.00472252955660224, + 0.03019702434539795, + 0.017965370789170265, + 0.11621786653995514, + 0.018368691205978394, + 0.11026273667812347, + 0.03277695178985596, + 0.015780290588736534, + 0.009292496368288994, + 0.004683696664869785, + 0.012896171770989895, + 0.0075363051146268845, + 0.009529911912977695, + 0.007860001176595688, + 0.013243630528450012, + 0.006850456818938255, + 0.007112190593034029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.032137081027030945 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982354536652565 + ], + "opposite_logprob_tokens": [ + -3.5298235416412354 + ], + "logprob_reward": -0.02982354536652565, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982354536652565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556312680244446, + 0.04607250541448593, + 0.037812571972608566, + 0.10234200209379196, + 0.10262331366539001, + 0.054262783378362656, + 0.06313927471637726, + 0.06283813714981079, + 0.24302232265472412, + 0.03225581720471382 + ], + "credit_map": [ + 0.04156262427568436, + 0.017393913120031357, + 0.06804633885622025, + 0.03967023640871048, + 0.01659552939236164, + 0.01800837740302086, + 0.026815185323357582, + 0.02753906510770321, + 0.006778201088309288, + 0.006693015806376934, + 0.0057252440601587296, + 0.006034431047737598, + 0.008624204434454441, + 0.004699768964201212, + 0.003790512215346098, + 0.0037271257024258375, + 0.004159650765359402, + 0.005219283979386091, + 0.005041849333792925, + 0.005048296879976988, + 0.004853142891079187, + 0.004821903072297573, + 0.0049707405269145966, + 0.0036977031268179417, + 0.007612010929733515, + 0.015626072883605957, + 0.012249617837369442, + 0.011254184879362583, + 0.012316972948610783, + 0.01927891932427883, + 0.013504358939826488, + 0.010499868541955948, + 0.01430893037468195, + 0.009139117784798145, + 0.0161188542842865, + 0.012450823560357094, + 0.011997055262327194, + 0.014754010364413261, + 0.012209624983370304, + 0.01164490170776844, + 0.010770057328045368, + 0.004861126653850079, + 0.00963639933615923, + 0.00564414681866765, + 0.003804184263572097, + 0.005810175556689501, + 0.007711939979344606, + 0.00602475181221962, + 0.01359625905752182, + 0.008965590037405491, + 0.005737170111387968, + 0.006523428950458765, + 0.010722441598773003, + 0.006649211049079895, + 0.0070706941187381744, + 0.0038744830526411533, + 0.012533679604530334, + 0.008470720611512661, + 0.007605194579809904, + 0.005491315852850676, + 0.00972184631973505, + 0.007863444276154041, + 0.006080225110054016, + 0.005071717780083418, + 0.02016138657927513, + 0.01075637899339199, + 0.08155226707458496, + 0.012911416590213776, + 0.07626878470182419, + 0.02307773008942604, + 0.012445325963199139, + 0.0058490438386797905, + 0.002797097899019718, + 0.005472855176776648, + 0.00395900197327137, + 0.005300513934344053, + 0.003507041372358799, + 0.004815146327018738, + 0.00379548710770905, + 0.0026086720172315836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02982354536652565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547126039862633 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012547126039862633, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547126039862633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21311776340007782, + 0.06148355454206467, + 0.09327416121959686, + 0.19207727909088135, + 0.05529167130589485, + 0.07139980792999268, + 0.14263106882572174, + 0.055483587086200714, + 0.06521451473236084, + 0.05002661421895027 + ], + "credit_map": [ + 0.025720831006765366, + 0.022519845515489578, + 0.03242278844118118, + 0.030229514464735985, + 0.020155662670731544, + 0.02622978575527668, + 0.029455406591296196, + 0.02638392522931099, + 0.004863569512963295, + 0.006615577265620232, + 0.010831929743289948, + 0.0078288484364748, + 0.010318023152649403, + 0.007618853822350502, + 0.005448931362479925, + 0.007957817055284977, + 0.007468341384083033, + 0.009300080128014088, + 0.011614327318966389, + 0.010984953492879868, + 0.008246458135545254, + 0.010419704020023346, + 0.02320372499525547, + 0.01203655730932951, + 0.015116643160581589, + 0.032606467604637146, + 0.026488380506634712, + 0.02586454339325428, + 0.020403893664479256, + 0.02902035415172577, + 0.023001566529273987, + 0.01957540214061737, + 0.009815474040806293, + 0.004108663648366928, + 0.00887343566864729, + 0.007014246191829443, + 0.007019046228379011, + 0.006953376345336437, + 0.00594033719971776, + 0.00556708499789238, + 0.011043244041502476, + 0.006415385752916336, + 0.00928120780736208, + 0.006701469887048006, + 0.008124260231852531, + 0.011166292242705822, + 0.009693575091660023, + 0.008974368683993816, + 0.029333122074604034, + 0.018316887319087982, + 0.016014905646443367, + 0.018471816554665565, + 0.01962634176015854, + 0.014671740122139454, + 0.01771731488406658, + 0.008478919975459576, + 0.009219005703926086, + 0.007215958554297686, + 0.007691584527492523, + 0.005386949982494116, + 0.00906720757484436, + 0.007962316274642944, + 0.005420092958956957, + 0.0035204628948122263, + 0.006373055279254913, + 0.00729770865291357, + 0.00753745436668396, + 0.009819019585847855, + 0.009321784600615501, + 0.010162226855754852, + 0.010822161100804806, + 0.0038811014965176582, + 0.0039334832690656185, + 0.005635952576994896, + 0.006380800157785416, + 0.006931603420525789, + 0.008530829101800919, + 0.008037963882088661, + 0.006936843041330576, + 0.0036391355097293854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.012547126039862633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08773549646139145, + 0.18847520649433136, + 0.17592474818229675, + 0.031117001548409462, + 0.051987797021865845, + 0.022396860644221306, + 0.18409959971904755, + 0.0916791558265686, + 0.1317002922296524, + 0.03488383814692497 + ], + "credit_map": [ + 0.013862951658666134, + 0.009201349690556526, + 0.02281356044113636, + 0.011022054590284824, + 0.00830859411507845, + 0.006079967599362135, + 0.007715811487287283, + 0.008731197565793991, + 0.01930324360728264, + 0.031986597925424576, + 0.029350681230425835, + 0.02633732557296753, + 0.027786901220679283, + 0.021023156121373177, + 0.016395270824432373, + 0.016292031854391098, + 0.020292267203330994, + 0.03065725415945053, + 0.021403536200523376, + 0.02496907487511635, + 0.0275273434817791, + 0.026150498539209366, + 0.014641838148236275, + 0.010282943956553936, + 0.0038057167548686266, + 0.005237223114818335, + 0.003764270804822445, + 0.0036982239689677954, + 0.003326834412291646, + 0.0047180913388729095, + 0.0033853924833238125, + 0.00318124913610518, + 0.007895240560173988, + 0.003730558091774583, + 0.010348953306674957, + 0.006952268071472645, + 0.0061610061675310135, + 0.006702839862555265, + 0.005265876185148954, + 0.004931057803332806, + 0.0033480310812592506, + 0.002240890869870782, + 0.00455879233777523, + 0.0028102684300392866, + 0.0023931339383125305, + 0.002376653952524066, + 0.002418889431282878, + 0.0022502008359879255, + 0.023951806128025055, + 0.027715986594557762, + 0.019145572558045387, + 0.017784137278795242, + 0.04834922403097153, + 0.01713895983994007, + 0.020167194306850433, + 0.009846722707152367, + 0.014423729851841927, + 0.010122903622686863, + 0.012763438746333122, + 0.013980264775454998, + 0.009429926052689552, + 0.014623045921325684, + 0.008665887638926506, + 0.007669956889003515, + 0.014491179026663303, + 0.012097533792257309, + 0.025923317298293114, + 0.02156425639986992, + 0.02704470232129097, + 0.011409776285290718, + 0.012829982675611973, + 0.006339551415294409, + 0.0021006125025451183, + 0.00622952077537775, + 0.003162776120007038, + 0.004062531981617212, + 0.005262314807623625, + 0.007167226169258356, + 0.0032805257942527533, + 0.0036183323245495558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 2 + ], + "selected_reward": -0.009785185568034649 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18978779017925262, + 0.1178901270031929, + 0.04383421689271927, + 0.047631170600652695, + 0.08906028419733047, + 0.10237230360507965, + 0.13485562801361084, + 0.0688910037279129, + 0.10001585632562637, + 0.10566163063049316 + ], + "credit_map": [ + 0.03332418575882912, + 0.013833951205015182, + 0.05862046033143997, + 0.025092944502830505, + 0.01162816397845745, + 0.014865504577755928, + 0.014033710584044456, + 0.018388871103525162, + 0.011000965721905231, + 0.013584146276116371, + 0.01952527090907097, + 0.013507054187357426, + 0.022694546729326248, + 0.017722757533192635, + 0.009481326676905155, + 0.010374053381383419, + 0.0036562138702720404, + 0.004411584697663784, + 0.006878717802464962, + 0.005572888068854809, + 0.004261173773556948, + 0.005339264404028654, + 0.008828829973936081, + 0.004885546397417784, + 0.0039029496256262064, + 0.00630927924066782, + 0.0059840502217411995, + 0.006158717442303896, + 0.005346409510821104, + 0.009464247152209282, + 0.005581271834671497, + 0.004884245805442333, + 0.013097979128360748, + 0.006601893808692694, + 0.014491748996078968, + 0.009417744353413582, + 0.010422280989587307, + 0.012499100528657436, + 0.011753723956644535, + 0.010775805450975895, + 0.020309118553996086, + 0.008898966945707798, + 0.02227582409977913, + 0.009809771552681923, + 0.010055717080831528, + 0.011621268466114998, + 0.010025939904153347, + 0.00937569234520197, + 0.02083260379731655, + 0.020061500370502472, + 0.01601913943886757, + 0.01845865696668625, + 0.025909502059221268, + 0.013081379234790802, + 0.014229722321033478, + 0.006263115908950567, + 0.012768367305397987, + 0.010269044898450375, + 0.008031202480196953, + 0.005150850396603346, + 0.009518101811408997, + 0.008168848231434822, + 0.009821165353059769, + 0.005163424648344517, + 0.010821959003806114, + 0.01320552360266447, + 0.012884855270385742, + 0.019213493913412094, + 0.0143592469394207, + 0.014256352558732033, + 0.010745367035269737, + 0.004529058001935482, + 0.00839563924819231, + 0.010800505056977272, + 0.009667606092989445, + 0.010690201073884964, + 0.015942392870783806, + 0.029458554461598396, + 0.009901710785925388, + 0.010805019177496433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.009784122928977013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a low-crouching first-person perspective, the camera advances in a shaky, handheld motion, cautiously navigating a scene of apocalyptic disaster. The ground is a chaotic mess, littered with powdered cement, twisted rebar, and the charred husks of military equipment. In the immediate foreground, a pair of hands clad in fingerless black tactical gloves can be seen, one pressed against the jagged concrete, the other gripping the butt of a futuristic assault rifle. In the middle distance, a burning vehicle billows thick, oily black smoke, while a gritty, gray dust kicked up by the chaos hangs heavily in the air, obscuring the skeletal remains of bombed-out high-rises against an oppressive, brassy sky. The entire frame periodically shakes with the deep, concussive vibrations of distant impacts, causing the ground to tremble violently and sending small pebbles and debris skittering across its surface. With each jolt, the camera lurches, momentarily losing focus, before steadily reorienting and continuing its slow, deliberate crawl forward. Harsh sunlight struggles to penetrate the thick haze, casting long, distorted shadows and creating sharp, gleaming highlights on shards of glass and discarded bullet casings, creating a desaturated and austere palette that heightens the atmosphere of extreme tension and visceral disorientation.", + "seed": 11014, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11014_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": -0.02438489254564047, + "vlm_logprob": -0.02438489254564047, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.02438489254564047, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05417373776435852 + ], + "opposite_logprob_tokens": [ + -3.179173707962036 + ], + "logprob_reward": -0.05417373776435852, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05417373776435852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07055195420980453, + 0.20258355140686035, + 0.059592220932245255, + 0.15711966156959534, + 0.18118003010749817, + 0.14768174290657043, + 0.024657834321260452, + 0.019670508801937103, + 0.026357250288128853, + 0.11060523986816406 + ], + "credit_map": [ + 0.012404965236783028, + 0.012278212234377861, + 0.0041718450374901295, + 0.005352651700377464, + 0.0069668833166360855, + 0.009174639359116554, + 0.00744371535256505, + 0.012759045697748661, + 0.023624824360013008, + 0.02010069042444229, + 0.0066032130271196365, + 0.008198757655918598, + 0.015143678523600101, + 0.06301693618297577, + 0.009458882734179497, + 0.05643656104803085, + 0.005501491948962212, + 0.006683036684989929, + 0.00898705329746008, + 0.008305368013679981, + 0.009232967160642147, + 0.0067474679090082645, + 0.007013089954853058, + 0.007121745031327009, + 0.007950574159622192, + 0.016456641256809235, + 0.009274507872760296, + 0.011026361025869846, + 0.06043170392513275, + 0.017718438059091568, + 0.008014427497982979, + 0.02624700218439102, + 0.01683221571147442, + 0.02105589583516121, + 0.026528822258114815, + 0.03263985738158226, + 0.017701851204037666, + 0.030785076320171356, + 0.01688540354371071, + 0.01875089295208454, + 0.004981652367860079, + 0.01720726490020752, + 0.0046728928573429585, + 0.006834121886640787, + 0.00798092782497406, + 0.00695271510630846, + 0.06240140646696091, + 0.03665076568722725, + 0.0020165599416941404, + 0.0025183965917676687, + 0.0015175349544733763, + 0.0025074975565075874, + 0.004163600038737059, + 0.003861453849822283, + 0.004458399023860693, + 0.0036143907345831394, + 0.0019391411915421486, + 0.0018908630590885878, + 0.0023806290701031685, + 0.0024672055151313543, + 0.0031138074118644, + 0.003395865438506007, + 0.002589366864413023, + 0.0018936297856271267, + 0.00267242593690753, + 0.00228428328409791, + 0.0029274129774421453, + 0.002866725903004408, + 0.004282806534320116, + 0.0048226406797766685, + 0.00366187933832407, + 0.0028390756342560053, + 0.015519966371357441, + 0.008832812309265137, + 0.016820039600133896, + 0.01705256849527359, + 0.017969803884625435, + 0.010112687945365906, + 0.014335541985929012, + 0.00996182020753622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.05417373776435852 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828405007719994 + ], + "opposite_logprob_tokens": [ + -3.5298283100128174 + ], + "logprob_reward": -0.029828405007719994, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828405007719994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07530490309000015, + 0.09781523048877716, + 0.2771338224411011, + 0.19274422526359558, + 0.05153666064143181, + 0.08042500913143158, + 0.03228263929486275, + 0.0409093014895916, + 0.05496164783835411, + 0.09688655287027359 + ], + "credit_map": [ + 0.00981885939836502, + 0.015123999677598476, + 0.005725356750190258, + 0.006440119817852974, + 0.015625683590769768, + 0.006412175949662924, + 0.008861742913722992, + 0.007296957075595856, + 0.009261850267648697, + 0.008557289838790894, + 0.004557269159704447, + 0.0039098141714930534, + 0.013493855483829975, + 0.026984304189682007, + 0.007047432474792004, + 0.02400340512394905, + 0.022094476968050003, + 0.03885842487215996, + 0.03352766111493111, + 0.032373543828725815, + 0.047741275280714035, + 0.03300650417804718, + 0.03483794257044792, + 0.03469400107860565, + 0.012253021821379662, + 0.018664173781871796, + 0.013648279011249542, + 0.01749131642282009, + 0.06450743973255157, + 0.022902898490428925, + 0.013225703500211239, + 0.03005138784646988, + 0.004316859412938356, + 0.00962716806679964, + 0.004626171197742224, + 0.004027627874165773, + 0.011016209609806538, + 0.006941355764865875, + 0.006164583843201399, + 0.0048166802152991295, + 0.004526423756033182, + 0.011473147198557854, + 0.007754653226584196, + 0.010575915686786175, + 0.007099212147295475, + 0.00596732459962368, + 0.02043413743376732, + 0.012594190426170826, + 0.0031018091831356287, + 0.00326696434058249, + 0.002709463471546769, + 0.0036430624313652515, + 0.005696811247617006, + 0.004338822793215513, + 0.005579998251050711, + 0.0039457036182284355, + 0.003999983426183462, + 0.004175335168838501, + 0.004705983679741621, + 0.006532995495945215, + 0.004898201674222946, + 0.00583537295460701, + 0.0066821034997701645, + 0.004079325590282679, + 0.005193747580051422, + 0.004604914225637913, + 0.006826090160757303, + 0.006935108453035355, + 0.009105929173529148, + 0.00877960491925478, + 0.00774599751457572, + 0.005770252086222172, + 0.014616092666983604, + 0.010170495137572289, + 0.01606576144695282, + 0.013953321613371372, + 0.014705890789628029, + 0.008564927615225315, + 0.010037029162049294, + 0.008773031644523144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.029828405007719994 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196668475866318 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196668475866318, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196668475866318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1189832016825676, + 0.2900928258895874, + 0.06105724349617958, + 0.04250466078519821, + 0.09055065363645554, + 0.15475919842720032, + 0.04321572557091713, + 0.0796099603176117, + 0.018358740955591202, + 0.10086781531572342 + ], + "credit_map": [ + 0.017405688762664795, + 0.022490529343485832, + 0.0068588582798838615, + 0.00827455148100853, + 0.012179790064692497, + 0.02385663613677025, + 0.015174023807048798, + 0.012743125669658184, + 0.03370514139533043, + 0.026364868506789207, + 0.01347460225224495, + 0.013722391799092293, + 0.02764664776623249, + 0.08115489780902863, + 0.014287121593952179, + 0.07973715662956238, + 0.005488956812769175, + 0.008387573063373566, + 0.009231368079781532, + 0.009139679372310638, + 0.006474878638982773, + 0.004616694059222937, + 0.008029937744140625, + 0.009688155725598335, + 0.0023792407009750605, + 0.004840944893658161, + 0.002661809790879488, + 0.003789973445236683, + 0.015143199823796749, + 0.004618225619196892, + 0.003859566291794181, + 0.0052117034792900085, + 0.007543602492660284, + 0.01449562143534422, + 0.011118673719465733, + 0.012253466062247753, + 0.01518221478909254, + 0.01383223943412304, + 0.009296712465584278, + 0.006828121840953827, + 0.0059168338775634766, + 0.012109678238630295, + 0.007903313264250755, + 0.010997400619089603, + 0.009314946830272675, + 0.00804717093706131, + 0.059554263949394226, + 0.04091557487845421, + 0.004176324233412743, + 0.004602636210620403, + 0.002292041201144457, + 0.0038177194073796272, + 0.008551436476409435, + 0.006971188820898533, + 0.006114039570093155, + 0.006690341047942638, + 0.007932011038064957, + 0.007207169663161039, + 0.009748590178787708, + 0.013133539818227291, + 0.009593171067535877, + 0.00992791261523962, + 0.013229046016931534, + 0.00883851945400238, + 0.001977602019906044, + 0.0015574664575979114, + 0.0017283030319958925, + 0.00217629037797451, + 0.0036804508417844772, + 0.002709107706323266, + 0.0023302563931792974, + 0.0021992637775838375, + 0.012915708124637604, + 0.009672441519796848, + 0.013175904750823975, + 0.023190811276435852, + 0.01330155786126852, + 0.007574374787509441, + 0.015394516289234161, + 0.005642488598823547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.018196668475866318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079574935138226 + ], + "opposite_logprob_tokens": [ + -4.511079788208008 + ], + "logprob_reward": -0.011079574935138226, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079574935138226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817504107952118, + 0.26280754804611206, + 0.05030743405222893, + 0.07488863915205002, + 0.17295542359352112, + 0.041587747633457184, + 0.032894033938646317, + 0.03533655405044556, + 0.02211792953312397, + 0.025354228913784027 + ], + "credit_map": [ + 0.03883222118020058, + 0.04151919484138489, + 0.017218023538589478, + 0.020030835643410683, + 0.02836952544748783, + 0.05779462307691574, + 0.029010292142629623, + 0.04897569864988327, + 0.02249331772327423, + 0.024930639192461967, + 0.009670689702033997, + 0.011503873392939568, + 0.045191146433353424, + 0.060316041111946106, + 0.02550567127764225, + 0.06319618225097656, + 0.003880174132063985, + 0.006722940132021904, + 0.006936944089829922, + 0.006829224526882172, + 0.005906982347369194, + 0.0049914829432964325, + 0.00966043584048748, + 0.005379251204431057, + 0.0041198437102139, + 0.008439469151198864, + 0.003639416303485632, + 0.004353265278041363, + 0.02863049879670143, + 0.007758596446365118, + 0.005599520169198513, + 0.012348026968538761, + 0.013976146467030048, + 0.036426618695259094, + 0.022084945812821388, + 0.021802805364131927, + 0.025014366954565048, + 0.019826283678412437, + 0.018763240426778793, + 0.015061015263199806, + 0.002769776852801442, + 0.007720516063272953, + 0.0026262525934726, + 0.003673277795314789, + 0.00624568248167634, + 0.005730105098336935, + 0.00546178687363863, + 0.007360349874943495, + 0.003287958214059472, + 0.0041084797121584415, + 0.00219698972068727, + 0.002705412684008479, + 0.006616736762225628, + 0.005251889582723379, + 0.004135847557336092, + 0.00459072133526206, + 0.003368303645402193, + 0.0033654950093477964, + 0.004042013082653284, + 0.0037717288359999657, + 0.005540638696402311, + 0.00728917820379138, + 0.004285949282348156, + 0.003673250088468194, + 0.0020782812498509884, + 0.0025096829049289227, + 0.0020998467225581408, + 0.0020908645819872618, + 0.00417056679725647, + 0.003659564536064863, + 0.0035305770579725504, + 0.001978545915335417, + 0.0031972178258001804, + 0.0024321021046489477, + 0.003875614609569311, + 0.004128839820623398, + 0.003241651225835085, + 0.0020099927205592394, + 0.003908803220838308, + 0.0025600064545869827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011079574935138226 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646076545119286 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646076545119286, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646076545119286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18568894267082214, + 0.12342085689306259, + 0.06362853944301605, + 0.03261859714984894, + 0.16695554554462433, + 0.117110475897789, + 0.05672752484679222, + 0.047845881432294846, + 0.06136506795883179, + 0.14463858306407928 + ], + "credit_map": [ + 0.01963181607425213, + 0.023550210520625114, + 0.01304612960666418, + 0.017072798684239388, + 0.018467850983142853, + 0.03320682421326637, + 0.028181007131934166, + 0.032532304525375366, + 0.013393091037869453, + 0.014614153653383255, + 0.008607837371528149, + 0.007552418392151594, + 0.020835325121879578, + 0.023607056587934494, + 0.008379906415939331, + 0.026431068778038025, + 0.006173259112983942, + 0.009273006580770016, + 0.007513143587857485, + 0.008643483743071556, + 0.009199891239404678, + 0.006833304185420275, + 0.009442893788218498, + 0.006549560930579901, + 0.0024782936088740826, + 0.0035801862832158804, + 0.0033216553274542093, + 0.004178253468126059, + 0.008571718819439411, + 0.0036293943412601948, + 0.003190459217876196, + 0.003668638411909342, + 0.021582018584012985, + 0.031439751386642456, + 0.013986661098897457, + 0.01819886453449726, + 0.032412339001894, + 0.02085692249238491, + 0.016007788479328156, + 0.012471196241676807, + 0.004760961513966322, + 0.011850160546600819, + 0.00466943858191371, + 0.006130567751824856, + 0.007385049946606159, + 0.006269362755119801, + 0.04434100165963173, + 0.03170393034815788, + 0.005169394426047802, + 0.006270979531109333, + 0.004084980580955744, + 0.006182396784424782, + 0.0068709030747413635, + 0.006843595765531063, + 0.011777541600167751, + 0.00952773168683052, + 0.00376566918566823, + 0.0033193014096468687, + 0.0060229492373764515, + 0.005548601970076561, + 0.006139745935797691, + 0.007166949566453695, + 0.008730615489184856, + 0.007152046542614698, + 0.006243355572223663, + 0.006345328874886036, + 0.006937381811439991, + 0.006875207647681236, + 0.010739834047853947, + 0.008833060041069984, + 0.007882015779614449, + 0.007508887443691492, + 0.018497364595532417, + 0.017888404428958893, + 0.01828838139772415, + 0.025746576488018036, + 0.015077120624482632, + 0.010191415436565876, + 0.02580617554485798, + 0.0131431445479393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.008646076545119286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a low-crouching first-person perspective, the camera advances in a shaky, handheld motion, cautiously navigating a scene of apocalyptic disaster. The ground is a chaotic mess, littered with powdered cement, twisted rebar, and the charred husks of military equipment. In the immediate foreground, a pair of hands clad in fingerless black tactical gloves can be seen, one pressed against the jagged concrete, the other gripping the butt of a futuristic assault rifle. In the middle distance, a burning vehicle billows thick, oily black smoke, while a gritty, gray dust kicked up by the chaos hangs heavily in the air, obscuring the skeletal remains of bombed-out high-rises against an oppressive, brassy sky. The entire frame periodically shakes with the deep, concussive vibrations of distant impacts, causing the ground to tremble violently and sending small pebbles and debris skittering across its surface. With each jolt, the camera lurches, momentarily losing focus, before steadily reorienting and continuing its slow, deliberate crawl forward. Harsh sunlight struggles to penetrate the thick haze, casting long, distorted shadows and creating sharp, gleaming highlights on shards of glass and discarded bullet casings, creating a desaturated and austere palette that heightens the atmosphere of extreme tension and visceral disorientation.", + "seed": 11015, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_10/index_4092_seed_11015_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": -0.02085722694173455, + "vlm_logprob": -0.02085722694173455, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.02085722694173455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034294504672288895 + ], + "opposite_logprob_tokens": [ + -4.034294605255127 + ], + "logprob_reward": -0.034294504672288895, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.034294504672288895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1975729912519455, + 0.04511209577322006, + 0.05725841224193573, + 0.04195961728692055, + 0.12972840666770935, + 0.03750148043036461, + 0.04940526559948921, + 0.04725750535726547, + 0.35357633233070374, + 0.04062793031334877 + ], + "credit_map": [ + 0.01402177382260561, + 0.02511204592883587, + 0.01971438340842724, + 0.012738347984850407, + 0.06295128166675568, + 0.035567302256822586, + 0.013877535238862038, + 0.013590303249657154, + 0.003833869704976678, + 0.005460880231112242, + 0.008556933142244816, + 0.004148646257817745, + 0.011006456799805164, + 0.006906283088028431, + 0.0030561177991330624, + 0.0021429110784083605, + 0.006847232114523649, + 0.014947373420000076, + 0.007193178869783878, + 0.004750824999064207, + 0.007176362443715334, + 0.008815761655569077, + 0.0038809855468571186, + 0.003646696684882045, + 0.005457574035972357, + 0.006814803462475538, + 0.0054293167777359486, + 0.0042009493336081505, + 0.0070704803802073, + 0.005379647947847843, + 0.003796890377998352, + 0.0038099545054137707, + 0.014602826908230782, + 0.026433829218149185, + 0.015360413119196892, + 0.025132544338703156, + 0.013149192556738853, + 0.01606443151831627, + 0.010498253628611565, + 0.00848690327256918, + 0.0033255512826144695, + 0.005119561683386564, + 0.005449145566672087, + 0.006033264100551605, + 0.003985365852713585, + 0.0036900590639561415, + 0.005155140999704599, + 0.004743389319628477, + 0.00549189280718565, + 0.007971346378326416, + 0.006566640455275774, + 0.007005883846431971, + 0.007370302453637123, + 0.0035546438302844763, + 0.006510062608867884, + 0.004934492986649275, + 0.004605673253536224, + 0.006113389972597361, + 0.00496744504198432, + 0.0063575939275324345, + 0.00823301263153553, + 0.006944621447473764, + 0.006058119237422943, + 0.0039776465855538845, + 0.01833331026136875, + 0.13929717242717743, + 0.12117411196231842, + 0.01867343671619892, + 0.02884533256292343, + 0.01431318186223507, + 0.006305394694209099, + 0.006634391378611326, + 0.004301887471228838, + 0.0056936247274279594, + 0.007025071419775486, + 0.006122404709458351, + 0.005676924716681242, + 0.005247536115348339, + 0.003624465549364686, + 0.0029360137414187193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.034294504672288895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372691571712494 + ], + "opposite_logprob_tokens": [ + -3.408726930618286 + ], + "logprob_reward": -0.03372691571712494, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372691571712494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21396759152412415, + 0.03324735537171364, + 0.05783621594309807, + 0.05538498982787132, + 0.26861053705215454, + 0.06456804275512695, + 0.04316575080156326, + 0.036894239485263824, + 0.11635129153728485, + 0.10997392237186432 + ], + "credit_map": [ + 0.02050873078405857, + 0.03370096534490585, + 0.029592014849185944, + 0.019611872732639313, + 0.044125035405159, + 0.029628397896885872, + 0.021580304950475693, + 0.015220284461975098, + 0.003515825141221285, + 0.004873685538768768, + 0.004889556206762791, + 0.0026974910870194435, + 0.005521747283637524, + 0.004523498937487602, + 0.00406144792214036, + 0.0031641069799661636, + 0.003838417585939169, + 0.015504274517297745, + 0.003556037787348032, + 0.0034559275954961777, + 0.007161235902458429, + 0.018459340557456017, + 0.002670137444511056, + 0.0031908464152365923, + 0.007078239694237709, + 0.007047944702208042, + 0.007681185379624367, + 0.007981515489518642, + 0.007313677109777927, + 0.007806210778653622, + 0.004968347027897835, + 0.005507874768227339, + 0.030469877645373344, + 0.05723387002944946, + 0.03104524128139019, + 0.05118279531598091, + 0.02668275311589241, + 0.03227388486266136, + 0.021751252934336662, + 0.017970852553844452, + 0.007300048600882292, + 0.010442768223583698, + 0.007830965332686901, + 0.009776773862540722, + 0.009643885307013988, + 0.007154673337936401, + 0.00593578303232789, + 0.006483155302703381, + 0.004777936264872551, + 0.006586494389921427, + 0.005114039871841669, + 0.005442514084279537, + 0.008660384453833103, + 0.0034308626782149076, + 0.005209117662161589, + 0.003944404423236847, + 0.0039670770056545734, + 0.004750936292111874, + 0.004442467354238033, + 0.00537679111585021, + 0.006502216216176748, + 0.0043177613988518715, + 0.0038116464857012033, + 0.0037253443151712418, + 0.007885250262916088, + 0.039130281656980515, + 0.03145328164100647, + 0.008764446713030338, + 0.009814238175749779, + 0.008205000311136246, + 0.005617555230855942, + 0.005481242202222347, + 0.011054097674787045, + 0.015800155699253082, + 0.013754736632108688, + 0.01815183274447918, + 0.012868820689618587, + 0.016558686271309853, + 0.013139178976416588, + 0.008646422065794468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.03372691571712494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362566277384758 + ], + "opposite_logprob_tokens": [ + -3.651362657546997 + ], + "logprob_reward": -0.026362566277384758, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362566277384758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20447148382663727, + 0.026390407234430313, + 0.033925577998161316, + 0.05015477165579796, + 0.05881476029753685, + 0.04243699461221695, + 0.029034670442342758, + 0.030946891754865646, + 0.4970729351043701, + 0.026751505210995674 + ], + "credit_map": [ + 0.015811847522854805, + 0.029454117640852928, + 0.0170110035687685, + 0.015677060931921005, + 0.06180444732308388, + 0.034953273832798004, + 0.016152722761034966, + 0.013607006520032883, + 0.002646283246576786, + 0.004502515308558941, + 0.0036621992476284504, + 0.00315878982655704, + 0.004286111798137426, + 0.002792000537738204, + 0.0031695233192294836, + 0.002172984415665269, + 0.003838403383269906, + 0.007788693066686392, + 0.005097606219351292, + 0.003042684867978096, + 0.0033592316322028637, + 0.006247181911021471, + 0.0022991651203483343, + 0.002252608770504594, + 0.004721940029412508, + 0.006084946915507317, + 0.006383984815329313, + 0.005692408885806799, + 0.00899791531264782, + 0.006443945690989494, + 0.0057656834833323956, + 0.006063942331820726, + 0.004877434112131596, + 0.010960057377815247, + 0.004764340352267027, + 0.006030841264873743, + 0.010844847187399864, + 0.006996141746640205, + 0.006824724841862917, + 0.007516373880207539, + 0.0046175094321370125, + 0.006445315666496754, + 0.005578942131251097, + 0.006430396344512701, + 0.005149342119693756, + 0.004729813430458307, + 0.005221057683229446, + 0.004264616873115301, + 0.0033662267960608006, + 0.004819558467715979, + 0.003634075401350856, + 0.004815280437469482, + 0.0047494834288954735, + 0.002375534735620022, + 0.003523975843563676, + 0.0017505348660051823, + 0.0034303050488233566, + 0.004171133507043123, + 0.003304310841485858, + 0.004492194391787052, + 0.00585008691996336, + 0.004066601861268282, + 0.002666871529072523, + 0.002965386491268873, + 0.022402601316571236, + 0.2007613629102707, + 0.17115628719329834, + 0.026149854063987732, + 0.04120917245745659, + 0.01929416134953499, + 0.007979301735758781, + 0.008120168000459671, + 0.002942450111731887, + 0.00480849202722311, + 0.003042232245206833, + 0.003274535993114114, + 0.0042408122681081295, + 0.0033549547661095858, + 0.002742605982348323, + 0.002345422049984336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.026362566277384758 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257708020508289 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257708020508289, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257708020508289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09332634508609772, + 0.01756829209625721, + 0.0911141186952591, + 0.025323398411273956, + 0.029895566403865814, + 0.023969516158103943, + 0.03367278724908829, + 0.07026172429323196, + 0.5841900706291199, + 0.030678093433380127 + ], + "credit_map": [ + 0.007090359460562468, + 0.01485373079776764, + 0.006059904117137194, + 0.007624783553183079, + 0.02820020727813244, + 0.015690859407186508, + 0.007562198676168919, + 0.006244299467653036, + 0.001439934130758047, + 0.002830970799550414, + 0.0032162668649107218, + 0.0017851950833573937, + 0.003427633084356785, + 0.0024093661922961473, + 0.001563844969496131, + 0.000895080913323909, + 0.005843631457537413, + 0.02615271881222725, + 0.003585733473300934, + 0.0041209496557712555, + 0.012995071709156036, + 0.032806508243083954, + 0.0031642818357795477, + 0.0024452218785881996, + 0.0026148189790546894, + 0.004115441348403692, + 0.003335688030347228, + 0.0026019373908638954, + 0.003386094933375716, + 0.003385227872058749, + 0.0029839612543582916, + 0.002900226740166545, + 0.0053959134966135025, + 0.006916028913110495, + 0.003294906346127391, + 0.0032437308691442013, + 0.003065024269744754, + 0.003414595266804099, + 0.00213428377173841, + 0.0024310823064297438, + 0.0023161459248512983, + 0.003533637383952737, + 0.004452589433640242, + 0.002908741356804967, + 0.004045517649501562, + 0.003637343179434538, + 0.0016652169870212674, + 0.0014103236608207226, + 0.004491709638386965, + 0.005327981896698475, + 0.004150340333580971, + 0.004956834949553013, + 0.005170648917555809, + 0.0034054189454764128, + 0.0035277416463941336, + 0.0026421132497489452, + 0.00856101606041193, + 0.01110762171447277, + 0.0076752654276788235, + 0.00958629697561264, + 0.012835063971579075, + 0.008223497308790684, + 0.006742643658071756, + 0.005530320107936859, + 0.02679910697042942, + 0.2354525923728943, + 0.19788144528865814, + 0.03089367039501667, + 0.04886028915643692, + 0.02281336672604084, + 0.01023742277175188, + 0.011252200230956078, + 0.002328583039343357, + 0.0055647725239396095, + 0.0041865017265081406, + 0.003440614091232419, + 0.004358280450105667, + 0.0038379409816116095, + 0.004382240120321512, + 0.002579160500317812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.005257708020508289 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 low-crouching first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a pair of hands wearing fingerless black tactical gloves visible in the immediate foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a futuristic assault rifle visible, with one hand gripping its butt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene include a burning vehicle billowing thick, oily black smoke in the middle distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera advance with a shaky, handheld motion that lurches and loses focus periodically due to vibrations?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04777834936976433, + 0.021109627559781075, + 0.023548807948827744, + 0.0407196544110775, + 0.06541233509778976, + 0.0595095194876194, + 0.029004856944084167, + 0.015252958051860332, + 0.6616342663764954, + 0.036029599606990814 + ], + "credit_map": [ + 0.004259140230715275, + 0.0066032386384904385, + 0.005680897273123264, + 0.003681568894535303, + 0.012384461238980293, + 0.007399259600788355, + 0.003462748369202018, + 0.004307032562792301, + 0.001964595867320895, + 0.002912776079028845, + 0.0039471169002354145, + 0.0028371186926960945, + 0.004023026674985886, + 0.0030844006687402725, + 0.0014751872513443232, + 0.0008654030389152467, + 0.0031842973548918962, + 0.00407141400501132, + 0.0029382179491221905, + 0.0025497872848063707, + 0.0030484190210700035, + 0.003522561863064766, + 0.002469312399625778, + 0.0017647967906668782, + 0.004847866948693991, + 0.006447609048336744, + 0.005600421689450741, + 0.004998138640075922, + 0.0068857185542583466, + 0.005010665860027075, + 0.004332343582063913, + 0.0025968842674046755, + 0.007305886130779982, + 0.015132305212318897, + 0.007969461381435394, + 0.011785050854086876, + 0.006991556845605373, + 0.007660758215934038, + 0.004390139598399401, + 0.0041771684773266315, + 0.0049002086743712425, + 0.005842703394591808, + 0.009782062843441963, + 0.00792081467807293, + 0.009114217013120651, + 0.005956095177680254, + 0.009723746217787266, + 0.006269664037972689, + 0.0035683447495102882, + 0.0047554983757436275, + 0.0034183254465460777, + 0.0035234636161476374, + 0.005768807139247656, + 0.0030001963023096323, + 0.0029871631413698196, + 0.001983056776225567, + 0.0020402977243065834, + 0.0018392482306808233, + 0.001709959004074335, + 0.0014474994968622923, + 0.002502607414498925, + 0.002699097152799368, + 0.0016342508606612682, + 0.0013799975859001279, + 0.030812440440058708, + 0.26528599858283997, + 0.22970570623874664, + 0.034894149750471115, + 0.05397754907608032, + 0.025623291730880737, + 0.010511106811463833, + 0.010823989287018776, + 0.0036827789153903723, + 0.006228428799659014, + 0.00434622447937727, + 0.004338480997830629, + 0.0057584564201533794, + 0.005866619758307934, + 0.0034834621474146843, + 0.0023251459933817387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.004644440021365881 + } + ], + "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/0000055/group_11/index_tensor([872, 872])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_tensor([872, 872])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..680c2f8ffdc043b47f6011351a8dd15dbaf32053 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_tensor([872, 872])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle close-up remains static, focused on a portion of a dark, polished mahogany tabletop, where a single, warm light source from an off-camera lamp casts a focused golden glow. Within this pool of light, two hands take center stage. One hand, belonging to an elderly individual, features prominent veins and neatly manicured nails, resting palm-down. Its index and middle fingers press gently yet firmly against the edge of a neatly stacked stack of off-white legal documents, held together by a simple metal clip. With a single, steady, deliberate movement, the hand gently pushes the entire sheet across the smooth, reflective wood, the paper sliding past with a soft, whispering sound. The other hand, younger, with smoother skin and slender fingers, rests motionless on the table, fingers slightly curled, in a receptive stillness as the paper slides toward it. The scene is imbued with a quiet, tense atmosphere, the warm, low-key lighting accentuating the subtle textures of skin, paper, and texture, transforming a simple transfer into a gesture of profound meaning and unspoken consent.", + "seed": 4375, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_872_seed_4375_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": -0.0389528397237882, + "vlm_logprob": -0.0389528397237882, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.0389528397237882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024152105674147606 + ], + "opposite_logprob_tokens": [ + -5.1491522789001465 + ], + "logprob_reward": -0.024152105674147606, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.024152105674147606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10520532727241516, + 0.13251887261867523, + 0.06225869804620743, + 0.031141206622123718, + 0.15569856762886047, + 0.0678953304886818, + 0.047709666192531586, + 0.11775606870651245, + 0.1291012018918991, + 0.15071505308151245 + ], + "credit_map": [ + 0.014366146177053452, + 0.009564491920173168, + 0.012948071584105492, + 0.010646145790815353, + 0.016386641189455986, + 0.018591968342661858, + 0.005971736740320921, + 0.01673012785613537, + 0.02336137183010578, + 0.011602208949625492, + 0.02225027047097683, + 0.015330314636230469, + 0.011964162811636925, + 0.01776762865483761, + 0.011978638358414173, + 0.01826428435742855, + 0.009712036699056625, + 0.005004362668842077, + 0.01109042577445507, + 0.00928422436118126, + 0.006379158701747656, + 0.006422901060432196, + 0.005136238876730204, + 0.00922934990376234, + 0.0044441139325499535, + 0.0035541895776987076, + 0.003848147578537464, + 0.0037171358708292246, + 0.00330616207793355, + 0.004587939474731684, + 0.0028790654614567757, + 0.004804450552910566, + 0.021468618884682655, + 0.010189213789999485, + 0.027675403282046318, + 0.03147006407380104, + 0.012122257612645626, + 0.012748293578624725, + 0.0191985834389925, + 0.020826127380132675, + 0.007658559828996658, + 0.005901722237467766, + 0.009352742694318295, + 0.0034007809590548277, + 0.01198587752878666, + 0.014000166207551956, + 0.0037167079281061888, + 0.01187877170741558, + 0.009935502894222736, + 0.00556509755551815, + 0.0047286152839660645, + 0.005190688651055098, + 0.0033993704710155725, + 0.007897540926933289, + 0.004982152488082647, + 0.006010700482875109, + 0.013004401698708534, + 0.018006527796387672, + 0.01296388078480959, + 0.019935142248868942, + 0.021934013813734055, + 0.01615387201309204, + 0.007915081456303596, + 0.007843141444027424, + 0.0220046229660511, + 0.020015830174088478, + 0.00923252198845148, + 0.011192092671990395, + 0.024383315816521645, + 0.018161745741963387, + 0.008882617577910423, + 0.015228440053761005, + 0.020679069682955742, + 0.013895252719521523, + 0.022632285952568054, + 0.018454883247613907, + 0.017383597791194916, + 0.017741648480296135, + 0.021838488057255745, + 0.01808982715010643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.024152105674147606 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111647140234709 + ], + "opposite_logprob_tokens": [ + -5.5041117668151855 + ], + "logprob_reward": -0.004111647140234709, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111647140234709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10077578574419022, + 0.22452662885189056, + 0.03332981467247009, + 0.02162378467619419, + 0.1739320605993271, + 0.05561669543385506, + 0.07161450386047363, + 0.04762604460120201, + 0.055027443915605545, + 0.21592721343040466 + ], + "credit_map": [ + 0.012377363629639149, + 0.0067065623588860035, + 0.011515635997056961, + 0.011267035268247128, + 0.012314350344240665, + 0.022052032873034477, + 0.007420911453664303, + 0.017121899873018265, + 0.02212379314005375, + 0.015193537808954716, + 0.04834252595901489, + 0.04932635277509689, + 0.019740330055356026, + 0.024695249274373055, + 0.016466429457068443, + 0.028638411313295364, + 0.0060201967135071754, + 0.002819010289385915, + 0.006295038387179375, + 0.005504818167537451, + 0.0026782508939504623, + 0.0023364676162600517, + 0.0028618189971894026, + 0.0048142122104763985, + 0.0021143765188753605, + 0.001974984770640731, + 0.0030498220585286617, + 0.0037706196308135986, + 0.0022064847871661186, + 0.003441113280132413, + 0.002050940180197358, + 0.003015444613993168, + 0.0302712544798851, + 0.00950497668236494, + 0.034553274512290955, + 0.03557835519313812, + 0.010414506308734417, + 0.015062123537063599, + 0.013216730207204819, + 0.02533085271716118, + 0.005056053400039673, + 0.004965926520526409, + 0.006250496953725815, + 0.004181314259767532, + 0.0053515746258199215, + 0.013333885930478573, + 0.008761084638535976, + 0.007716355379670858, + 0.016490057110786438, + 0.00799939688295126, + 0.007972154766321182, + 0.006542541552335024, + 0.005904132500290871, + 0.012758713215589523, + 0.0069640642032027245, + 0.006983441766351461, + 0.006298926658928394, + 0.006664234679192305, + 0.0054211788810789585, + 0.007384131196886301, + 0.007702929433435202, + 0.005908628460019827, + 0.0036060546990484, + 0.004639961291104555, + 0.005243383347988129, + 0.00497051514685154, + 0.008802820928394794, + 0.006572522222995758, + 0.009354195557534695, + 0.0057409354485571384, + 0.004566830117255449, + 0.009776240214705467, + 0.026452170684933662, + 0.020922444760799408, + 0.04384208470582962, + 0.035297609865665436, + 0.021172253414988518, + 0.020726798102259636, + 0.029699139297008514, + 0.017814703285694122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 4 + ], + "selected_reward": -0.004111647140234709 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311134427785873 + ], + "opposite_logprob_tokens": [ + -3.773311138153076 + ], + "logprob_reward": -0.023311134427785873, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311134427785873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09794779121875763, + 0.17384959757328033, + 0.08218348771333694, + 0.13156448304653168, + 0.11091510206460953, + 0.21087723970413208, + 0.0706976056098938, + 0.04030218347907066, + 0.03408120572566986, + 0.047581251710653305 + ], + "credit_map": [ + 0.009289314970374107, + 0.007603647653013468, + 0.01460876315832138, + 0.012835167348384857, + 0.014602459967136383, + 0.017040401697158813, + 0.007274908013641834, + 0.01469313632696867, + 0.013983634300529957, + 0.012175978161394596, + 0.028954053297638893, + 0.030310137197375298, + 0.015821730718016624, + 0.03229714557528496, + 0.017013566568493843, + 0.023293351754546165, + 0.013811235316097736, + 0.00743557745590806, + 0.014927227981388569, + 0.011727962642908096, + 0.006882290355861187, + 0.008120434358716011, + 0.007892237044870853, + 0.011386524885892868, + 0.016057120636105537, + 0.009786332957446575, + 0.014114690944552422, + 0.014199876226484776, + 0.01670260541141033, + 0.028854141011834145, + 0.009062042459845543, + 0.022787680849432945, + 0.015276946127414703, + 0.006342997308820486, + 0.024895867332816124, + 0.020032288506627083, + 0.007092780899256468, + 0.008618650957942009, + 0.013114611618220806, + 0.015540962107479572, + 0.02252483181655407, + 0.013839982450008392, + 0.034963712096214294, + 0.010008437559008598, + 0.026970859616994858, + 0.040863197296857834, + 0.010170867666602135, + 0.05153534933924675, + 0.016335463151335716, + 0.009577900171279907, + 0.007601167540997267, + 0.00638539157807827, + 0.005304660648107529, + 0.01353072002530098, + 0.0035915763583034277, + 0.008370723575353622, + 0.008875742554664612, + 0.005748230963945389, + 0.00336121185682714, + 0.004595628939568996, + 0.00538642518222332, + 0.006296409759670496, + 0.0025140189100056887, + 0.003524513216689229, + 0.0034351127687841654, + 0.0025566802360117435, + 0.007126553915441036, + 0.004813851788640022, + 0.0038387926761060953, + 0.0026967099402099848, + 0.003220704384148121, + 0.006392796989530325, + 0.004842991940677166, + 0.005716527812182903, + 0.008790519088506699, + 0.007944324985146523, + 0.004016756080091, + 0.005693354178220034, + 0.0060189999639987946, + 0.004557779990136623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.023311134427785873 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284509420394897 + ], + "opposite_logprob_tokens": [ + -2.0178451538085938 + ], + "logprob_reward": -0.14284509420394897, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284509420394897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15181314945220947, + 0.23026548326015472, + 0.04717414453625679, + 0.05447209253907204, + 0.19343198835849762, + 0.041704557836055756, + 0.11674338579177856, + 0.051110804080963135, + 0.061861466616392136, + 0.051422882825136185 + ], + "credit_map": [ + 0.012454358860850334, + 0.010933849029242992, + 0.02803908847272396, + 0.026656584814190865, + 0.015172227285802364, + 0.027593472972512245, + 0.010274864733219147, + 0.020688705146312714, + 0.022682756185531616, + 0.01808135211467743, + 0.03186071664094925, + 0.016245845705270767, + 0.030874913558363914, + 0.045353516936302185, + 0.030515098944306374, + 0.03465129807591438, + 0.008369292132556438, + 0.003961742855608463, + 0.006507402285933495, + 0.00689679617062211, + 0.0052700587548315525, + 0.00585947697982192, + 0.0029327881056815386, + 0.007376587484031916, + 0.006284609436988831, + 0.006450873799622059, + 0.00615253672003746, + 0.004116484429687262, + 0.007216005120426416, + 0.010623558424413204, + 0.004295186139643192, + 0.00933284405618906, + 0.020321309566497803, + 0.014777389355003834, + 0.02000325545668602, + 0.018326636403799057, + 0.023311970755457878, + 0.023211365565657616, + 0.04223267361521721, + 0.031247399747371674, + 0.00476655550301075, + 0.002955676754936576, + 0.004827891010791063, + 0.0021731664892286062, + 0.004786043893545866, + 0.007206048350781202, + 0.002306935843080282, + 0.01268224511295557, + 0.015767555683851242, + 0.010068963281810284, + 0.013013077899813652, + 0.013051806949079037, + 0.0076389373280107975, + 0.03498299419879913, + 0.008007269352674484, + 0.014212789945304394, + 0.006372142117470503, + 0.007832207717001438, + 0.005894964095205069, + 0.009259283542633057, + 0.007265844848006964, + 0.006029049400240183, + 0.004414469934999943, + 0.0040428475476801395, + 0.008809982798993587, + 0.007154247257858515, + 0.008332624100148678, + 0.009272110648453236, + 0.010649098083376884, + 0.00748610682785511, + 0.005634028930217028, + 0.004523268900811672, + 0.004144411534070969, + 0.006956109777092934, + 0.005452895071357489, + 0.0052495235577225685, + 0.006116875447332859, + 0.0072264340706169605, + 0.007341222837567329, + 0.008935419842600822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.14284509420394897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00034421717282384634 + ], + "opposite_logprob_tokens": [ + -8.000344276428223 + ], + "logprob_reward": -0.00034421717282384634, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.00034421717282384634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15578320622444153, + 0.16211318969726562, + 0.033856313675642014, + 0.022037958726286888, + 0.11871350556612015, + 0.058380451053380966, + 0.11492226272821426, + 0.1154923364520073, + 0.07514365762472153, + 0.14355716109275818 + ], + "credit_map": [ + 0.02089199796319008, + 0.010943006724119186, + 0.027871420606970787, + 0.028591345995664597, + 0.012830236926674843, + 0.02074599824845791, + 0.012103009037673473, + 0.021806180477142334, + 0.014282133430242538, + 0.011683455668389797, + 0.03138834238052368, + 0.039835549890995026, + 0.013725045137107372, + 0.013362004421651363, + 0.019331257790327072, + 0.0185054000467062, + 0.004931942094117403, + 0.004822074435651302, + 0.0032071261666715145, + 0.003330421634018421, + 0.00523604778572917, + 0.004213208332657814, + 0.0025841540191322565, + 0.005531339440494776, + 0.0031081358902156353, + 0.0027923528105020523, + 0.0024344250559806824, + 0.002239219844341278, + 0.004264262039214373, + 0.002191771985962987, + 0.001784532330930233, + 0.0032232587691396475, + 0.012518452480435371, + 0.007469726726412773, + 0.023321310058236122, + 0.01695449836552143, + 0.010159128345549107, + 0.01207334641367197, + 0.018332913517951965, + 0.01788412779569626, + 0.005572427995502949, + 0.004790015984326601, + 0.008835055865347385, + 0.003411410376429558, + 0.0043004173785448074, + 0.010280258022248745, + 0.003083892399445176, + 0.0181069765239954, + 0.03498891741037369, + 0.013745391741394997, + 0.011433783918619156, + 0.009289867244660854, + 0.009800040163099766, + 0.01894388347864151, + 0.00807100534439087, + 0.008649366907775402, + 0.013629350811243057, + 0.019944244995713234, + 0.01603027991950512, + 0.020358489826321602, + 0.011827096343040466, + 0.014702343381941319, + 0.00952757615596056, + 0.009472958743572235, + 0.00968120712786913, + 0.009971044026315212, + 0.00851721502840519, + 0.011994496919214725, + 0.010440760292112827, + 0.010040810331702232, + 0.007182150613516569, + 0.007315979804843664, + 0.020640037953853607, + 0.011741172522306442, + 0.016696836799383163, + 0.020727435126900673, + 0.018384262919425964, + 0.015185906551778316, + 0.02262250706553459, + 0.01755901798605919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00034421717282384634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up remains static, focused on a portion of a dark, polished mahogany tabletop, where a single, warm light source from an off-camera lamp casts a focused golden glow. Within this pool of light, two hands take center stage. One hand, belonging to an elderly individual, features prominent veins and neatly manicured nails, resting palm-down. Its index and middle fingers press gently yet firmly against the edge of a neatly stacked stack of off-white legal documents, held together by a simple metal clip. With a single, steady, deliberate movement, the hand gently pushes the entire sheet across the smooth, reflective wood, the paper sliding past with a soft, whispering sound. The other hand, younger, with smoother skin and slender fingers, rests motionless on the table, fingers slightly curled, in a receptive stillness as the paper slides toward it. The scene is imbued with a quiet, tense atmosphere, the warm, low-key lighting accentuating the subtle textures of skin, paper, and texture, transforming a simple transfer into a gesture of profound meaning and unspoken consent.", + "seed": 4376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_872_seed_4376_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": -0.03659297479316592, + "vlm_logprob": -0.03659297479316592, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.03659297479316592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022202610969543457 + ], + "opposite_logprob_tokens": [ + -4.522202491760254 + ], + "logprob_reward": -0.022202610969543457, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.022202610969543457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1276315450668335, + 0.4033781886100769, + 0.03514637425541878, + 0.015404262579977512, + 0.12383077293634415, + 0.050016216933727264, + 0.05510207265615463, + 0.07411357760429382, + 0.0716620534658432, + 0.043714966624975204 + ], + "credit_map": [ + 0.01632636971771717, + 0.010178319178521633, + 0.025037774816155434, + 0.0235638115555048, + 0.014855829067528248, + 0.013016595505177975, + 0.011902698315680027, + 0.012750150635838509, + 0.0273980051279068, + 0.016489438712596893, + 0.12762059271335602, + 0.03591389209032059, + 0.027973711490631104, + 0.04564167559146881, + 0.015569056384265423, + 0.10677181929349899, + 0.0030104711186140776, + 0.003018041606992483, + 0.00474906899034977, + 0.002531250473111868, + 0.006111702881753445, + 0.007906440645456314, + 0.003962238784879446, + 0.003857158124446869, + 0.001204655272886157, + 0.001128573203459382, + 0.0018492024391889572, + 0.0020577663090080023, + 0.0021856706589460373, + 0.0022103183437138796, + 0.0025318313855677843, + 0.002236244035884738, + 0.009055368602275848, + 0.006580293644219637, + 0.00997145939618349, + 0.014753655530512333, + 0.010011098347604275, + 0.028109246864914894, + 0.020949658006429672, + 0.024400008842349052, + 0.006054964382201433, + 0.005390145815908909, + 0.006333198864012957, + 0.009268300607800484, + 0.0037988892290741205, + 0.00598580576479435, + 0.004783228971064091, + 0.008401686325669289, + 0.003890736261382699, + 0.005502517335116863, + 0.00728862127289176, + 0.009776377119123936, + 0.007133591920137405, + 0.0072794207371771336, + 0.010588872246444225, + 0.0036419364623725414, + 0.006042864639312029, + 0.006971321534365416, + 0.015957225114107132, + 0.013610562309622765, + 0.00708372425287962, + 0.004859004635363817, + 0.00838813278824091, + 0.011200745590031147, + 0.008185028098523617, + 0.006581160705536604, + 0.012520299293100834, + 0.009400510229170322, + 0.009978876449167728, + 0.005308065097779036, + 0.008133878000080585, + 0.011554236523807049, + 0.004268210846930742, + 0.004097056575119495, + 0.008311967365443707, + 0.007107105106115341, + 0.0050564175471663475, + 0.002997979521751404, + 0.005342880263924599, + 0.0065333484672009945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.022202610969543457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645274978131056 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.007645274978131056, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645274978131056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15186351537704468, + 0.3518346846103668, + 0.03980134800076485, + 0.03981783613562584, + 0.19523371756076813, + 0.031053565442562103, + 0.02478843182325363, + 0.08810863643884659, + 0.040501561015844345, + 0.03699661046266556 + ], + "credit_map": [ + 0.01704346388578415, + 0.01773146726191044, + 0.012232236564159393, + 0.03430576249957085, + 0.019811341539025307, + 0.014479435048997402, + 0.012627778574824333, + 0.02363203652203083, + 0.023398518562316895, + 0.012240748852491379, + 0.1169208213686943, + 0.03680337965488434, + 0.036326710134744644, + 0.03546833246946335, + 0.010998628102242947, + 0.07967754453420639, + 0.0043536084704101086, + 0.003277824493125081, + 0.0037267704028636217, + 0.005257393699139357, + 0.005536745302379131, + 0.007876444607973099, + 0.0036919005215168, + 0.006080659106373787, + 0.00442923977971077, + 0.004981885198503733, + 0.0023741882760077715, + 0.006331354845315218, + 0.004185935482382774, + 0.00762730510905385, + 0.004844626411795616, + 0.005043300334364176, + 0.009366455487906933, + 0.008007018826901913, + 0.005012740846723318, + 0.007102950010448694, + 0.011012041941285133, + 0.07618355005979538, + 0.014478722587227821, + 0.06407024711370468, + 0.005097659770399332, + 0.004356906749308109, + 0.0037041057366877794, + 0.0037274975329637527, + 0.003005306702107191, + 0.004096767399460077, + 0.0028586520347744226, + 0.004206672310829163, + 0.002278925385326147, + 0.0025268401950597763, + 0.003504873253405094, + 0.00504910247400403, + 0.002532361540943384, + 0.003571131732314825, + 0.0038123445119708776, + 0.0015128529630601406, + 0.006347516551613808, + 0.007330748718231916, + 0.01933283545076847, + 0.021480271592736244, + 0.005042996723204851, + 0.00535719795152545, + 0.010388062335550785, + 0.012829003855586052, + 0.00397019786760211, + 0.004753983113914728, + 0.0064317225478589535, + 0.005487962160259485, + 0.003822687081992626, + 0.0033756771590560675, + 0.005318580660969019, + 0.007340752054005861, + 0.0030427833553403616, + 0.0034381907898932695, + 0.008355552330613136, + 0.005484933964908123, + 0.0032280576415359974, + 0.003251285059377551, + 0.004695114679634571, + 0.005500693339854479 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.007645274978131056 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797697886824608 + ], + "opposite_logprob_tokens": [ + -4.6347975730896 + ], + "logprob_reward": -0.009797697886824608, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797697886824608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03706500306725502, + 0.11645802110433578, + 0.05936472862958908, + 0.010650035925209522, + 0.6937180161476135, + 0.010653151199221611, + 0.014389652758836746, + 0.019711501896381378, + 0.0297185480594635, + 0.008271344937384129 + ], + "credit_map": [ + 0.00585172139108181, + 0.004170364700257778, + 0.005346484482288361, + 0.0044940016232430935, + 0.005637148395180702, + 0.0030851298943161964, + 0.004028476309031248, + 0.004451673477888107, + 0.0076968916691839695, + 0.004161751363426447, + 0.03809495270252228, + 0.012530895881354809, + 0.010443249717354774, + 0.012246541678905487, + 0.0037924794014543295, + 0.0274912528693676, + 0.006789489183574915, + 0.006103506777435541, + 0.005040431395173073, + 0.005411754362285137, + 0.0073136924766004086, + 0.013060194440186024, + 0.0041937874630093575, + 0.01145186647772789, + 0.0016017763409763575, + 0.0011978812981396914, + 0.0007949282298795879, + 0.0012383662397041917, + 0.0015205633826553822, + 0.0019602086395025253, + 0.0010914955055341125, + 0.0012448164634406567, + 0.017895521596074104, + 0.016583546996116638, + 0.010089553892612457, + 0.016791414469480515, + 0.03782479465007782, + 0.306884229183197, + 0.026278547942638397, + 0.26137039065361023, + 0.0008753553847782314, + 0.000766575220040977, + 0.0012803005520254374, + 0.0021626341622322798, + 0.0007610971224494278, + 0.0020093282219022512, + 0.0011034751078113914, + 0.0016943849623203278, + 0.0015022202860563993, + 0.0014014022890478373, + 0.002493423642590642, + 0.002790226135402918, + 0.0011581009021028876, + 0.0018916110275313258, + 0.0024161715991795063, + 0.0007364957127720118, + 0.0015948355430737138, + 0.0022926260717213154, + 0.003027004422619939, + 0.0032061729580163956, + 0.0014865685952827334, + 0.0015308178262785077, + 0.0034471568651497364, + 0.003126319497823715, + 0.003305725520476699, + 0.002670570509508252, + 0.004977386444807053, + 0.0037891333922743797, + 0.004626034758985043, + 0.002142219338566065, + 0.0034745358861982822, + 0.0047329384833574295, + 0.0008753560832701623, + 0.0007467413088306785, + 0.0014124769950285554, + 0.00119037926197052, + 0.001095535117201507, + 0.0006317752413451672, + 0.0009879636345431209, + 0.0013311172369867563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.009797697886824608 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428235024213791 + ], + "opposite_logprob_tokens": [ + -2.0178234577178955 + ], + "logprob_reward": -0.1428235024213791, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428235024213791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042568597942590714, + 0.1008838340640068, + 0.03671341389417648, + 0.014153361320495605, + 0.5551589131355286, + 0.046228062361478806, + 0.03136514127254486, + 0.10195325314998627, + 0.05041908845305443, + 0.020556285977363586 + ], + "credit_map": [ + 0.00688506243750453, + 0.004533988889306784, + 0.003810869064182043, + 0.008810977451503277, + 0.003685042494907975, + 0.0037182262167334557, + 0.005243838299065828, + 0.005880595650523901, + 0.006950754206627607, + 0.003391775768250227, + 0.031790684908628464, + 0.010328474454581738, + 0.016398223116993904, + 0.010861704126000404, + 0.002923283027485013, + 0.01823892816901207, + 0.0036013575736433268, + 0.003616822650656104, + 0.0063818092457950115, + 0.0029303745832294226, + 0.0049369195476174355, + 0.004786364734172821, + 0.0035294913686811924, + 0.0069302748888731, + 0.002079887082800269, + 0.0013769642682746053, + 0.0011515178484842181, + 0.001833436544984579, + 0.0019095046445727348, + 0.0023430685978382826, + 0.0016517677577212453, + 0.0018072144594043493, + 0.017886817455291748, + 0.01640169695019722, + 0.01631208322942257, + 0.0247227530926466, + 0.033658161759376526, + 0.21700848639011383, + 0.04661739617586136, + 0.18255148828029633, + 0.004933733027428389, + 0.005796547047793865, + 0.006098151672631502, + 0.007684305775910616, + 0.004483803175389767, + 0.007154867518693209, + 0.003972156438976526, + 0.0061044953763484955, + 0.0025150435976684093, + 0.003667708020657301, + 0.004878632724285126, + 0.0054746284149587154, + 0.00346179585903883, + 0.003687974764034152, + 0.004959606099873781, + 0.0027197522576898336, + 0.010107007808983326, + 0.013491010293364525, + 0.015383526682853699, + 0.016034655272960663, + 0.009295331314206123, + 0.0075898850336670876, + 0.01572677306830883, + 0.014325068332254887, + 0.0038128432352095842, + 0.005046255420893431, + 0.006057226564735174, + 0.007202260196208954, + 0.005370210390537977, + 0.004771299194544554, + 0.009650289081037045, + 0.008508709259331226, + 0.0021094365511089563, + 0.0028227628208696842, + 0.002329583978280425, + 0.002157878363505006, + 0.002353915711864829, + 0.0018297394271939993, + 0.00351673923432827, + 0.0034362308215349913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.1428235024213791 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004957877099514008 + ], + "opposite_logprob_tokens": [ + -7.625495910644531 + ], + "logprob_reward": -0.0004957877099514008, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0004957877099514008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1381036341190338, + 0.3288349211215973, + 0.04007900133728981, + 0.023763131350278854, + 0.2601017653942108, + 0.027010053396224976, + 0.026589233428239822, + 0.0679653137922287, + 0.04108807444572449, + 0.04646480455994606 + ], + "credit_map": [ + 0.01486031711101532, + 0.019615067169070244, + 0.010740742087364197, + 0.02398594841361046, + 0.023498117923736572, + 0.01455424539744854, + 0.013084623962640762, + 0.017764579504728317, + 0.021340614184737206, + 0.011975444853305817, + 0.10772540420293808, + 0.036889780312776566, + 0.04179385304450989, + 0.032474905252456665, + 0.010424291715025902, + 0.06621065735816956, + 0.0047618611715734005, + 0.004972487688064575, + 0.003779954742640257, + 0.004884186200797558, + 0.006473541259765625, + 0.006455328315496445, + 0.0037233128678053617, + 0.005028329789638519, + 0.00230066804215312, + 0.00278081395663321, + 0.0016818239819258451, + 0.003989721182733774, + 0.00316934147849679, + 0.004540398716926575, + 0.0024072665255516768, + 0.0028931000269949436, + 0.008529137820005417, + 0.006704721599817276, + 0.007216831669211388, + 0.009384162724018097, + 0.01719571091234684, + 0.10542955249547958, + 0.015148837119340897, + 0.0904928594827652, + 0.003579996759071946, + 0.003200778504833579, + 0.00432899035513401, + 0.004446632228791714, + 0.0022208804730325937, + 0.0032309042289853096, + 0.0022446566727012396, + 0.0037572174333035946, + 0.003058906178921461, + 0.0033968002535402775, + 0.00380663201212883, + 0.005052205640822649, + 0.002429857850074768, + 0.003146676579490304, + 0.004089152440428734, + 0.0016090048011392355, + 0.007471415214240551, + 0.005235370714217424, + 0.015384487807750702, + 0.015424362383782864, + 0.004743972327560186, + 0.003726877737790346, + 0.006064007990062237, + 0.009914825670421124, + 0.005421193782240152, + 0.00413753604516387, + 0.004730668850243092, + 0.005464478861540556, + 0.006180807016789913, + 0.003035112749785185, + 0.0062409741804003716, + 0.005877306219190359, + 0.0052447388879954815, + 0.005456042475998402, + 0.007079613860696554, + 0.00631696917116642, + 0.0067049674689769745, + 0.0030082371085882187, + 0.005972965620458126, + 0.006681273691356182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0004957877099514008 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up remains static, focused on a portion of a dark, polished mahogany tabletop, where a single, warm light source from an off-camera lamp casts a focused golden glow. Within this pool of light, two hands take center stage. One hand, belonging to an elderly individual, features prominent veins and neatly manicured nails, resting palm-down. Its index and middle fingers press gently yet firmly against the edge of a neatly stacked stack of off-white legal documents, held together by a simple metal clip. With a single, steady, deliberate movement, the hand gently pushes the entire sheet across the smooth, reflective wood, the paper sliding past with a soft, whispering sound. The other hand, younger, with smoother skin and slender fingers, rests motionless on the table, fingers slightly curled, in a receptive stillness as the paper slides toward it. The scene is imbued with a quiet, tense atmosphere, the warm, low-key lighting accentuating the subtle textures of skin, paper, and texture, transforming a simple transfer into a gesture of profound meaning and unspoken consent.", + "seed": 4377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_872_seed_4377_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": -0.06611022921279072, + "vlm_logprob": -0.06611022921279072, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.06611022921279072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04948912188410759 + ], + "opposite_logprob_tokens": [ + -3.7994890213012695 + ], + "logprob_reward": -0.04948912188410759, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04948912188410759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1346520185470581, + 0.040493905544281006, + 0.07442813366651535, + 0.018248502165079117, + 0.054361872375011444, + 0.1852356195449829, + 0.05688222870230675, + 0.04186416417360306, + 0.03336027264595032, + 0.3604733347892761 + ], + "credit_map": [ + 0.011307453736662865, + 0.008825354278087616, + 0.025048548355698586, + 0.024143872782588005, + 0.012931348755955696, + 0.014582316391170025, + 0.01840357854962349, + 0.019409526139497757, + 0.003624082775786519, + 0.004126506391912699, + 0.005238781683146954, + 0.007610417902469635, + 0.0036606888752430677, + 0.006832040846347809, + 0.005901720374822617, + 0.003499664133414626, + 0.007891803979873657, + 0.010994751006364822, + 0.011261211708188057, + 0.012166028842329979, + 0.00728302588686347, + 0.009249841794371605, + 0.008414196781814098, + 0.007167263887822628, + 0.0022384398616850376, + 0.002260434441268444, + 0.003393123857676983, + 0.0031243441626429558, + 0.0018573551205918193, + 0.0018696255283430219, + 0.0017887683352455497, + 0.0017164084129035473, + 0.003233705647289753, + 0.0028685047291219234, + 0.01848371885716915, + 0.018892748281359673, + 0.002017750171944499, + 0.0025633249897509813, + 0.0032903028186410666, + 0.003011812688782811, + 0.007118759676814079, + 0.004827126860618591, + 0.04023537039756775, + 0.07945328950881958, + 0.0037807903718203306, + 0.005380304530262947, + 0.03312039002776146, + 0.011319555342197418, + 0.005577193573117256, + 0.004030301235616207, + 0.005919271148741245, + 0.013417214155197144, + 0.004404014442116022, + 0.003612825646996498, + 0.01508969534188509, + 0.004831706173717976, + 0.004136208910495043, + 0.003228542860597372, + 0.004433300346136093, + 0.007535272743552923, + 0.003347335383296013, + 0.005040606949478388, + 0.008134941570460796, + 0.006007950287312269, + 0.0029890851583331823, + 0.0023882300592958927, + 0.003130370518192649, + 0.005956275854259729, + 0.0021826205775141716, + 0.00505487434566021, + 0.008377887308597565, + 0.003280924865975976, + 0.008800463750958443, + 0.006758331321179867, + 0.022187180817127228, + 0.09149318188428879, + 0.017750155180692673, + 0.03692994639277458, + 0.1604476273059845, + 0.016106393188238144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.04948912188410759 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018206385895609856 + ], + "opposite_logprob_tokens": [ + -4.018206596374512 + ], + "logprob_reward": -0.018206385895609856, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018206385895609856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07357140630483627, + 0.04425398260354996, + 0.07263314723968506, + 0.02250405214726925, + 0.037837959825992584, + 0.1787584275007248, + 0.02007727324962616, + 0.21427318453788757, + 0.22623515129089355, + 0.10985545814037323 + ], + "credit_map": [ + 0.007588757202029228, + 0.008259251713752747, + 0.010689739137887955, + 0.010730062611401081, + 0.008820357732474804, + 0.008487979881465435, + 0.011072748340666294, + 0.007922503165900707, + 0.003395762527361512, + 0.0039932033978402615, + 0.005645313300192356, + 0.00892609916627407, + 0.002311066258698702, + 0.004144673235714436, + 0.009810401126742363, + 0.006027462892234325, + 0.008215547539293766, + 0.008890584111213684, + 0.010507915169000626, + 0.011789589188992977, + 0.0055443462915718555, + 0.005844875704497099, + 0.012777713127434254, + 0.00906257051974535, + 0.0017308705719187856, + 0.0015890436479821801, + 0.003337536007165909, + 0.003354020416736603, + 0.0027107414789497852, + 0.00206854660063982, + 0.0032842622604221106, + 0.004429028835147619, + 0.0035762565676122904, + 0.0028517574537545443, + 0.0062582227401435375, + 0.0068645114079117775, + 0.00536623178049922, + 0.004535699728876352, + 0.004189420025795698, + 0.004195858258754015, + 0.008351615630090237, + 0.007031918037682772, + 0.038842346519231796, + 0.07324264943599701, + 0.005661358591169119, + 0.006338165141642094, + 0.03283287212252617, + 0.006457487121224403, + 0.0012338989181444049, + 0.002022769767791033, + 0.0015555830905213952, + 0.005196571350097656, + 0.0018105547642335296, + 0.001478330697864294, + 0.005554726347327232, + 0.0012248358689248562, + 0.011654388159513474, + 0.00848427601158619, + 0.01838306337594986, + 0.05282703787088394, + 0.012070223689079285, + 0.017786510288715363, + 0.07937125116586685, + 0.013696415349841118, + 0.010724020190536976, + 0.010480426251888275, + 0.021847635507583618, + 0.04241851344704628, + 0.016472667455673218, + 0.019253412261605263, + 0.08525355905294418, + 0.01978490874171257, + 0.0027353146579116583, + 0.0020433764439076185, + 0.005165980197489262, + 0.019195135682821274, + 0.004363186191767454, + 0.026715872809290886, + 0.04364993795752525, + 0.005986643955111504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.018206385895609856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795691817998886 + ], + "opposite_logprob_tokens": [ + -4.634795665740967 + ], + "logprob_reward": -0.009795691817998886, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795691817998886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16020382940769196, + 0.10470757633447647, + 0.057821933180093765, + 0.0732649713754654, + 0.09526018053293228, + 0.28963515162467957, + 0.016602760180830956, + 0.032742127776145935, + 0.09545134007930756, + 0.07431011646986008 + ], + "credit_map": [ + 0.013354255817830563, + 0.013557582162320614, + 0.03223586454987526, + 0.02689698338508606, + 0.014941815286874771, + 0.016599470749497414, + 0.02261153794825077, + 0.02000631019473076, + 0.008753783069550991, + 0.011590267531573772, + 0.014040550217032433, + 0.0142271239310503, + 0.01085073221474886, + 0.018904490396380424, + 0.017845394089818, + 0.008495235815644264, + 0.006462351884692907, + 0.0072844428941607475, + 0.008530132472515106, + 0.008373013697564602, + 0.004516896791756153, + 0.0048262691125273705, + 0.010089767165482044, + 0.007739060092717409, + 0.006175069138407707, + 0.007245632819831371, + 0.011541505344212055, + 0.012341560795903206, + 0.009993113577365875, + 0.007184053771197796, + 0.009449117816984653, + 0.009334924630820751, + 0.006143687758594751, + 0.005999422166496515, + 0.02768007665872574, + 0.02748255431652069, + 0.008211791515350342, + 0.00734022306278348, + 0.005823657847940922, + 0.006578768603503704, + 0.01025348249822855, + 0.00694983871653676, + 0.06519242376089096, + 0.127224400639534, + 0.005029060412198305, + 0.0071274880319833755, + 0.054444555193185806, + 0.013413894921541214, + 0.0019670717883855104, + 0.0023002037778496742, + 0.0021156815346330404, + 0.002420330187305808, + 0.0021820717956870794, + 0.0014367853291332722, + 0.0022873207926750183, + 0.0018932931125164032, + 0.003521138336509466, + 0.0030187778174877167, + 0.0032649447675794363, + 0.005734551697969437, + 0.002480138326063752, + 0.004931774456053972, + 0.005177256651222706, + 0.004613547120243311, + 0.010013298131525517, + 0.008264963515102863, + 0.01111134234815836, + 0.0122981583699584, + 0.007610441651195288, + 0.018308281898498535, + 0.016871515661478043, + 0.010973337106406689, + 0.003065295983105898, + 0.0029927927535027266, + 0.004260185174643993, + 0.021365590393543243, + 0.004931256640702486, + 0.005577641539275646, + 0.02680262178182602, + 0.00531473895534873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.009795691817998886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25213900208473206 + ], + "opposite_logprob_tokens": [ + -1.5021389722824097 + ], + "logprob_reward": -0.25213900208473206, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25213900208473206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07204223424196243, + 0.05315003544092178, + 0.14023028314113617, + 0.11243397742509842, + 0.12266622483730316, + 0.2548186779022217, + 0.036903899163007736, + 0.023133564740419388, + 0.018452808260917664, + 0.16616831719875336 + ], + "credit_map": [ + 0.005553703289479017, + 0.0054251342080533504, + 0.013668077066540718, + 0.01222799252718687, + 0.006582531146705151, + 0.007260218728333712, + 0.014272933825850487, + 0.007051641121506691, + 0.0046082064509391785, + 0.003808517474681139, + 0.007331748027354479, + 0.009156376123428345, + 0.0030281732324510813, + 0.006476437207311392, + 0.010833550244569778, + 0.007907026447355747, + 0.012989492155611515, + 0.017482981085777283, + 0.020200388506054878, + 0.015630539506673813, + 0.013880088925361633, + 0.017299793660640717, + 0.028357962146401405, + 0.014389041811227798, + 0.009444775991141796, + 0.00872822105884552, + 0.01691906526684761, + 0.01899980567395687, + 0.01293803658336401, + 0.009577220305800438, + 0.017347892746329308, + 0.018478963524103165, + 0.008070952259004116, + 0.007294426206499338, + 0.033672790974378586, + 0.03377113863825798, + 0.013092197477817535, + 0.007337556220591068, + 0.008858712390065193, + 0.010568451136350632, + 0.010392284020781517, + 0.007747413590550423, + 0.05416346341371536, + 0.10868581384420395, + 0.006205243524163961, + 0.007262165192514658, + 0.04542193189263344, + 0.014940355904400349, + 0.0030753109604120255, + 0.0034586142282932997, + 0.005686766467988491, + 0.006815179716795683, + 0.004313124809414148, + 0.003607063088566065, + 0.006411091890186071, + 0.0035367486998438835, + 0.0020676879212260246, + 0.0019820453599095345, + 0.0023946170695126057, + 0.0037318433169275522, + 0.0022884649224579334, + 0.0031896685250103474, + 0.0041153221391141415, + 0.003363915951922536, + 0.0016375803388655186, + 0.0019669889006763697, + 0.002415209077298641, + 0.002958582015708089, + 0.0016208213055506349, + 0.0030957036651670933, + 0.0038366205990314484, + 0.0009213027660734951, + 0.004346834030002356, + 0.003720322158187628, + 0.011135613545775414, + 0.04408331960439682, + 0.00904566515237093, + 0.011195004917681217, + 0.07279554754495621, + 0.009846006520092487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.25213900208473206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009209443815052509 + ], + "opposite_logprob_tokens": [ + -7.00092077255249 + ], + "logprob_reward": -0.0009209443815052509, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009209443815052509, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.142891064286232, + 0.06939234584569931, + 0.057482969015836716, + 0.06395111232995987, + 0.03639357537031174, + 0.0259857214987278, + 0.053699903190135956, + 0.06311006844043732, + 0.20787009596824646, + 0.27922311425209045 + ], + "credit_map": [ + 0.014092843048274517, + 0.009886760264635086, + 0.02777213416993618, + 0.023641027510166168, + 0.011722264811396599, + 0.014903964474797249, + 0.021805962547659874, + 0.019066113978624344, + 0.006875599268823862, + 0.008770689368247986, + 0.010239965282380581, + 0.011127782985568047, + 0.005234124604612589, + 0.009991721250116825, + 0.0109785795211792, + 0.006173884496092796, + 0.00682533485814929, + 0.00570058124139905, + 0.008651971817016602, + 0.007609813939779997, + 0.005496181081980467, + 0.004964373540133238, + 0.010060990229249, + 0.00817372091114521, + 0.00567669328302145, + 0.006473131012171507, + 0.009518658742308617, + 0.01050228625535965, + 0.005724489223212004, + 0.007459989748895168, + 0.008139727637171745, + 0.010456140153110027, + 0.0034279488027095795, + 0.004045007284730673, + 0.0074345264583826065, + 0.007857048884034157, + 0.0033135402482002974, + 0.003310824977234006, + 0.0032003538217395544, + 0.003804325358942151, + 0.002327685011550784, + 0.0032845381647348404, + 0.003503174288198352, + 0.0037457661237567663, + 0.0029359834734350443, + 0.0034179578069597483, + 0.0035055933985859156, + 0.0032650218345224857, + 0.004917254205793142, + 0.00295333843678236, + 0.0056613171473145485, + 0.013533033430576324, + 0.004012663848698139, + 0.003232126124203205, + 0.015321473591029644, + 0.004068699199706316, + 0.005781915504485369, + 0.005241192877292633, + 0.005703618284314871, + 0.014642099849879742, + 0.00510505773127079, + 0.006385176442563534, + 0.014476841315627098, + 0.005774162244051695, + 0.00581729132682085, + 0.005714743863791227, + 0.011394345201551914, + 0.044851724058389664, + 0.010309161618351936, + 0.031409069895744324, + 0.09040801227092743, + 0.007965755648911, + 0.0061973086558282375, + 0.004330994561314583, + 0.014218469150364399, + 0.06599009782075882, + 0.012319518253207207, + 0.04287225008010864, + 0.12469432502985, + 0.008600135333836079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0009209443815052509 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle close-up remains static, focused on a portion of a dark, polished mahogany tabletop, where a single, warm light source from an off-camera lamp casts a focused golden glow. Within this pool of light, two hands take center stage. One hand, belonging to an elderly individual, features prominent veins and neatly manicured nails, resting palm-down. Its index and middle fingers press gently yet firmly against the edge of a neatly stacked stack of off-white legal documents, held together by a simple metal clip. With a single, steady, deliberate movement, the hand gently pushes the entire sheet across the smooth, reflective wood, the paper sliding past with a soft, whispering sound. The other hand, younger, with smoother skin and slender fingers, rests motionless on the table, fingers slightly curled, in a receptive stillness as the paper slides toward it. The scene is imbued with a quiet, tense atmosphere, the warm, low-key lighting accentuating the subtle textures of skin, paper, and texture, transforming a simple transfer into a gesture of profound meaning and unspoken consent.", + "seed": 4378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_11/index_872_seed_4378_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": -0.04056179446633905, + "vlm_logprob": -0.04056179446633905, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.04056179446633905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04792701080441475 + ], + "opposite_logprob_tokens": [ + -4.047926902770996 + ], + "logprob_reward": -0.04792701080441475, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.04792701080441475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3273774981498718, + 0.057330790907144547, + 0.026148241013288498, + 0.03539713844656944, + 0.060614459216594696, + 0.05081808194518089, + 0.03865153715014458, + 0.16729383170604706, + 0.17643187940120697, + 0.059936486184597015 + ], + "credit_map": [ + 0.03332410007715225, + 0.03218221664428711, + 0.056116510182619095, + 0.0645337626338005, + 0.018336975947022438, + 0.04539989307522774, + 0.02805669791996479, + 0.04942736774682999, + 0.004216108936816454, + 0.008502951823174953, + 0.008967808447778225, + 0.013404680415987968, + 0.005301427561789751, + 0.005134626757353544, + 0.0054459827952086926, + 0.006357208359986544, + 0.003303154371678829, + 0.003276761854067445, + 0.0041133142076432705, + 0.004666334018111229, + 0.002590413438156247, + 0.0023136944510042667, + 0.0030910689383745193, + 0.0027935022953897715, + 0.00457785464823246, + 0.004811963066458702, + 0.005504343658685684, + 0.0066427732817828655, + 0.0027697577606886625, + 0.003667872864753008, + 0.0031031121034175158, + 0.004319461528211832, + 0.007842033170163631, + 0.010531412437558174, + 0.008116454817354679, + 0.010339286178350449, + 0.005209002643823624, + 0.0054383790120482445, + 0.007170889992266893, + 0.005967005621641874, + 0.006690237205475569, + 0.007199770305305719, + 0.006910280790179968, + 0.00702083483338356, + 0.0037083118222653866, + 0.007258972153067589, + 0.005091764964163303, + 0.006937914527952671, + 0.004572589416056871, + 0.007748911622911692, + 0.005497088190168142, + 0.005985296797007322, + 0.0034608207643032074, + 0.0050295619294047356, + 0.00249313423410058, + 0.0038641360588371754, + 0.02507038228213787, + 0.025401566177606583, + 0.024840299040079117, + 0.023838136345148087, + 0.011705093085765839, + 0.01975872926414013, + 0.015188471414148808, + 0.021491164341568947, + 0.02146804891526699, + 0.024476617574691772, + 0.026766596361994743, + 0.0297292061150074, + 0.016799671575427055, + 0.01826261729001999, + 0.017137568444013596, + 0.021791571751236916, + 0.01198748778551817, + 0.00854798499494791, + 0.008038199506700039, + 0.01006210595369339, + 0.0038107531145215034, + 0.007281136699020863, + 0.0044302623718976974, + 0.005778559949249029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.04792701080441475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006757147144526243 + ], + "opposite_logprob_tokens": [ + -5.0067572593688965 + ], + "logprob_reward": -0.006757147144526243, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006757147144526243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22400280833244324, + 0.1389325112104416, + 0.12540583312511444, + 0.09213543683290482, + 0.04293200746178627, + 0.044092971831560135, + 0.04074332118034363, + 0.1605759561061859, + 0.0630388930439949, + 0.06814029067754745 + ], + "credit_map": [ + 0.029859019443392754, + 0.029224051162600517, + 0.03396853804588318, + 0.04041355475783348, + 0.015223272144794464, + 0.02284603752195835, + 0.01911318302154541, + 0.03335513547062874, + 0.01787531189620495, + 0.021673135459423065, + 0.021593721583485603, + 0.023122737184166908, + 0.01206617709249258, + 0.013331027701497078, + 0.011925109662115574, + 0.01734527386724949, + 0.015666894614696503, + 0.01603422313928604, + 0.019682297483086586, + 0.023213379085063934, + 0.014546670019626617, + 0.011483239941298962, + 0.011108042672276497, + 0.013671070337295532, + 0.009432482533156872, + 0.013622705824673176, + 0.01576637662947178, + 0.01774740032851696, + 0.008843774907290936, + 0.010181068442761898, + 0.0075805592350661755, + 0.008961062878370285, + 0.004869522526860237, + 0.004611831624060869, + 0.006253066007047892, + 0.007920736446976662, + 0.0037418773863464594, + 0.005229902919381857, + 0.004534835461527109, + 0.005770231131464243, + 0.006273113191127777, + 0.009045488201081753, + 0.00504302280023694, + 0.007314356043934822, + 0.0043112412095069885, + 0.00403554504737258, + 0.003359749913215637, + 0.004710452631115913, + 0.004066476132720709, + 0.006957024335861206, + 0.0056652287021279335, + 0.007279523182660341, + 0.0036228757817298174, + 0.004733780864626169, + 0.0034263532143086195, + 0.0049920533783733845, + 0.02288360707461834, + 0.023578403517603874, + 0.024358535185456276, + 0.024072227999567986, + 0.010930284857749939, + 0.01890578493475914, + 0.014378021471202374, + 0.02146909199655056, + 0.0063545480370521545, + 0.011017388664186, + 0.010236700065433979, + 0.013114829547703266, + 0.0044266823679208755, + 0.005738541949540377, + 0.006316075101494789, + 0.0058341254480183125, + 0.009240270592272282, + 0.010160024277865887, + 0.008437348529696465, + 0.00977171678096056, + 0.006927225273102522, + 0.008619057014584541, + 0.006489665247499943, + 0.008494979701936245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.006757147144526243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020606480538845062 + ], + "opposite_logprob_tokens": [ + -3.895606517791748 + ], + "logprob_reward": -0.020606480538845062, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020606480538845062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30139702558517456, + 0.05237710475921631, + 0.04028628021478653, + 0.04699582979083061, + 0.054835397750139236, + 0.15977627038955688, + 0.059684790670871735, + 0.19850699603557587, + 0.05115512013435364, + 0.03498518094420433 + ], + "credit_map": [ + 0.02569335140287876, + 0.0325370728969574, + 0.04706091806292534, + 0.052910350263118744, + 0.020379211753606796, + 0.04990716278553009, + 0.0260623786598444, + 0.04684658721089363, + 0.005214556120336056, + 0.007549055851995945, + 0.008227565325796604, + 0.010502662509679794, + 0.00619756244122982, + 0.004086565226316452, + 0.004693802446126938, + 0.005905334372073412, + 0.00473603792488575, + 0.006974321790039539, + 0.007124691735953093, + 0.00780402310192585, + 0.004677418619394302, + 0.003111850470304489, + 0.0032002723310142756, + 0.002657666802406311, + 0.00536193186417222, + 0.008839930407702923, + 0.006890021730214357, + 0.01025466714054346, + 0.0043266331776976585, + 0.004022079054266214, + 0.0033415756188333035, + 0.003958991263061762, + 0.005147377494722605, + 0.006934585515409708, + 0.009168710559606552, + 0.009321518242359161, + 0.004870064556598663, + 0.007332250941544771, + 0.005173637066036463, + 0.006887257564812899, + 0.01929781213402748, + 0.034897319972515106, + 0.020335393026471138, + 0.026637254282832146, + 0.014563468284904957, + 0.01613706722855568, + 0.010680973529815674, + 0.017226992174983025, + 0.0069332304410636425, + 0.014020738191902637, + 0.007983684539794922, + 0.010905943810939789, + 0.005019627511501312, + 0.005071984138339758, + 0.003904094221070409, + 0.005845491774380207, + 0.028410281985998154, + 0.029738547280430794, + 0.030653759837150574, + 0.030802467837929726, + 0.013679811730980873, + 0.022008473053574562, + 0.017468929290771484, + 0.025744732469320297, + 0.005057700909674168, + 0.007817262783646584, + 0.008619220927357674, + 0.010926502756774426, + 0.004131901543587446, + 0.004815710708498955, + 0.004640982486307621, + 0.005145840346813202, + 0.0055871387012302876, + 0.005675340536981821, + 0.0039480458945035934, + 0.004602806176990271, + 0.002487055491656065, + 0.005241645034402609, + 0.0021582769695669413, + 0.005284874700009823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.020606480538845062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707902491092682 + ], + "opposite_logprob_tokens": [ + -2.1270790100097656 + ], + "logprob_reward": -0.12707902491092682, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707902491092682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33182233572006226, + 0.07490980625152588, + 0.04036853834986687, + 0.05176336690783501, + 0.06848067790269852, + 0.12540759146213531, + 0.060189273208379745, + 0.05016620084643364, + 0.17419297993183136, + 0.022699208930134773 + ], + "credit_map": [ + 0.033980730921030045, + 0.03268137574195862, + 0.04775611311197281, + 0.03833458945155144, + 0.025285329669713974, + 0.05654352903366089, + 0.03224163502454758, + 0.0649990439414978, + 0.007107425015419722, + 0.009351760149002075, + 0.011249154806137085, + 0.01600990816950798, + 0.007238407153636217, + 0.0076227812096476555, + 0.006165986880660057, + 0.010164380073547363, + 0.004465088248252869, + 0.005799107253551483, + 0.008186426013708115, + 0.008442466147243977, + 0.003808328416198492, + 0.0035760707687586546, + 0.0033031227067112923, + 0.002787928096950054, + 0.006541580427438021, + 0.007840474136173725, + 0.009177085943520069, + 0.009880117140710354, + 0.003375658066943288, + 0.004413262475281954, + 0.005512088071554899, + 0.005023101344704628, + 0.0071234325878322124, + 0.010178238153457642, + 0.010694519616663456, + 0.012231306172907352, + 0.005515153519809246, + 0.007970942184329033, + 0.007152671925723553, + 0.007614416535943747, + 0.013816559687256813, + 0.025870144367218018, + 0.01751876063644886, + 0.020150108262896538, + 0.011087180115282536, + 0.01545607391744852, + 0.008430260233581066, + 0.013078494928777218, + 0.006742429919540882, + 0.013698119670152664, + 0.008438815362751484, + 0.010240464471280575, + 0.005016271490603685, + 0.005874262657016516, + 0.004136735573410988, + 0.006042174529284239, + 0.007778425235301256, + 0.006759296637028456, + 0.007473199162632227, + 0.008667022921144962, + 0.0030797093641012907, + 0.0052893939428031445, + 0.005325922742486, + 0.005793231073766947, + 0.02241280861198902, + 0.025970229879021645, + 0.027210453525185585, + 0.027635637670755386, + 0.016179395839571953, + 0.01839793473482132, + 0.014547860249876976, + 0.021838661283254623, + 0.0034672620240598917, + 0.004429018124938011, + 0.0030411100015044212, + 0.003376095090061426, + 0.0015663348603993654, + 0.0026219317223876715, + 0.0012901778100058436, + 0.002907279646024108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.12707902491092682 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 high-angle close-up that remains static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene focused on a dark, polished mahogany tabletop with a single warm light source casting a golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two hands visible: one elderly with prominent veins and manicured nails, and one younger with smoother skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the elderly hand push a stack of off-white legal documents held by a metal clip across the table while the younger hand remains motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting low-key and warm, accentuating the textures of the skin, paper, and wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00043930893298238516 + ], + "opposite_logprob_tokens": [ + -7.750439167022705 + ], + "logprob_reward": -0.00043930893298238516, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.00043930893298238516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2776857316493988, + 0.09572459012269974, + 0.058665014803409576, + 0.09517979621887207, + 0.03699926286935806, + 0.034843649715185165, + 0.05317476764321327, + 0.17709259688854218, + 0.11182546615600586, + 0.0588090606033802 + ], + "credit_map": [ + 0.035819731652736664, + 0.03445155546069145, + 0.044149309396743774, + 0.048516206443309784, + 0.014640793204307556, + 0.03651029244065285, + 0.022150078788399696, + 0.041447773575782776, + 0.010499411262571812, + 0.013462847098708153, + 0.01545126736164093, + 0.019233014434576035, + 0.0076740519143640995, + 0.010613609105348587, + 0.006917967461049557, + 0.011872416362166405, + 0.007598686031997204, + 0.007473296020179987, + 0.009387626312673092, + 0.012156595475971699, + 0.006438475102186203, + 0.00498906709253788, + 0.004811843391507864, + 0.005809429567307234, + 0.00848136842250824, + 0.013216733001172543, + 0.016217339783906937, + 0.018436571583151817, + 0.009276757016777992, + 0.010096737183630466, + 0.009247109293937683, + 0.010207181796431541, + 0.004124890081584454, + 0.004436927381902933, + 0.005344565492123365, + 0.00636096578091383, + 0.0025869300588965416, + 0.005098012741655111, + 0.004565167706459761, + 0.00448180316016078, + 0.004185141995549202, + 0.003980822395533323, + 0.005575934890657663, + 0.0054787201806902885, + 0.00310767674818635, + 0.005065582692623138, + 0.003894193796440959, + 0.00355557631701231, + 0.005569747183471918, + 0.011618055403232574, + 0.006756958086043596, + 0.00876277219504118, + 0.004583511035889387, + 0.006147707812488079, + 0.004106973763555288, + 0.0056290412321686745, + 0.027269242331385612, + 0.027519192546606064, + 0.0270122978836298, + 0.02642061375081539, + 0.011577533558011055, + 0.018978094682097435, + 0.015050982125103474, + 0.02326464280486107, + 0.012763689272105694, + 0.016245540231466293, + 0.014280888251960278, + 0.015169252641499043, + 0.007662741467356682, + 0.014907407574355602, + 0.011050187982618809, + 0.019745759665966034, + 0.009517254307866096, + 0.006448653060942888, + 0.00905994139611721, + 0.009167157113552094, + 0.00520471204072237, + 0.0076256729662418365, + 0.004055644851177931, + 0.007730026729404926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.00043930893298238516 + } + ], + "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/0000055/group_12/index_tensor([21, 21])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_tensor([21, 21])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1bffbf192cc7b02096aeaafbcc1012db229fb8a1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_tensor([21, 21])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a level angle, capturing the dramatic moment at a remarkably high frame rate, focuses on a crow perched on a bare, weathered, cracked branch. In the bright, cool, diffuse light, its glossy black body stands in stark contrast to the soft, out-of-focus, grayish-white gloom of the sky. Suddenly, as if startled by an invisible thunderclap, a flicker of alarm flashes in its small, obsidian eyes, and its body tenses. Then, with a sudden burst of power, its wings spread upward and backward at astonishing speed. In extreme slow motion, each individual black flight feather unfurls in the air like a fan, its delicate blue-violet iridescence shimmering in the light, revealing the intricate structure of its vanes and rachises. With the powerful flapping of wings, one can almost feel the shock wave, its body is pushed upward by the huge force, and its claws instantly release the branch. The camera then tilts upward smoothly to capture its back as it flies resolutely into the sky. The whole picture, in the cold black, white and gray tones, creates a dramatic atmosphere full of tension, primitive and poetic.", + "seed": 14038, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_21_seed_14038_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": -0.01215047282166779, + "vlm_logprob": -0.01215047282166779, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.01215047282166779, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02018914557993412 + ], + "opposite_logprob_tokens": [ + -4.39518928527832 + ], + "logprob_reward": -0.02018914557993412, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.02018914557993412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07604469358921051, + 0.05290890112519264, + 0.13686734437942505, + 0.03783419355750084, + 0.0791698694229126, + 0.03130804002285004, + 0.06357137858867645, + 0.03051002323627472, + 0.046707574278116226, + 0.44507795572280884 + ], + "credit_map": [ + 0.010021612979471684, + 0.014764695428311825, + 0.006168936379253864, + 0.004340172745287418, + 0.007412211503833532, + 0.008479221723973751, + 0.006692221388220787, + 0.018165621906518936, + 0.007066550198942423, + 0.01728886179625988, + 0.0051347739063203335, + 0.0023480479139834642, + 0.0037122112698853016, + 0.0050087571144104, + 0.003511963412165642, + 0.008837730623781681, + 0.027225201949477196, + 0.010601530782878399, + 0.007576455362141132, + 0.00683988444507122, + 0.010879095643758774, + 0.03465884178876877, + 0.011355385184288025, + 0.027730954810976982, + 0.0034044934436678886, + 0.0035662134177982807, + 0.003165989648550749, + 0.00259514432400465, + 0.005198532249778509, + 0.004942722152918577, + 0.004407263360917568, + 0.010553835891187191, + 0.00863448902964592, + 0.009197253733873367, + 0.0074005271308124065, + 0.007671610452234745, + 0.01471493300050497, + 0.011214207857847214, + 0.007752457167953253, + 0.0125843845307827, + 0.004048667382448912, + 0.004155431874096394, + 0.004514687694609165, + 0.004211431369185448, + 0.003090228885412216, + 0.003567332634702325, + 0.0027558349538594484, + 0.0049644229002296925, + 0.011958596296608448, + 0.007704750634729862, + 0.006780281662940979, + 0.01059348601847887, + 0.007534914184361696, + 0.007982354611158371, + 0.005639163311570883, + 0.005377836059778929, + 0.0040625715628266335, + 0.00458653224632144, + 0.002790350466966629, + 0.004689588211476803, + 0.004628736525774002, + 0.0031789452768862247, + 0.002985578728839755, + 0.003587719053030014, + 0.007998190820217133, + 0.005847612861543894, + 0.007338790688663721, + 0.005781017243862152, + 0.004014922771602869, + 0.003741147229447961, + 0.0038740120362490416, + 0.008111882023513317, + 0.07562314718961716, + 0.060068197548389435, + 0.06901298463344574, + 0.04042978584766388, + 0.038208361715078354, + 0.05412403494119644, + 0.03552690148353577, + 0.07208456844091415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.02018914557993412 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032093231566250324 + ], + "opposite_logprob_tokens": [ + -5.753209114074707 + ], + "logprob_reward": -0.0032093231566250324, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032093231566250324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.186256542801857, + 0.21036124229431152, + 0.030250171199440956, + 0.041293222457170486, + 0.1073666363954544, + 0.029082419350743294, + 0.09404654055833817, + 0.05956518277525902, + 0.07984920591115952, + 0.16192883253097534 + ], + "credit_map": [ + 0.02335689403116703, + 0.060805510729551315, + 0.023832328617572784, + 0.009221935644745827, + 0.014752035029232502, + 0.01647055707871914, + 0.014203168451786041, + 0.023614121600985527, + 0.02724446915090084, + 0.08171714097261429, + 0.025427857413887978, + 0.010805711150169373, + 0.013242127373814583, + 0.01796267367899418, + 0.012640339322388172, + 0.02132090926170349, + 0.005523813422769308, + 0.0031948979012668133, + 0.0012835368979722261, + 0.0012186478124931455, + 0.0024146733339875937, + 0.0051797120831906796, + 0.0021439308766275644, + 0.009290959686040878, + 0.005020110867917538, + 0.004997496493160725, + 0.004210530314594507, + 0.003437843406572938, + 0.004876132123172283, + 0.00513130659237504, + 0.005621196702122688, + 0.007998606190085411, + 0.009120848961174488, + 0.012941943481564522, + 0.015049980953335762, + 0.01008609589189291, + 0.025611933320760727, + 0.015274928882718086, + 0.009968530386686325, + 0.009312367998063564, + 0.003465029178187251, + 0.003628079080954194, + 0.0037152667064219713, + 0.0044617364183068275, + 0.004372240975499153, + 0.0030917241238057613, + 0.0033908486366271973, + 0.0029574919026345015, + 0.01271080318838358, + 0.01255849190056324, + 0.013921483419835567, + 0.013915895484387875, + 0.012159214355051517, + 0.010432049632072449, + 0.00925450585782528, + 0.009094098582863808, + 0.007899409160017967, + 0.007377146743237972, + 0.006148945074528456, + 0.010169019922614098, + 0.007094390690326691, + 0.006601880304515362, + 0.005480966996401548, + 0.008793425746262074, + 0.016875561326742172, + 0.00963409710675478, + 0.01364594604820013, + 0.008098771795630455, + 0.008947223424911499, + 0.007412448059767485, + 0.005519462749361992, + 0.009715699590742588, + 0.03300904482603073, + 0.02397255413234234, + 0.02508706785738468, + 0.011385458521544933, + 0.013317792676389217, + 0.014903842471539974, + 0.00927864108234644, + 0.030974410474300385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0032093231566250324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019483646610751748 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.0019483646610751748, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019483646610751748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1375667154788971, + 0.049037691205739975, + 0.06073009967803955, + 0.04684725031256676, + 0.10456394404172897, + 0.09856423735618591, + 0.08045002073049545, + 0.08460775762796402, + 0.13390201330184937, + 0.20373030006885529 + ], + "credit_map": [ + 0.01473467331379652, + 0.0565168671309948, + 0.020293377339839935, + 0.00935919675976038, + 0.007665303535759449, + 0.011450081132352352, + 0.005478966981172562, + 0.012068240903317928, + 0.007546591106802225, + 0.007225268520414829, + 0.004685648251324892, + 0.0028494521975517273, + 0.0035686884075403214, + 0.007277993485331535, + 0.00500772288069129, + 0.010876324027776718, + 0.014015484601259232, + 0.004652923904359341, + 0.0021623012144118547, + 0.0021389247849583626, + 0.004245946649461985, + 0.016192680224776268, + 0.0035390653647482395, + 0.013782767578959465, + 0.005468951538205147, + 0.004293350502848625, + 0.003406787756830454, + 0.0032595794182270765, + 0.009917397052049637, + 0.005991326179355383, + 0.006808845326304436, + 0.007701006717979908, + 0.013871870003640652, + 0.010976557619869709, + 0.013565401546657085, + 0.01214013434946537, + 0.014583873562514782, + 0.013943623751401901, + 0.010643663816154003, + 0.014838816598057747, + 0.013077935203909874, + 0.01240062341094017, + 0.010261467657983303, + 0.00998312421143055, + 0.01342066377401352, + 0.014100814238190651, + 0.010126035660505295, + 0.015193569473922253, + 0.012056606821715832, + 0.010059423744678497, + 0.008775602094829082, + 0.010605736635625362, + 0.011398347094655037, + 0.010957556776702404, + 0.008472233079373837, + 0.008124507032334805, + 0.012913303449749947, + 0.010067254304885864, + 0.008350932039320469, + 0.013940881937742233, + 0.00996338203549385, + 0.008070539683103561, + 0.008494279347360134, + 0.012807182036340237, + 0.020701371133327484, + 0.013970272615551949, + 0.021388085559010506, + 0.011288240551948547, + 0.013017518445849419, + 0.01532301027327776, + 0.013000579550862312, + 0.025212939828634262, + 0.03442413732409477, + 0.028336480259895325, + 0.03054937720298767, + 0.01989937759935856, + 0.019218789413571358, + 0.02452658861875534, + 0.018137037754058838, + 0.028638510033488274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0019483646610751748 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001715499092824757 + ], + "opposite_logprob_tokens": [ + -6.376715660095215 + ], + "logprob_reward": -0.001715499092824757, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001715499092824757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06909909844398499, + 0.051143161952495575, + 0.014823583886027336, + 0.02810046821832657, + 0.30873140692710876, + 0.0344407819211483, + 0.04511827230453491, + 0.03775465115904808, + 0.08026073127985, + 0.3305278718471527 + ], + "credit_map": [ + 0.0079610301181674, + 0.024953577667474747, + 0.008399276062846184, + 0.004079947713762522, + 0.003963158465921879, + 0.006416692398488522, + 0.003284533740952611, + 0.010040879249572754, + 0.007156002800911665, + 0.01661568507552147, + 0.006239129696041346, + 0.0026503982953727245, + 0.002893600380048156, + 0.004178476054221392, + 0.002617795020341873, + 0.00879207719117403, + 0.0027389214374125004, + 0.00124569539912045, + 0.0013999161310493946, + 0.0010173224145546556, + 0.0016036767046898603, + 0.0026643124874681234, + 0.0011297372402623296, + 0.003024002304300666, + 0.003742866450920701, + 0.00310262362472713, + 0.0025779162533581257, + 0.0015816993545740843, + 0.0031999368220567703, + 0.0037792969960719347, + 0.0034369677305221558, + 0.006679159123450518, + 0.025958774611353874, + 0.04264456778764725, + 0.03393325209617615, + 0.03793613985180855, + 0.06631559133529663, + 0.04013429582118988, + 0.0297401063144207, + 0.032068684697151184, + 0.004916589707136154, + 0.006451297551393509, + 0.0038636690005660057, + 0.003318897681310773, + 0.0038419882766902447, + 0.004220492672175169, + 0.0028979023918509483, + 0.004929943475872278, + 0.006951092276722193, + 0.005512756295502186, + 0.004379236605018377, + 0.006749188993126154, + 0.006852404214441776, + 0.007140515837818384, + 0.003461232176050544, + 0.004071846138685942, + 0.005826534703373909, + 0.004780696239322424, + 0.0040453230030834675, + 0.0051746792159974575, + 0.003983745351433754, + 0.005599294789135456, + 0.003884203266352415, + 0.004460177384316921, + 0.007543794345110655, + 0.009079212322831154, + 0.012648588977754116, + 0.009161715395748615, + 0.008512008003890514, + 0.010263740085065365, + 0.0069737909361720085, + 0.016077885404229164, + 0.07075491547584534, + 0.032387878745794296, + 0.05723104253411293, + 0.030335545539855957, + 0.017563823610544205, + 0.029235221445560455, + 0.023481953889131546, + 0.06953748315572739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.001715499092824757 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369003161787987 + ], + "opposite_logprob_tokens": [ + -3.4086899757385254 + ], + "logprob_reward": -0.03369003161787987, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369003161787987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036922626197338104, + 0.047948647290468216, + 0.025399431586265564, + 0.04343281686306, + 0.36436715722084045, + 0.08354653418064117, + 0.0265409704297781, + 0.017500123009085655, + 0.031013576313853264, + 0.3233281672000885 + ], + "credit_map": [ + 0.005271332338452339, + 0.004586808383464813, + 0.0037941038608551025, + 0.0024215118028223515, + 0.00394856184720993, + 0.004539755638688803, + 0.00290264212526381, + 0.0094579067081213, + 0.006816216744482517, + 0.013419200666248798, + 0.005308231338858604, + 0.0026415216270834208, + 0.003049619495868683, + 0.005456689745187759, + 0.002970347413793206, + 0.008286818861961365, + 0.004993930924683809, + 0.0019687791354954243, + 0.0012655567843466997, + 0.0011663249460980296, + 0.002376165706664324, + 0.006505591794848442, + 0.001860962132923305, + 0.005262117832899094, + 0.00582784553989768, + 0.004512428771704435, + 0.003173269098624587, + 0.0029338530730456114, + 0.004255285486578941, + 0.005958127323538065, + 0.005590950604528189, + 0.011181051842868328, + 0.04316394403576851, + 0.045314718037843704, + 0.03273310512304306, + 0.03223434090614319, + 0.07072771340608597, + 0.05178997293114662, + 0.03773923218250275, + 0.05066410079598427, + 0.011290417052805424, + 0.009450545534491539, + 0.011696971952915192, + 0.012781921774148941, + 0.012715260498225689, + 0.009828870184719563, + 0.007153750397264957, + 0.00862879492342472, + 0.005295459646731615, + 0.0031961952336132526, + 0.002261866582557559, + 0.0040388889610767365, + 0.003008997766301036, + 0.003724585520103574, + 0.002753037493675947, + 0.0022619368974119425, + 0.0034437039867043495, + 0.002322134794667363, + 0.0016955233877524734, + 0.002058582380414009, + 0.0017968134488910437, + 0.001656919252127409, + 0.001730885123834014, + 0.002795559586957097, + 0.005227736197412014, + 0.0036590301897376776, + 0.005338496528565884, + 0.0027186519000679255, + 0.0023655048571527004, + 0.0030740266665816307, + 0.0021670146379619837, + 0.00646311417222023, + 0.06989725679159164, + 0.03361419960856438, + 0.05550407990813255, + 0.031213432550430298, + 0.016275795176625252, + 0.023655783385038376, + 0.02198411151766777, + 0.07118348032236099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.03369003161787987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, capturing the dramatic moment at a remarkably high frame rate, focuses on a crow perched on a bare, weathered, cracked branch. In the bright, cool, diffuse light, its glossy black body stands in stark contrast to the soft, out-of-focus, grayish-white gloom of the sky. Suddenly, as if startled by an invisible thunderclap, a flicker of alarm flashes in its small, obsidian eyes, and its body tenses. Then, with a sudden burst of power, its wings spread upward and backward at astonishing speed. In extreme slow motion, each individual black flight feather unfurls in the air like a fan, its delicate blue-violet iridescence shimmering in the light, revealing the intricate structure of its vanes and rachises. With the powerful flapping of wings, one can almost feel the shock wave, its body is pushed upward by the huge force, and its claws instantly release the branch. The camera then tilts upward smoothly to capture its back as it flies resolutely into the sky. The whole picture, in the cold black, white and gray tones, creates a dramatic atmosphere full of tension, primitive and poetic.", + "seed": 14039, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_21_seed_14039_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": -0.0672950322041288, + "vlm_logprob": -0.0672950322041288, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.0672950322041288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22979898750782013 + ], + "opposite_logprob_tokens": [ + -1.6047990322113037 + ], + "logprob_reward": -0.22979898750782013, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22979898750782013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08735848218202591, + 0.1279701292514801, + 0.02849765494465828, + 0.07770660519599915, + 0.5411232113838196, + 0.10610147565603256, + 0.0067139435559511185, + 0.006673451513051987, + 0.006253069266676903, + 0.011601986363530159 + ], + "credit_map": [ + 0.011812996119260788, + 0.005283089354634285, + 0.0055518988519907, + 0.003257657866925001, + 0.003739693434908986, + 0.021873271092772484, + 0.003548126434907317, + 0.032291751354932785, + 0.02451455034315586, + 0.012109076604247093, + 0.012154213152825832, + 0.00990022998303175, + 0.007315087132155895, + 0.018096253275871277, + 0.00603890186175704, + 0.03784181922674179, + 0.005010724067687988, + 0.0028514894656836987, + 0.0028713333886116743, + 0.0015939169097691774, + 0.0022030449472367764, + 0.0037903464399278164, + 0.0016923205694183707, + 0.008484480902552605, + 0.0072916303761303425, + 0.010327964089810848, + 0.007450405042618513, + 0.007201268337666988, + 0.008018274791538715, + 0.014058594591915607, + 0.006902574561536312, + 0.016455894336104393, + 0.015336012467741966, + 0.0171505156904459, + 0.008368049748241901, + 0.010668990202248096, + 0.015606692060828209, + 0.23133158683776855, + 0.010997586883604527, + 0.23166373372077942, + 0.011041326448321342, + 0.00722269294783473, + 0.017095055431127548, + 0.006266945507377386, + 0.01568649522960186, + 0.018032880499958992, + 0.014518627896904945, + 0.016237448900938034, + 0.0009387015015818179, + 0.0006544122006744146, + 0.00040319637628272176, + 0.00038179210969246924, + 0.0006659067003056407, + 0.0009518551523797214, + 0.0006617473554797471, + 0.002056331839412451, + 0.0008724948274902999, + 0.0007982290699146688, + 0.0008947051828727126, + 0.0005095885135233402, + 0.0006008794880472124, + 0.0007122252718545496, + 0.001134185236878693, + 0.0011511436896398664, + 0.0007021302008070052, + 0.0007689832127653062, + 0.0008200328447856009, + 0.0009073290857486427, + 0.0007900719065219164, + 0.0006300754030235112, + 0.0010423909407109022, + 0.0005920553230680525, + 0.0010340975131839514, + 0.0020887122955173254, + 0.0031944175716489553, + 0.0016150962328538299, + 0.0010358900763094425, + 0.0010376159334555268, + 0.0008685527136549354, + 0.0007276036776602268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.22979898750782013 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028397017158567905 + ], + "opposite_logprob_tokens": [ + -5.8778395652771 + ], + "logprob_reward": -0.0028397017158567905, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028397017158567905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07346139848232269, + 0.11289175599813461, + 0.05135510861873627, + 0.15140996873378754, + 0.2848493456840515, + 0.11267846077680588, + 0.08428146690130234, + 0.024397598579525948, + 0.04674654081463814, + 0.05792834609746933 + ], + "credit_map": [ + 0.010531335137784481, + 0.01540778111666441, + 0.00985864270478487, + 0.007455121725797653, + 0.008875710889697075, + 0.006446497980505228, + 0.005472433753311634, + 0.00941387563943863, + 0.02151699922978878, + 0.009441998787224293, + 0.012281984090805054, + 0.010967548936605453, + 0.00699236337095499, + 0.019396336749196053, + 0.006637682672590017, + 0.025656843557953835, + 0.008549332618713379, + 0.005741309840232134, + 0.004029096104204655, + 0.0031823655590415, + 0.004238322842866182, + 0.005805168766528368, + 0.003064175834879279, + 0.01674533635377884, + 0.01690256968140602, + 0.018148139119148254, + 0.011122378520667553, + 0.010396059602499008, + 0.011347223073244095, + 0.032075267285108566, + 0.009915744885802269, + 0.041502587497234344, + 0.009044576436281204, + 0.010887534357607365, + 0.005547397769987583, + 0.00810258835554123, + 0.009968435391783714, + 0.11665112525224686, + 0.00741442060098052, + 0.1172332614660263, + 0.012513051740825176, + 0.007244743872433901, + 0.01616060920059681, + 0.00543793523684144, + 0.016404978930950165, + 0.02067701332271099, + 0.014344975352287292, + 0.019895147532224655, + 0.0162749495357275, + 0.008598282933235168, + 0.0043640537187457085, + 0.005313910078257322, + 0.00688673323020339, + 0.011471971869468689, + 0.0079301493242383, + 0.02344142086803913, + 0.0031882005278021097, + 0.0038243713788688183, + 0.003675578162074089, + 0.001314256340265274, + 0.0023629069328308105, + 0.002913833362981677, + 0.0034692026674747467, + 0.003649250604212284, + 0.007262378465384245, + 0.008896473795175552, + 0.007205234374850988, + 0.005254687741398811, + 0.004116980824619532, + 0.004166337661445141, + 0.005598912946879864, + 0.004245533607900143, + 0.004777475725859404, + 0.009780913591384888, + 0.013136807829141617, + 0.010032434947788715, + 0.006348032038658857, + 0.004860684275627136, + 0.005385346710681915, + 0.0036066502798348665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.0028397017158567905 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001342587056569755 + ], + "opposite_logprob_tokens": [ + -6.6263427734375 + ], + "logprob_reward": -0.001342587056569755, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001342587056569755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16471107304096222, + 0.17432473599910736, + 0.038657043129205704, + 0.08610016107559204, + 0.17117398977279663, + 0.10233628004789352, + 0.14370709657669067, + 0.039957743138074875, + 0.025206636637449265, + 0.05382522940635681 + ], + "credit_map": [ + 0.02435552328824997, + 0.01119248103350401, + 0.011918723583221436, + 0.005727214738726616, + 0.006492188666015863, + 0.036286111921072006, + 0.006397685036063194, + 0.062341149896383286, + 0.033270176500082016, + 0.016995534300804138, + 0.016924023628234863, + 0.014390810392796993, + 0.009938240051269531, + 0.02419794164597988, + 0.008495619520545006, + 0.05011238530278206, + 0.006857738364487886, + 0.0043963221833109856, + 0.004035711754113436, + 0.002401925390586257, + 0.0031535779125988483, + 0.004740155767649412, + 0.0025820275768637657, + 0.010489584878087044, + 0.008440636098384857, + 0.011597869917750359, + 0.008246456272900105, + 0.007570281624794006, + 0.008042123168706894, + 0.01612531580030918, + 0.007352109998464584, + 0.018725372850894928, + 0.009990893304347992, + 0.010745641775429249, + 0.003933157306164503, + 0.006235775072127581, + 0.0086662070825696, + 0.06210118159651756, + 0.0064097619615495205, + 0.06309138983488083, + 0.011499223299324512, + 0.007505996618419886, + 0.014896361157298088, + 0.005427193827927113, + 0.015314863994717598, + 0.01760992780327797, + 0.013846831396222115, + 0.016235895454883575, + 0.02808990515768528, + 0.01485045813024044, + 0.008172804489731789, + 0.009486989118158817, + 0.012473339214920998, + 0.01908336952328682, + 0.013011009432375431, + 0.038539230823516846, + 0.00661990512162447, + 0.006403795443475246, + 0.00597456656396389, + 0.0019302924629300833, + 0.003132789395749569, + 0.005040375981479883, + 0.004212297033518553, + 0.006643723696470261, + 0.0027473177760839462, + 0.0043384963646531105, + 0.0031783576123416424, + 0.0039033042266964912, + 0.0020713263656944036, + 0.0025245421566069126, + 0.003431701334193349, + 0.003011590801179409, + 0.004500625655055046, + 0.009817611426115036, + 0.0142067177221179, + 0.008617752231657505, + 0.005082235671579838, + 0.004164583049714565, + 0.004341418389230967, + 0.003094290615990758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.001342587056569755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.14175795018672943, + 0.13811631500720978, + 0.02779022604227066, + 0.0550508126616478, + 0.41179654002189636, + 0.09443960338830948, + 0.015124042518436909, + 0.035486411303281784, + 0.049622368067502975, + 0.030815741047263145 + ], + "credit_map": [ + 0.020471878349781036, + 0.009505466558039188, + 0.009990581311285496, + 0.005488774739205837, + 0.006641787942498922, + 0.03204841539263725, + 0.00593154039233923, + 0.051679499447345734, + 0.025999782606959343, + 0.012551458552479744, + 0.013531901873648167, + 0.012291559018194675, + 0.00804738886654377, + 0.02023683674633503, + 0.007055311463773251, + 0.03840208053588867, + 0.004262347240000963, + 0.0027678431943058968, + 0.002280927961692214, + 0.0016854234272614121, + 0.001997065730392933, + 0.004743089899420738, + 0.0014872936299070716, + 0.0085662342607975, + 0.005445245187729597, + 0.0077271475456655025, + 0.006791483145207167, + 0.004820536822080612, + 0.005177853163331747, + 0.009472467936575413, + 0.00441923039034009, + 0.011196845211088657, + 0.013419023714959621, + 0.015199074521660805, + 0.0066382307559251785, + 0.009391835890710354, + 0.012055291794240475, + 0.1730491667985916, + 0.008533865213394165, + 0.17350997030735016, + 0.010655789636075497, + 0.006338753271847963, + 0.01115178782492876, + 0.004670118913054466, + 0.012443630956113338, + 0.020312458276748657, + 0.008776020258665085, + 0.020091041922569275, + 0.0021406887099146843, + 0.0014928518794476986, + 0.001231987727805972, + 0.0018841688288375735, + 0.0017036661738529801, + 0.0017229451332241297, + 0.0019511290593072772, + 0.0029966039583086967, + 0.005473658442497253, + 0.004390723071992397, + 0.003961327951401472, + 0.0030918947886675596, + 0.0030211873818188906, + 0.003863482503220439, + 0.00467561325058341, + 0.007008520420640707, + 0.006288890726864338, + 0.005633303429931402, + 0.007128244265913963, + 0.006568122189491987, + 0.005891192704439163, + 0.0047049871645867825, + 0.008796757087111473, + 0.004610870964825153, + 0.003129249205812812, + 0.006707509979605675, + 0.010682973079383373, + 0.0031096069142222404, + 0.0015351721085608006, + 0.0017889081500470638, + 0.002261117100715637, + 0.001601203577592969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "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 main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029522329568863 + ], + "opposite_logprob_tokens": [ + -2.350295305252075 + ], + "logprob_reward": -0.10029522329568863, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029522329568863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04308100789785385, + 0.02322523295879364, + 0.009484328329563141, + 0.04383232071995735, + 0.49758294224739075, + 0.11164437979459763, + 0.13859090209007263, + 0.03624800220131874, + 0.046942804008722305, + 0.04936808720231056 + ], + "credit_map": [ + 0.006958224344998598, + 0.004801279399544001, + 0.0037728010211139917, + 0.0037928845267742872, + 0.0038932841271162033, + 0.0065120914950966835, + 0.003979741130024195, + 0.009370699524879456, + 0.005040935706347227, + 0.0024670944549143314, + 0.002089221030473709, + 0.0011498249368742108, + 0.0013534358004108071, + 0.0036993829999119043, + 0.0009555047145113349, + 0.006469833664596081, + 0.0012750421883538365, + 0.0010915080783888698, + 0.0008625883492641151, + 0.0007032963912934065, + 0.0007020636694505811, + 0.001738306600600481, + 0.0007172782206907868, + 0.0023942457046359777, + 0.006064639426767826, + 0.006548439152538776, + 0.0035741969477385283, + 0.004201559815555811, + 0.004395296797156334, + 0.006596725899726152, + 0.00348353898152709, + 0.008967922069132328, + 0.008965231478214264, + 0.0093540009111166, + 0.007665663491934538, + 0.007744161412119865, + 0.010956533253192902, + 0.22274035215377808, + 0.008230212144553661, + 0.22192677855491638, + 0.007193868048489094, + 0.004603229463100433, + 0.011102586053311825, + 0.0034329784102737904, + 0.007977886125445366, + 0.034561604261398315, + 0.008685920387506485, + 0.03408629819750786, + 0.0229384433478117, + 0.014561587013304234, + 0.00814104825258255, + 0.008078615181148052, + 0.016470830887556076, + 0.01847640611231327, + 0.01046565268188715, + 0.03945830836892128, + 0.005769891198724508, + 0.0048373700119555, + 0.0051010772585868835, + 0.0019543843809515238, + 0.0031429550144821405, + 0.004774778615683317, + 0.004477472510188818, + 0.006190072745084763, + 0.006786743178963661, + 0.005054120440036058, + 0.006868971046060324, + 0.00608036108314991, + 0.005207960959523916, + 0.004932527896016836, + 0.007626296486705542, + 0.00438582431524992, + 0.005187894683331251, + 0.01008681207895279, + 0.015585090965032578, + 0.005503315012902021, + 0.002820066176354885, + 0.0035031531006097794, + 0.004070627503097057, + 0.0026111288461834192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.10029522329568863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, capturing the dramatic moment at a remarkably high frame rate, focuses on a crow perched on a bare, weathered, cracked branch. In the bright, cool, diffuse light, its glossy black body stands in stark contrast to the soft, out-of-focus, grayish-white gloom of the sky. Suddenly, as if startled by an invisible thunderclap, a flicker of alarm flashes in its small, obsidian eyes, and its body tenses. Then, with a sudden burst of power, its wings spread upward and backward at astonishing speed. In extreme slow motion, each individual black flight feather unfurls in the air like a fan, its delicate blue-violet iridescence shimmering in the light, revealing the intricate structure of its vanes and rachises. With the powerful flapping of wings, one can almost feel the shock wave, its body is pushed upward by the huge force, and its claws instantly release the branch. The camera then tilts upward smoothly to capture its back as it flies resolutely into the sky. The whole picture, in the cold black, white and gray tones, creates a dramatic atmosphere full of tension, primitive and poetic.", + "seed": 14040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_21_seed_14040_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": -0.01582146661821753, + "vlm_logprob": -0.01582146661821753, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.01582146661821753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01227844599634409 + ], + "opposite_logprob_tokens": [ + -6.0122785568237305 + ], + "logprob_reward": -0.01227844599634409, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.01227844599634409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058398742228746414, + 0.04895123466849327, + 0.12154518812894821, + 0.029918793588876724, + 0.07704180479049683, + 0.08638841658830643, + 0.09107235074043274, + 0.17263706028461456, + 0.1346539556980133, + 0.1793925166130066 + ], + "credit_map": [ + 0.008412600494921207, + 0.005286167375743389, + 0.006016450934112072, + 0.0043603768572211266, + 0.00685725174844265, + 0.008115929551422596, + 0.008041653782129288, + 0.01130831241607666, + 0.007058406248688698, + 0.005655103363096714, + 0.010269920341670513, + 0.0032374802976846695, + 0.003910808824002743, + 0.006370593793690205, + 0.003887962782755494, + 0.008560959249734879, + 0.017140474170446396, + 0.01421878207474947, + 0.0267248023301363, + 0.008028736338019371, + 0.011763920076191425, + 0.01733374036848545, + 0.009033029899001122, + 0.01730170287191868, + 0.004347793757915497, + 0.003480038605630398, + 0.004012479446828365, + 0.00222735945135355, + 0.003906841389834881, + 0.0038925393018871546, + 0.002391149988397956, + 0.005660593509674072, + 0.006965672131627798, + 0.00915231928229332, + 0.007560227997601032, + 0.007520686369389296, + 0.01839226298034191, + 0.009494373574852943, + 0.007781148422509432, + 0.010175115428864956, + 0.00910002551972866, + 0.009095757268369198, + 0.005881977267563343, + 0.010036158375442028, + 0.013579383492469788, + 0.011889850720763206, + 0.009495465084910393, + 0.017309801653027534, + 0.015483474358916283, + 0.00882740318775177, + 0.012026962824165821, + 0.008732525631785393, + 0.012606252916157246, + 0.009658747352659702, + 0.010834990069270134, + 0.01290199812501669, + 0.03702201321721077, + 0.024442771449685097, + 0.029457345604896545, + 0.023095646873116493, + 0.012865358963608742, + 0.01653088815510273, + 0.011532790958881378, + 0.01769024133682251, + 0.018306225538253784, + 0.011318663135170937, + 0.019778935238718987, + 0.027023717761039734, + 0.016960233449935913, + 0.012874186038970947, + 0.012222415767610073, + 0.016169577836990356, + 0.017324356362223625, + 0.022522026672959328, + 0.023045919835567474, + 0.03878873959183693, + 0.01577901281416416, + 0.018420446664094925, + 0.02223379537463188, + 0.021278228610754013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.01227844599634409 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751819048076868 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006751819048076868, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751819048076868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04134386405348778, + 0.02228042297065258, + 0.12370749562978745, + 0.08676028996706009, + 0.10991070419549942, + 0.12056364119052887, + 0.07924985140562057, + 0.08224131166934967, + 0.03934195265173912, + 0.2946004271507263 + ], + "credit_map": [ + 0.009104058146476746, + 0.003646266181021929, + 0.005080339498817921, + 0.002816730411723256, + 0.005666193086653948, + 0.006117860320955515, + 0.004877359606325626, + 0.004035057965666056, + 0.004524648189544678, + 0.0024830643087625504, + 0.002863439964130521, + 0.002050767419859767, + 0.0026031937450170517, + 0.003729713847860694, + 0.0017046063439920545, + 0.0023209890350699425, + 0.01628781668841839, + 0.015596677549183369, + 0.029372122138738632, + 0.008281855843961239, + 0.012090098112821579, + 0.01661335490643978, + 0.008733057416975498, + 0.016732508316636086, + 0.014150714501738548, + 0.010441791266202927, + 0.013964422978460789, + 0.007488830015063286, + 0.012679710984230042, + 0.013887610286474228, + 0.005058485548943281, + 0.009088722988963127, + 0.008492511697113514, + 0.014383845031261444, + 0.00867808610200882, + 0.007927951402962208, + 0.03840389847755432, + 0.012550292536616325, + 0.007234679535031319, + 0.012239431031048298, + 0.014779701828956604, + 0.013736071065068245, + 0.011921536177396774, + 0.014911523088812828, + 0.016693366691470146, + 0.01539877150207758, + 0.01426768023520708, + 0.01885499246418476, + 0.01261911354959011, + 0.009437242522835732, + 0.010433435440063477, + 0.005418200045824051, + 0.009849999099969864, + 0.008519196882843971, + 0.009119292721152306, + 0.013853373937308788, + 0.011630541644990444, + 0.00891910120844841, + 0.011630176566541195, + 0.01416187733411789, + 0.007084508426487446, + 0.00896841287612915, + 0.007833750918507576, + 0.012012941762804985, + 0.004886027425527573, + 0.003964210860431194, + 0.005969506688416004, + 0.007561482489109039, + 0.0036387848667800426, + 0.0028341275174170732, + 0.00443158857524395, + 0.006056226324290037, + 0.033614322543144226, + 0.056584808975458145, + 0.045042410492897034, + 0.05122528597712517, + 0.031183846294879913, + 0.03335944935679436, + 0.020721932873129845, + 0.022868353873491287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.006751819048076868 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001944795367307961 + ], + "opposite_logprob_tokens": [ + -6.2519450187683105 + ], + "logprob_reward": -0.001944795367307961, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001944795367307961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03592987731099129, + 0.018861111253499985, + 0.05730853229761124, + 0.05206514522433281, + 0.1315833181142807, + 0.13325689733028412, + 0.16729924082756042, + 0.15735122561454773, + 0.09128028899431229, + 0.1550643891096115 + ], + "credit_map": [ + 0.006183083169162273, + 0.00360431638546288, + 0.0063514080829918385, + 0.0022955923341214657, + 0.0036636851727962494, + 0.004921368323266506, + 0.0028984774835407734, + 0.006011942867189646, + 0.0035560601390898228, + 0.0023091640323400497, + 0.0028524522203952074, + 0.0015929037472233176, + 0.0018297563074156642, + 0.002740125171840191, + 0.0014335103332996368, + 0.00254714023321867, + 0.007800104562193155, + 0.006024775095283985, + 0.014722812920808792, + 0.0032337636221200228, + 0.005086359102278948, + 0.008277985267341137, + 0.004088196437805891, + 0.008074531331658363, + 0.008056437596678734, + 0.00856119766831398, + 0.008507872931659222, + 0.004003501497209072, + 0.005728859920054674, + 0.0071756551042199135, + 0.004202562849968672, + 0.005829056724905968, + 0.009846925735473633, + 0.017181148752570152, + 0.010142819955945015, + 0.008999188430607319, + 0.04621203616261482, + 0.014836626127362251, + 0.009348937310278416, + 0.01501564122736454, + 0.015149871818721294, + 0.013501693494617939, + 0.015502742491662502, + 0.013214044272899628, + 0.014805473387241364, + 0.016652172431349754, + 0.020335152745246887, + 0.02409573644399643, + 0.027348477393388748, + 0.020167624577879906, + 0.0210048146545887, + 0.01015748642385006, + 0.01869797892868519, + 0.01853187009692192, + 0.020567534491419792, + 0.030823426321148872, + 0.02058509737253189, + 0.015063243918120861, + 0.02453278750181198, + 0.022918403148651123, + 0.012569530867040157, + 0.01968435011804104, + 0.016018884256482124, + 0.02597893215715885, + 0.009633229114115238, + 0.008576441556215286, + 0.014466037973761559, + 0.02381831966340542, + 0.00912948977202177, + 0.006440673489123583, + 0.007820973172783852, + 0.011395122855901718, + 0.021864790469408035, + 0.03688213229179382, + 0.026875728741288185, + 0.021697303280234337, + 0.014719953760504723, + 0.015784405171871185, + 0.007765740156173706, + 0.009474320337176323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.001944795367307961 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027196360751986504, + 0.01661853678524494, + 0.040325574576854706, + 0.07360934466123581, + 0.09282102435827255, + 0.2041899710893631, + 0.09799695760011673, + 0.07099759578704834, + 0.08084133267402649, + 0.29540330171585083 + ], + "credit_map": [ + 0.003517222823575139, + 0.002513520885258913, + 0.0027405929286032915, + 0.00159729877486825, + 0.0026937706861644983, + 0.0031230130698531866, + 0.005049053113907576, + 0.0059618866071105, + 0.0024607840459793806, + 0.0018678827909752727, + 0.0019516166066750884, + 0.0013286868343129754, + 0.0019809301011264324, + 0.0023563229478895664, + 0.0017392364097759128, + 0.002933075651526451, + 0.005085390992462635, + 0.005146302282810211, + 0.009086604230105877, + 0.0031149170827120543, + 0.004010980948805809, + 0.005312077235430479, + 0.0028084295336157084, + 0.00576087087392807, + 0.011462748050689697, + 0.012272517196834087, + 0.012098900973796844, + 0.005997927859425545, + 0.008568741381168365, + 0.010630487464368343, + 0.005450157914310694, + 0.007127857301384211, + 0.006225140765309334, + 0.011656149290502071, + 0.008995476178824902, + 0.006157862488180399, + 0.02894262969493866, + 0.011347455903887749, + 0.007359645329415798, + 0.012136653997004032, + 0.01696670986711979, + 0.012898019514977932, + 0.014181687496602535, + 0.015043667517602444, + 0.04434040188789368, + 0.0471668466925621, + 0.01612837053835392, + 0.03746425732970238, + 0.013803370296955109, + 0.008468526415526867, + 0.014467086642980576, + 0.005750840064138174, + 0.009033462032675743, + 0.010605471208691597, + 0.012371107004582882, + 0.023497093468904495, + 0.007961331866681576, + 0.00842070858925581, + 0.01079723984003067, + 0.013391831889748573, + 0.005658602807670832, + 0.009199369698762894, + 0.005095421336591244, + 0.01047308836132288, + 0.00957357045263052, + 0.008058622479438782, + 0.011877479963004589, + 0.0180357713252306, + 0.00889156386256218, + 0.0062093609012663364, + 0.007696860935539007, + 0.010498100891709328, + 0.03829086571931839, + 0.07530352473258972, + 0.05626130849123001, + 0.03740471228957176, + 0.02472378872334957, + 0.023534204810857773, + 0.018390661105513573, + 0.021494217216968536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.0031918552704155445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054940417408943176 + ], + "opposite_logprob_tokens": [ + -2.9299404621124268 + ], + "logprob_reward": -0.054940417408943176, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054940417408943176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021159280091524124, + 0.01962338015437126, + 0.024591142311692238, + 0.03842375800013542, + 0.07949712127447128, + 0.11425282061100006, + 0.17435023188591003, + 0.19630186259746552, + 0.05579273775219917, + 0.2760077118873596 + ], + "credit_map": [ + 0.0037839498836547136, + 0.0016612597974017262, + 0.0020587765611708164, + 0.0016022218624129891, + 0.002880803542211652, + 0.0025520729832351208, + 0.0037569378037005663, + 0.002863257424905896, + 0.002835723338648677, + 0.0022000023163855076, + 0.0031985589303076267, + 0.0012177161406725645, + 0.0023647197522222996, + 0.002504462143406272, + 0.001517539843916893, + 0.0037846574559807777, + 0.00401287479326129, + 0.002992381574586034, + 0.004382984712719917, + 0.0020289712119847536, + 0.002363994950428605, + 0.003265607636421919, + 0.0020136607345193624, + 0.0035306671634316444, + 0.005537938792258501, + 0.006582061760127544, + 0.005630264990031719, + 0.0036403320264071226, + 0.005470141302794218, + 0.005383655894547701, + 0.0023971491027623415, + 0.0037822118028998375, + 0.007007818203419447, + 0.009651081636548042, + 0.008757966570556164, + 0.006615818478167057, + 0.01915745809674263, + 0.009017279371619225, + 0.007458738517016172, + 0.011830957606434822, + 0.015022829174995422, + 0.012781263329088688, + 0.013508331961929798, + 0.01094583235681057, + 0.01006306055933237, + 0.012657075189054012, + 0.014505673199892044, + 0.024768758565187454, + 0.02970248833298683, + 0.0202272217720747, + 0.02252850867807865, + 0.011957742273807526, + 0.015007799491286278, + 0.01972685009241104, + 0.023038731887936592, + 0.03216088190674782, + 0.026296034455299377, + 0.01925201155245304, + 0.02856203354895115, + 0.026077106595039368, + 0.020750999450683594, + 0.023383408784866333, + 0.020932072773575783, + 0.031048189848661423, + 0.007016913499683142, + 0.006759955082088709, + 0.008049697615206242, + 0.011037097312510014, + 0.0059826369397342205, + 0.004582260735332966, + 0.005913932342082262, + 0.006450238637626171, + 0.035483986139297485, + 0.06830130517482758, + 0.04687129333615303, + 0.049027763307094574, + 0.02460445463657379, + 0.02300817146897316, + 0.0143727483227849, + 0.01433799508959055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.054940417408943176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, capturing the dramatic moment at a remarkably high frame rate, focuses on a crow perched on a bare, weathered, cracked branch. In the bright, cool, diffuse light, its glossy black body stands in stark contrast to the soft, out-of-focus, grayish-white gloom of the sky. Suddenly, as if startled by an invisible thunderclap, a flicker of alarm flashes in its small, obsidian eyes, and its body tenses. Then, with a sudden burst of power, its wings spread upward and backward at astonishing speed. In extreme slow motion, each individual black flight feather unfurls in the air like a fan, its delicate blue-violet iridescence shimmering in the light, revealing the intricate structure of its vanes and rachises. With the powerful flapping of wings, one can almost feel the shock wave, its body is pushed upward by the huge force, and its claws instantly release the branch. The camera then tilts upward smoothly to capture its back as it flies resolutely into the sky. The whole picture, in the cold black, white and gray tones, creates a dramatic atmosphere full of tension, primitive and poetic.", + "seed": 14041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_12/index_21_seed_14041_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": -0.018679501581937076, + "vlm_logprob": -0.018679501581937076, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.018679501581937076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014419402927160263 + ], + "opposite_logprob_tokens": [ + -5.3894195556640625 + ], + "logprob_reward": -0.014419402927160263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.014419402927160263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11583466827869415, + 0.05040404945611954, + 0.07285258919000626, + 0.05449505150318146, + 0.16278865933418274, + 0.11195732653141022, + 0.10658996552228928, + 0.11606129258871078, + 0.10241053253412247, + 0.10660579800605774 + ], + "credit_map": [ + 0.015208634547889233, + 0.01138316374272108, + 0.01847631111741066, + 0.010529580526053905, + 0.018368100747466087, + 0.015554463490843773, + 0.011711104772984982, + 0.01460331305861473, + 0.0063976445235311985, + 0.007001375313848257, + 0.006409234367311001, + 0.004488499369472265, + 0.007851758040487766, + 0.0071377879939973354, + 0.00604692567139864, + 0.005070829298347235, + 0.009860971942543983, + 0.013356281444430351, + 0.009264039807021618, + 0.007388904690742493, + 0.009777553379535675, + 0.00918496772646904, + 0.007333818823099136, + 0.006686056032776833, + 0.005574111361056566, + 0.005729979835450649, + 0.007503310218453407, + 0.004587887786328793, + 0.010378511622548103, + 0.006830228492617607, + 0.007344044279307127, + 0.006546975579112768, + 0.027910141274333, + 0.010269686579704285, + 0.04067107290029526, + 0.00912861991673708, + 0.01093324925750494, + 0.016115546226501465, + 0.015206299722194672, + 0.03255404159426689, + 0.009549878537654877, + 0.009930869564414024, + 0.02046351693570614, + 0.011183987371623516, + 0.013415887020528316, + 0.009636701084673405, + 0.013311544433236122, + 0.024464940652251244, + 0.011079737916588783, + 0.010314215905964375, + 0.014707865193486214, + 0.01734994351863861, + 0.013328603468835354, + 0.009189896285533905, + 0.008712085895240307, + 0.0219076257199049, + 0.014415071345865726, + 0.014364656060934067, + 0.017544042319059372, + 0.02116195112466812, + 0.012517007067799568, + 0.010408378206193447, + 0.012165600433945656, + 0.013484591618180275, + 0.014978179708123207, + 0.01450591441243887, + 0.011778856627643108, + 0.009464097209274769, + 0.015553317964076996, + 0.014107702299952507, + 0.012372107245028019, + 0.009650361724197865, + 0.013855510391294956, + 0.015703171491622925, + 0.01659265160560608, + 0.01660962961614132, + 0.011889180168509483, + 0.0113298324868083, + 0.008691092021763325, + 0.0119347358122468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.014419402927160263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555130757391453 + ], + "opposite_logprob_tokens": [ + -4.387555122375488 + ], + "logprob_reward": -0.012555130757391453, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555130757391453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06339092552661896, + 0.06414207071065903, + 0.059170980006456375, + 0.158080592751503, + 0.11380863934755325, + 0.1817542314529419, + 0.11991322785615921, + 0.05650096759200096, + 0.10373906791210175, + 0.07949932664632797 + ], + "credit_map": [ + 0.010019520297646523, + 0.006583148613572121, + 0.009490745142102242, + 0.004742770455777645, + 0.011179243214428425, + 0.009206338785588741, + 0.005967952776700258, + 0.006201206706464291, + 0.006162066478282213, + 0.011190610937774181, + 0.010307321324944496, + 0.004961060825735331, + 0.01376663614064455, + 0.008082927204668522, + 0.005616370588541031, + 0.004055073484778404, + 0.008574524894356728, + 0.009318511001765728, + 0.013347984291613102, + 0.007405729498714209, + 0.004492851439863443, + 0.006679497193545103, + 0.0047925724647939205, + 0.0045593068934977055, + 0.012436605989933014, + 0.008482415229082108, + 0.009395221248269081, + 0.0054358020424842834, + 0.006391386967152357, + 0.050280965864658356, + 0.006130521651357412, + 0.059527672827243805, + 0.020229285582900047, + 0.008265557698905468, + 0.02613925375044346, + 0.009991290979087353, + 0.010230311192572117, + 0.010799815878272057, + 0.007773746736347675, + 0.020379381254315376, + 0.022929640486836433, + 0.014488170854747295, + 0.026940196752548218, + 0.01835012249648571, + 0.02467586100101471, + 0.012396503239870071, + 0.03301559016108513, + 0.0289581548422575, + 0.014638992957770824, + 0.010132879950106144, + 0.020157085731625557, + 0.02614772506058216, + 0.006699759047478437, + 0.010476425290107727, + 0.0064793918281793594, + 0.025180967524647713, + 0.005343817174434662, + 0.006579664070159197, + 0.007082589901983738, + 0.006437878590077162, + 0.013718421570956707, + 0.005984433460980654, + 0.005039901938289404, + 0.006314261816442013, + 0.013104155659675598, + 0.01664412021636963, + 0.013578693382441998, + 0.012382172048091888, + 0.012148914858698845, + 0.011358056217432022, + 0.01322020310908556, + 0.011302751488983631, + 0.012841661460697651, + 0.017090480774641037, + 0.010866808705031872, + 0.009551330469548702, + 0.009939378127455711, + 0.008579579181969166, + 0.004658313002437353, + 0.005971770267933607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 6 + ], + "selected_reward": -0.012555130757391453 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200802555307746 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200802555307746, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200802555307746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050102975219488144, + 0.03543660044670105, + 0.06530825793743134, + 0.3260202705860138, + 0.08980466425418854, + 0.13430504500865936, + 0.12031079083681107, + 0.05512292683124542, + 0.09354707598686218, + 0.030041446909308434 + ], + "credit_map": [ + 0.00833864789456129, + 0.0063234842382371426, + 0.008641015738248825, + 0.004994021262973547, + 0.005628741346299648, + 0.006406501401215792, + 0.004690254107117653, + 0.005080304108560085, + 0.0034169212449342012, + 0.00563679076731205, + 0.0035200705751776695, + 0.0034088187385350466, + 0.006655130535364151, + 0.004948295187205076, + 0.004578635562211275, + 0.00327193271368742, + 0.009889347478747368, + 0.01185887586325407, + 0.014443612657487392, + 0.005315317306667566, + 0.005744428373873234, + 0.007500835694372654, + 0.004261169116944075, + 0.006294657476246357, + 0.02653113752603531, + 0.017129186540842056, + 0.024515049532055855, + 0.008838895708322525, + 0.011514924466609955, + 0.0993824452161789, + 0.01220972090959549, + 0.12589889764785767, + 0.01520437840372324, + 0.00717284856364131, + 0.018822288140654564, + 0.008567893877625465, + 0.008560431189835072, + 0.008579119108617306, + 0.006329196505248547, + 0.01656850054860115, + 0.023956436663866043, + 0.014618946239352226, + 0.021634133532643318, + 0.018089301884174347, + 0.012688374146819115, + 0.014549716375768185, + 0.014249172993004322, + 0.014518951065838337, + 0.018576931208372116, + 0.011917559430003166, + 0.01907939463853836, + 0.02161175198853016, + 0.008725753054022789, + 0.013767248950898647, + 0.00690575921908021, + 0.019726382568478584, + 0.00647705839946866, + 0.006816511508077383, + 0.007860726676881313, + 0.01064674835652113, + 0.008072489872574806, + 0.004986373707652092, + 0.005076965782791376, + 0.005186046473681927, + 0.010720543563365936, + 0.013522690162062645, + 0.012712782248854637, + 0.01145639643073082, + 0.010400939732789993, + 0.0108255036175251, + 0.009542207233607769, + 0.014365999959409237, + 0.005544198211282492, + 0.005891138222068548, + 0.004641248844563961, + 0.004147516563534737, + 0.003359248861670494, + 0.0026797924656420946, + 0.0018780557438731194, + 0.001900242641568184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.002200802555307746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196520334109664 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.002196520334109664, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196520334109664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042940132319927216, + 0.03395671397447586, + 0.07375471293926239, + 0.10990612953901291, + 0.1345241516828537, + 0.08316849917173386, + 0.12430501729249954, + 0.09112128615379333, + 0.1968357115983963, + 0.10948769748210907 + ], + "credit_map": [ + 0.007536998018622398, + 0.004384091589599848, + 0.009246163070201874, + 0.0034688569139689207, + 0.005351052153855562, + 0.005907730665057898, + 0.0038018482737243176, + 0.00324338604696095, + 0.003961711190640926, + 0.004291543737053871, + 0.005385385826230049, + 0.0034542283974587917, + 0.006226132158190012, + 0.004528420977294445, + 0.003997553139925003, + 0.002111733192577958, + 0.01066984236240387, + 0.01102836150676012, + 0.019075704738497734, + 0.005869163665920496, + 0.006107422988861799, + 0.009166744537651539, + 0.005372732877731323, + 0.006464737933129072, + 0.008834421634674072, + 0.007094358094036579, + 0.006682679057121277, + 0.004532807040959597, + 0.007032760418951511, + 0.03374277055263519, + 0.006272675935178995, + 0.035713646560907364, + 0.016745010390877724, + 0.011999346315860748, + 0.025292331352829933, + 0.010095721110701561, + 0.018054429441690445, + 0.015158415772020817, + 0.014636282809078693, + 0.022542594000697136, + 0.015629051253199577, + 0.006641244515776634, + 0.0139284897595644, + 0.010757268406450748, + 0.009302973747253418, + 0.007871907204389572, + 0.008994285017251968, + 0.010043271817266941, + 0.015304830856621265, + 0.01697411760687828, + 0.017869295552372932, + 0.017722485587000847, + 0.019010521471500397, + 0.011384272016584873, + 0.011153050698339939, + 0.014886436052620411, + 0.010316804982721806, + 0.00836219172924757, + 0.010170516557991505, + 0.013469462282955647, + 0.02120680920779705, + 0.009838947094976902, + 0.006781000178307295, + 0.01097553875297308, + 0.019646409898996353, + 0.022393595427274704, + 0.023622900247573853, + 0.0215796809643507, + 0.03637408837676048, + 0.026626883074641228, + 0.02639961801469326, + 0.020192516967654228, + 0.020289750769734383, + 0.018250714987516403, + 0.017880193889141083, + 0.013028927147388458, + 0.011933539994060993, + 0.009417240507900715, + 0.007834995165467262, + 0.010852321051061153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 6 + ], + "selected_reward": -0.002196520334109664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a crow perched on a bare, weathered, cracked branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the crow have a glossy black body with blue-violet iridescent flight feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crow spread its wings upward and backward at high speed, revealing individual feather structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crow release its claws from the branch and fly upward into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt upward smoothly to follow the crow's ascent?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202565133571625 + ], + "opposite_logprob_tokens": [ + -2.812025547027588 + ], + "logprob_reward": -0.06202565133571625, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202565133571625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040918949991464615, + 0.03322641924023628, + 0.08315467834472656, + 0.15538737177848816, + 0.07470715045928955, + 0.06762079149484634, + 0.11436288803815842, + 0.138140007853508, + 0.19636598229408264, + 0.09611577540636063 + ], + "credit_map": [ + 0.006624413188546896, + 0.0035391636192798615, + 0.009940088726580143, + 0.0032828357070684433, + 0.004380961414426565, + 0.005827546119689941, + 0.003588512074202299, + 0.0037354296073317528, + 0.0043816654942929745, + 0.0040591200813651085, + 0.006933008786290884, + 0.003097841050475836, + 0.004564725793898106, + 0.004119004588574171, + 0.003314393572509289, + 0.002756661269813776, + 0.013742310926318169, + 0.012267675250768661, + 0.024095341563224792, + 0.005783642176538706, + 0.0060766637325286865, + 0.010390051640570164, + 0.004233869258314371, + 0.00656511913985014, + 0.013206197880208492, + 0.00819234736263752, + 0.012316019274294376, + 0.004825628362596035, + 0.005181255284696817, + 0.046405620872974396, + 0.006247585639357567, + 0.059012722223997116, + 0.010292927734553814, + 0.004934426862746477, + 0.019094426184892654, + 0.006860881112515926, + 0.007090146187692881, + 0.0065165855921804905, + 0.00569075345993042, + 0.014227006584405899, + 0.011699232272803783, + 0.006894821301102638, + 0.009924807585775852, + 0.008623641915619373, + 0.006451823748648167, + 0.009008700959384441, + 0.0055860853753983974, + 0.009431679733097553, + 0.012670896016061306, + 0.01352018304169178, + 0.016980985179543495, + 0.021119175478816032, + 0.01475886907428503, + 0.013156488537788391, + 0.008073627017438412, + 0.014082670211791992, + 0.0151012372225523, + 0.015979254618287086, + 0.018997661769390106, + 0.02747584693133831, + 0.022753631696105003, + 0.011796163395047188, + 0.01161421649158001, + 0.014422004111111164, + 0.026043476536870003, + 0.032902590930461884, + 0.027804464101791382, + 0.019717814400792122, + 0.030686086043715477, + 0.020491035655140877, + 0.017044462263584137, + 0.021676044911146164, + 0.01818157732486725, + 0.015176433138549328, + 0.012682153843343258, + 0.009643523953855038, + 0.015128742903470993, + 0.01023754384368658, + 0.007598474621772766, + 0.007467320654541254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.06202565133571625 + } + ], + "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/0000055/group_13/index_tensor([6570, 6570])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_tensor([6570, 6570])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..938fc510fd98322f581dc68398a8a88ffc5d4a95 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_tensor([6570, 6570])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This vibrant, anime-style, low-angle, dynamic tracking shot follows a firefighter as they relentlessly charge into the inferno of a collapsing building. Their hulking, soot-stained shifters, with sharp lines and deep shadows, are streaked with luminous yellow reflective stripes to contrast against the flickering lights. The camera quickly captures an extreme close-up of their eyes, framed by the soot-streaked visors of their helmets. Their large, expressive anime eyes are a bright, resolute azure blue, reflecting the roaring, stylized flames around them. With pupils constricted and brows furrowed, they survey the perilous surroundings with unwavering determination, their intense gaze piercing the thick smoke and glowing embers that fill the air. Flames roar dramatically in the shadows of the battery, tinged with orange and yellow, licking at charred support beams and causing the air to flicker with heat, creating a tense and heroic atmosphere as the firefighters battle the overwhelming blaze in their rescue mission.", + "seed": 6177, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6177_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": -0.17007182501256465, + "vlm_logprob": -0.17007182501256465, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.17007182501256465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43529751896858215 + ], + "opposite_logprob_tokens": [ + -1.0602974891662598 + ], + "logprob_reward": -0.43529751896858215, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43529751896858215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0788017138838768, + 0.42277097702026367, + 0.10993663966655731, + 0.027793414890766144, + 0.0259256549179554, + 0.09876994788646698, + 0.06350510567426682, + 0.041576966643333435, + 0.04971947520971298, + 0.08120008558034897 + ], + "credit_map": [ + 0.008298702538013458, + 0.01294768787920475, + 0.01011087279766798, + 0.01178960781544447, + 0.012448114342987537, + 0.008220450952649117, + 0.00840462651103735, + 0.006581644527614117, + 0.026975108310580254, + 0.15497002005577087, + 0.1312856674194336, + 0.026370009407401085, + 0.03986762464046478, + 0.011394009925425053, + 0.020276902243494987, + 0.011631668545305729, + 0.006941455416381359, + 0.0316021591424942, + 0.013284520246088505, + 0.013768174685537815, + 0.015889061614871025, + 0.008734933100640774, + 0.009976422414183617, + 0.009739907458424568, + 0.0024301090743392706, + 0.005634667817503214, + 0.003093091072514653, + 0.0035550135653465986, + 0.004138654563575983, + 0.003062526462599635, + 0.0033004439901560545, + 0.002578908344730735, + 0.0022491205018013716, + 0.004270877223461866, + 0.003335650311782956, + 0.0025115900207310915, + 0.0045981635339558125, + 0.0023001793306320906, + 0.0031189739238470793, + 0.003541101235896349, + 0.009411408565938473, + 0.014936159364879131, + 0.01100334245711565, + 0.013064988888800144, + 0.016409117728471756, + 0.01279812678694725, + 0.010426623746752739, + 0.010720188729465008, + 0.007298015058040619, + 0.014140580780804157, + 0.004583409521728754, + 0.0063742371276021, + 0.010975035838782787, + 0.007369866594672203, + 0.008224384859204292, + 0.0045395782217383385, + 0.0036134133115410805, + 0.004705384373664856, + 0.0054023051634430885, + 0.0041060736402869225, + 0.00688429968431592, + 0.005794738885015249, + 0.004863508511334658, + 0.00620724493637681, + 0.009416453540325165, + 0.006035804748535156, + 0.005126390140503645, + 0.003385662566870451, + 0.007367983460426331, + 0.0039022217970341444, + 0.011295877397060394, + 0.0031890792306512594, + 0.015941083431243896, + 0.0057799979113042355, + 0.012242565862834454, + 0.005610513500869274, + 0.008039469830691814, + 0.004783430136740208, + 0.021980706602334976, + 0.006822321098297834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.43529751896858215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31353944540023804 + ], + "opposite_logprob_tokens": [ + -1.3135395050048828 + ], + "logprob_reward": -0.31353944540023804, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31353944540023804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1749527007341385, + 0.11325555294752121, + 0.12757281959056854, + 0.06019481644034386, + 0.03571850433945656, + 0.07131583988666534, + 0.10553614795207977, + 0.03689388558268547, + 0.10971550643444061, + 0.16484422981739044 + ], + "credit_map": [ + 0.014404552057385445, + 0.04367811977863312, + 0.01765415258705616, + 0.018662508577108383, + 0.026333916932344437, + 0.020349454134702682, + 0.02113746851682663, + 0.01273252721875906, + 0.00763475988060236, + 0.021976737305521965, + 0.01613246090710163, + 0.019385354593396187, + 0.019053269177675247, + 0.006576051004230976, + 0.013573109172284603, + 0.008923799730837345, + 0.007956231012940407, + 0.02912391722202301, + 0.017125921323895454, + 0.02348637953400612, + 0.011898650787770748, + 0.012646512128412724, + 0.012847594916820526, + 0.012487598694860935, + 0.004150621127337217, + 0.01615752838551998, + 0.007851792499423027, + 0.00715237483382225, + 0.007484905421733856, + 0.005868132691830397, + 0.004917788319289684, + 0.006611670833081007, + 0.0028578797355294228, + 0.004512672312557697, + 0.0031431985553354025, + 0.0049529895186424255, + 0.006841974798589945, + 0.0036739136558026075, + 0.005535920616239309, + 0.004199954681098461, + 0.006583387032151222, + 0.012800399214029312, + 0.008408688940107822, + 0.008633495308458805, + 0.013182366266846657, + 0.007203655783087015, + 0.0077729858458042145, + 0.006730859633535147, + 0.011798294261097908, + 0.01270567998290062, + 0.008023749105632305, + 0.013778057880699635, + 0.0184783972799778, + 0.015803344547748566, + 0.012235183268785477, + 0.01271344069391489, + 0.004974174778908491, + 0.004366105888038874, + 0.005145709495991468, + 0.003506413660943508, + 0.0048925126902759075, + 0.005744867026805878, + 0.0043989853002130985, + 0.003865116974338889, + 0.014425183646380901, + 0.020307404920458794, + 0.011979063041508198, + 0.009767604060471058, + 0.019745297729969025, + 0.006841977126896381, + 0.016104884445667267, + 0.01054407935589552, + 0.044917766004800797, + 0.007643070537596941, + 0.007351722568273544, + 0.005987107288092375, + 0.012909368611872196, + 0.008455573581159115, + 0.06500400602817535, + 0.012575596570968628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.31353944540023804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674650490283966 + ], + "opposite_logprob_tokens": [ + -3.0486745834350586 + ], + "logprob_reward": -0.048674650490283966, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674650490283966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10987043380737305, + 0.11386900395154953, + 0.10701882094144821, + 0.06310267746448517, + 0.05457914620637894, + 0.11981231719255447, + 0.10145255923271179, + 0.04770708829164505, + 0.12062346935272217, + 0.1619645357131958 + ], + "credit_map": [ + 0.012160834856331348, + 0.018454330042004585, + 0.011572607792913914, + 0.01123910304158926, + 0.01912541314959526, + 0.012844222597777843, + 0.01212639082223177, + 0.012347526848316193, + 0.013087661005556583, + 0.014992577955126762, + 0.021370699629187584, + 0.02145611308515072, + 0.011321868747472763, + 0.009885001927614212, + 0.012525183148682117, + 0.009229891933500767, + 0.008739376440644264, + 0.018050506711006165, + 0.01551895309239626, + 0.012636146508157253, + 0.013608690351247787, + 0.016168732196092606, + 0.010278381407260895, + 0.012018037959933281, + 0.0060447934083640575, + 0.01091996394097805, + 0.008178197778761387, + 0.005069679114967585, + 0.00872819684445858, + 0.005967135541141033, + 0.008102432824671268, + 0.010092276148498058, + 0.005568033084273338, + 0.008501355536282063, + 0.006468266248703003, + 0.0069144004955887794, + 0.008577902801334858, + 0.005544132087379694, + 0.006167924962937832, + 0.006837128661572933, + 0.012592654675245285, + 0.022982073947787285, + 0.007947172969579697, + 0.012716103345155716, + 0.030665474012494087, + 0.008764727041125298, + 0.015113147906959057, + 0.009030953049659729, + 0.011061642318964005, + 0.015606595203280449, + 0.008447369560599327, + 0.007260494865477085, + 0.017919396981596947, + 0.018784593790769577, + 0.011717009358108044, + 0.010655457153916359, + 0.0047117238864302635, + 0.008720156736671925, + 0.005024566315114498, + 0.0041369604878127575, + 0.0075261639431118965, + 0.004152228590101004, + 0.005723991431295872, + 0.0077112955041229725, + 0.01522284559905529, + 0.020225754007697105, + 0.011810806579887867, + 0.014937917701900005, + 0.021848930045962334, + 0.00760372169315815, + 0.016792170703411102, + 0.012181315571069717, + 0.04228293150663376, + 0.00947941280901432, + 0.007391416467726231, + 0.006162918172776699, + 0.012950612232089043, + 0.008535636588931084, + 0.06309571862220764, + 0.012065877206623554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.048674650490283966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306374117732048 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.04306374117732048, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306374117732048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16382195055484772, + 0.266252338886261, + 0.26539385318756104, + 0.014427917078137398, + 0.019043773412704468, + 0.022798584774136543, + 0.08086807280778885, + 0.05347226560115814, + 0.05778529495000839, + 0.05613590031862259 + ], + "credit_map": [ + 0.014557593502104282, + 0.03642071783542633, + 0.02217911370098591, + 0.019456861540675163, + 0.021203769370913506, + 0.020000910386443138, + 0.016158943995833397, + 0.013844051398336887, + 0.02289973385632038, + 0.05842900276184082, + 0.10064362734556198, + 0.022703927010297775, + 0.02212393283843994, + 0.007779156323522329, + 0.023198571056127548, + 0.00847441703081131, + 0.02137080766260624, + 0.0540875568985939, + 0.04186743125319481, + 0.024700099602341652, + 0.042751699686050415, + 0.01731603406369686, + 0.04491838067770004, + 0.018381841480731964, + 0.000993825145997107, + 0.0026853785384446383, + 0.0018415957456454635, + 0.0017394099850207567, + 0.0019674210343509912, + 0.0014021192910149693, + 0.001483244588598609, + 0.002314923098310828, + 0.0017176482360810041, + 0.00421617878600955, + 0.0021597843151539564, + 0.0014298588503152132, + 0.0028843432664871216, + 0.002203643321990967, + 0.0023439081851392984, + 0.0020884075202047825, + 0.0027413717471063137, + 0.0035047729033976793, + 0.002509416313841939, + 0.0018416527891531587, + 0.004730673972517252, + 0.002513229614123702, + 0.0028744912706315517, + 0.0020829744171351194, + 0.007952294312417507, + 0.011522339656949043, + 0.00983676128089428, + 0.01148295495659113, + 0.010925762355327606, + 0.01097806729376316, + 0.010316388681530952, + 0.00785350427031517, + 0.005360865965485573, + 0.005828955676406622, + 0.006274380721151829, + 0.0063101742416620255, + 0.007134394720196724, + 0.009307852014899254, + 0.00623790780082345, + 0.007017733063548803, + 0.009804481640458107, + 0.008479578420519829, + 0.006682203616946936, + 0.004693574737757444, + 0.009273363277316093, + 0.0037827976047992706, + 0.010362920351326466, + 0.004706376697868109, + 0.009751923382282257, + 0.006348968483507633, + 0.00482700252905488, + 0.0037185309920459986, + 0.010078812949359417, + 0.005370130762457848, + 0.010918732732534409, + 0.0051217940635979176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.04306374117732048 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1343972086906433, + 0.23860380053520203, + 0.1047564223408699, + 0.020005015656352043, + 0.044687479734420776, + 0.0383450984954834, + 0.053916655480861664, + 0.035018548369407654, + 0.028307080268859863, + 0.30196264386177063 + ], + "credit_map": [ + 0.010506498627364635, + 0.039023011922836304, + 0.011790762655436993, + 0.012712462805211544, + 0.01928899623453617, + 0.015305216424167156, + 0.0160785261541605, + 0.009691745974123478, + 0.01808176562190056, + 0.07178846001625061, + 0.07902909070253372, + 0.015698878094553947, + 0.025555478408932686, + 0.007661042734980583, + 0.013305425643920898, + 0.007483671884983778, + 0.00828047376126051, + 0.010199991054832935, + 0.017468877136707306, + 0.01971539668738842, + 0.01048015896230936, + 0.013406708836555481, + 0.014553945511579514, + 0.010650883428752422, + 0.0018675214378163218, + 0.0026410126592963934, + 0.003821156220510602, + 0.0016658048843964934, + 0.003066588658839464, + 0.001816816395148635, + 0.002225245116278529, + 0.002900871215388179, + 0.0038963239639997482, + 0.007512269075959921, + 0.003437721636146307, + 0.0050025880336761475, + 0.0095595084130764, + 0.003602696117013693, + 0.006616436410695314, + 0.005059937480837107, + 0.004448296502232552, + 0.008292854763567448, + 0.002421035896986723, + 0.0034930387046188116, + 0.009016986936330795, + 0.0031731603667140007, + 0.0047157141380012035, + 0.0027840142138302326, + 0.006892184726893902, + 0.005930326879024506, + 0.004155365750193596, + 0.004184065852314234, + 0.00891873985528946, + 0.009392457082867622, + 0.009914415888488293, + 0.004529104568064213, + 0.003439472522586584, + 0.004621800035238266, + 0.004304633475840092, + 0.004165344871580601, + 0.00494001992046833, + 0.0059671602211892605, + 0.00398834515362978, + 0.0035917754285037518, + 0.004630066454410553, + 0.0032822720240801573, + 0.00356547930277884, + 0.0025414125993847847, + 0.003719160333275795, + 0.002313140546903014, + 0.005614120978862047, + 0.002641429426148534, + 0.0535736121237278, + 0.03583753481507301, + 0.017748067155480385, + 0.018089158460497856, + 0.05577336251735687, + 0.014626407995820045, + 0.07842359691858292, + 0.027890918776392937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.009783769026398659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This vibrant, anime-style, low-angle, dynamic tracking shot follows a firefighter as they relentlessly charge into the inferno of a collapsing building. Their hulking, soot-stained shifters, with sharp lines and deep shadows, are streaked with luminous yellow reflective stripes to contrast against the flickering lights. The camera quickly captures an extreme close-up of their eyes, framed by the soot-streaked visors of their helmets. Their large, expressive anime eyes are a bright, resolute azure blue, reflecting the roaring, stylized flames around them. With pupils constricted and brows furrowed, they survey the perilous surroundings with unwavering determination, their intense gaze piercing the thick smoke and glowing embers that fill the air. Flames roar dramatically in the shadows of the battery, tinged with orange and yellow, licking at charred support beams and causing the air to flicker with heat, creating a tense and heroic atmosphere as the firefighters battle the overwhelming blaze in their rescue mission.", + "seed": 6178, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6178_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": -0.10263873171061277, + "vlm_logprob": -0.10263873171061277, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.10263873171061277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13013482093811035 + ], + "opposite_logprob_tokens": [ + -2.1301348209381104 + ], + "logprob_reward": -0.13013482093811035, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13013482093811035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1255349963903427, + 0.04459928721189499, + 0.05959659814834595, + 0.09947749227285385, + 0.24977648258209229, + 0.07793013751506805, + 0.22531481087207794, + 0.04059991613030434, + 0.0320633165538311, + 0.0451069138944149 + ], + "credit_map": [ + 0.011672255583107471, + 0.021769985556602478, + 0.013559863902628422, + 0.009146003052592278, + 0.021470889449119568, + 0.01886088028550148, + 0.013963425531983376, + 0.015091704204678535, + 0.004838285502046347, + 0.005831955000758171, + 0.006872633006423712, + 0.004744932055473328, + 0.0074502089992165565, + 0.005189716815948486, + 0.005145482253283262, + 0.004526078701019287, + 0.018369875848293304, + 0.01928800903260708, + 0.0048865363933146, + 0.0026388908736407757, + 0.0055619957856833935, + 0.002943200757727027, + 0.003235431155189872, + 0.0026726610958576202, + 0.01744389906525612, + 0.007597205229103565, + 0.009531893767416477, + 0.012892416678369045, + 0.01076660305261612, + 0.007775552570819855, + 0.02655804716050625, + 0.00691188033670187, + 0.08074191212654114, + 0.018645722419023514, + 0.009943226352334023, + 0.02754010446369648, + 0.021634135395288467, + 0.06049276143312454, + 0.020184874534606934, + 0.010593763552606106, + 0.012860807590186596, + 0.006162385921925306, + 0.01259666308760643, + 0.005946044810116291, + 0.007820680737495422, + 0.013692965731024742, + 0.010569967329502106, + 0.008280628360807896, + 0.055551350116729736, + 0.0178241990506649, + 0.03236467018723488, + 0.017326388508081436, + 0.02298455685377121, + 0.02710733935236931, + 0.02209564484655857, + 0.0300606619566679, + 0.009757546707987785, + 0.004629419185221195, + 0.0036301533691585064, + 0.0033248986583203077, + 0.004316185601055622, + 0.0061287651769816875, + 0.00336634018458426, + 0.00544661283493042, + 0.0032014644239097834, + 0.0056567988358438015, + 0.0046949745155870914, + 0.0036053189542144537, + 0.0038007376715540886, + 0.0034402182791382074, + 0.0032246699556708336, + 0.004439135082066059, + 0.0047086626291275024, + 0.004932391457259655, + 0.005089794751256704, + 0.00543868076056242, + 0.003882014425471425, + 0.010550904087722301, + 0.006211430300027132, + 0.004293039906769991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.13013482093811035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814103364944458 + ], + "opposite_logprob_tokens": [ + -1.4064103364944458 + ], + "logprob_reward": -0.2814103364944458, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814103364944458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09464997053146362, + 0.04319026693701744, + 0.08424107730388641, + 0.12054260820150375, + 0.25522398948669434, + 0.047981683164834976, + 0.0870673805475235, + 0.15079134702682495, + 0.07905079424381256, + 0.03726084902882576 + ], + "credit_map": [ + 0.008575439453125, + 0.017703866586089134, + 0.012846916913986206, + 0.008872064761817455, + 0.015382104553282261, + 0.012921975925564766, + 0.009760962799191475, + 0.008586657233536243, + 0.004457660485059023, + 0.005459282547235489, + 0.004776250571012497, + 0.004586627706885338, + 0.007940692827105522, + 0.006325158756226301, + 0.004164178390055895, + 0.005480417050421238, + 0.018993649631738663, + 0.018161535263061523, + 0.014645843766629696, + 0.003884343896061182, + 0.01032288558781147, + 0.006861941423267126, + 0.0060034701600670815, + 0.00536741828545928, + 0.019672973081469536, + 0.010020379908382893, + 0.006648157257586718, + 0.013190225698053837, + 0.01805000752210617, + 0.012606369331479073, + 0.0329156368970871, + 0.007438864093273878, + 0.08226912468671799, + 0.017951123416423798, + 0.009163228794932365, + 0.026708953082561493, + 0.024989115074276924, + 0.06577498465776443, + 0.01811179332435131, + 0.010255688801407814, + 0.013594078831374645, + 0.0033480743877589703, + 0.0035006895195692778, + 0.004044673405587673, + 0.004138168413192034, + 0.007109193596988916, + 0.006857689004391432, + 0.0053891208954155445, + 0.013631636276841164, + 0.008053384721279144, + 0.01065513864159584, + 0.008289755322039127, + 0.011888900771737099, + 0.006420545745640993, + 0.013040155172348022, + 0.015087876468896866, + 0.030964182689785957, + 0.015905404463410378, + 0.016675665974617004, + 0.015348362736403942, + 0.018390154466032982, + 0.01957344077527523, + 0.01428613718599081, + 0.01964801736176014, + 0.005804093554615974, + 0.010446475818753242, + 0.014256375841796398, + 0.008229768835008144, + 0.013955289497971535, + 0.005788036622107029, + 0.009266827255487442, + 0.011303937062621117, + 0.002954842522740364, + 0.0033620696049183607, + 0.0033040859270840883, + 0.00439637154340744, + 0.0018749621231108904, + 0.010431463830173016, + 0.0029192299116402864, + 0.008017828688025475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.2814103364944458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981891855597496 + ], + "opposite_logprob_tokens": [ + -3.5298190116882324 + ], + "logprob_reward": -0.02981891855597496, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981891855597496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24527010321617126, + 0.11817395687103271, + 0.04348428547382355, + 0.0527370311319828, + 0.05044756829738617, + 0.04190398380160332, + 0.04890856519341469, + 0.04089346155524254, + 0.04450272023677826, + 0.3136783540248871 + ], + "credit_map": [ + 0.021525368094444275, + 0.036880698055028915, + 0.03611581400036812, + 0.017505154013633728, + 0.039039354771375656, + 0.033930789679288864, + 0.027850350365042686, + 0.03242256119847298, + 0.01096161175519228, + 0.016031187027692795, + 0.008344273082911968, + 0.011643432080745697, + 0.020189225673675537, + 0.014735706150531769, + 0.014446300454437733, + 0.02182222157716751, + 0.005976960528641939, + 0.004404675215482712, + 0.0037220388185232878, + 0.0036019315011799335, + 0.007531201466917992, + 0.005178875755518675, + 0.006753132678568363, + 0.006315469741821289, + 0.007257323712110519, + 0.005225706845521927, + 0.005629206076264381, + 0.006558847147971392, + 0.007863161154091358, + 0.007045569829642773, + 0.008329598233103752, + 0.0048276158049702644, + 0.008399725891649723, + 0.008560854010283947, + 0.003925000783056021, + 0.005026060622185469, + 0.008134457282721996, + 0.007775804493576288, + 0.004100903403013945, + 0.0045247613452374935, + 0.006426261737942696, + 0.00400175154209137, + 0.006107942666858435, + 0.005952034145593643, + 0.0036517190746963024, + 0.006829293444752693, + 0.005021261982619762, + 0.0039137182757258415, + 0.008068423718214035, + 0.0039025910664349794, + 0.0071697598323225975, + 0.00544720608741045, + 0.0053986841812729836, + 0.004588800016790628, + 0.007080338429659605, + 0.007252759765833616, + 0.00506621366366744, + 0.004322999157011509, + 0.004664690233767033, + 0.004383417312055826, + 0.0059506334364414215, + 0.004185948055237532, + 0.006369884591549635, + 0.005949674639850855, + 0.004833849612623453, + 0.0061580948531627655, + 0.0045508695766329765, + 0.0036684731021523476, + 0.008956472389400005, + 0.005009915214031935, + 0.005960394162684679, + 0.005364651326090097, + 0.016648534685373306, + 0.014733390882611275, + 0.022944746538996696, + 0.021829498931765556, + 0.01248359028249979, + 0.12574347853660583, + 0.02377568744122982, + 0.0755193904042244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.02981891855597496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062045931816101074 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.062045931816101074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062045931816101074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2281845360994339, + 0.05633097514510155, + 0.03328516334295273, + 0.05791476368904114, + 0.36534377932548523, + 0.08214209228754044, + 0.0566868931055069, + 0.013673308305442333, + 0.05997059866786003, + 0.046467896550893784 + ], + "credit_map": [ + 0.019793320447206497, + 0.03565136715769768, + 0.036368150264024734, + 0.016921797767281532, + 0.034989386796951294, + 0.03353763744235039, + 0.02510586753487587, + 0.02581699565052986, + 0.005490854382514954, + 0.008606256917119026, + 0.005111822858452797, + 0.00448022224009037, + 0.008814334869384766, + 0.007431419566273689, + 0.00730147073045373, + 0.00909459125250578, + 0.007148400414735079, + 0.007611962966620922, + 0.0036181563045829535, + 0.0017709317617118359, + 0.005052871536463499, + 0.002106876578181982, + 0.0032458454370498657, + 0.002730119042098522, + 0.00785959791392088, + 0.005878988653421402, + 0.00536717101931572, + 0.006481893826276064, + 0.00923620630055666, + 0.007286850828677416, + 0.011604825966060162, + 0.004199226852506399, + 0.10404399782419205, + 0.02789420261979103, + 0.014979875646531582, + 0.03827807679772377, + 0.04509170353412628, + 0.08395182341337204, + 0.03476184606552124, + 0.01634225621819496, + 0.013251770287752151, + 0.008013084530830383, + 0.006427999585866928, + 0.007234425749629736, + 0.010282070375978947, + 0.01133689284324646, + 0.015713172033429146, + 0.009882675483822823, + 0.007739737629890442, + 0.007650142535567284, + 0.006908513139933348, + 0.0062151323072612286, + 0.008195350877940655, + 0.004689620342105627, + 0.006453627720475197, + 0.008834765292704105, + 0.0017013134201988578, + 0.0014423979446291924, + 0.0019305500900372863, + 0.001561804790981114, + 0.0018866932950913906, + 0.0012406676542013884, + 0.0021325992420315742, + 0.001777280354872346, + 0.004274946637451649, + 0.009596388787031174, + 0.008701811544597149, + 0.0059914072044193745, + 0.0102086728438735, + 0.005900489632040262, + 0.0067042154259979725, + 0.008592662401497364, + 0.002546879230067134, + 0.003411276498809457, + 0.0029482622630894184, + 0.004216982051730156, + 0.0021019973792135715, + 0.01745706982910633, + 0.0037586044054478407, + 0.010026822797954082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.062045931816101074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15249936282634735, + 0.03688051179051399, + 0.035644806921482086, + 0.02622026391327381, + 0.412310391664505, + 0.08756288141012192, + 0.07998616248369217, + 0.015412257052958012, + 0.12689784169197083, + 0.02658558264374733 + ], + "credit_map": [ + 0.013480637222528458, + 0.02516257017850876, + 0.023134613409638405, + 0.010622844099998474, + 0.026232454925775528, + 0.02249150723218918, + 0.015382246114313602, + 0.015992484986782074, + 0.003633312415331602, + 0.005734983365982771, + 0.0035792645066976547, + 0.0035363142378628254, + 0.00602023396641016, + 0.004673656541854143, + 0.004446031991392374, + 0.0052567156963050365, + 0.009687797166407108, + 0.01108449138700962, + 0.0024482612498104572, + 0.0017483694246038795, + 0.0043520499020814896, + 0.0020754816941916943, + 0.00213247281499207, + 0.0021158824674785137, + 0.0029505270067602396, + 0.0029460012447088957, + 0.0028469993267208338, + 0.0034625609405338764, + 0.004833511542528868, + 0.004205186385661364, + 0.0032898569479584694, + 0.0016856200527399778, + 0.1354387402534485, + 0.027370771393179893, + 0.014706586487591267, + 0.04420087859034538, + 0.03598780184984207, + 0.10631430894136429, + 0.03256441280245781, + 0.015726877376437187, + 0.014550634659826756, + 0.009986733086407185, + 0.007272080518305302, + 0.008007241412997246, + 0.01261518057435751, + 0.012981961481273174, + 0.013039138168096542, + 0.009109911508858204, + 0.009828243404626846, + 0.010903652757406235, + 0.005992645863443613, + 0.006975777447223663, + 0.016415471211075783, + 0.006529966369271278, + 0.010881864465773106, + 0.012458540499210358, + 0.0021558820735663176, + 0.0014514380600303411, + 0.0014408668503165245, + 0.0019313381053507328, + 0.002041140804067254, + 0.0018419602420181036, + 0.0026529713068157434, + 0.0018966590287163854, + 0.008739656768739223, + 0.018011946231126785, + 0.013907616958022118, + 0.012094549834728241, + 0.027774911373853683, + 0.010801986791193485, + 0.01661921665072441, + 0.01894795522093773, + 0.002526798751205206, + 0.0037277955561876297, + 0.002894992008805275, + 0.004269142635166645, + 0.003151254029944539, + 0.0036335778422653675, + 0.002806918229907751, + 0.0035751056857407093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.009783650748431683 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This vibrant, anime-style, low-angle, dynamic tracking shot follows a firefighter as they relentlessly charge into the inferno of a collapsing building. Their hulking, soot-stained shifters, with sharp lines and deep shadows, are streaked with luminous yellow reflective stripes to contrast against the flickering lights. The camera quickly captures an extreme close-up of their eyes, framed by the soot-streaked visors of their helmets. Their large, expressive anime eyes are a bright, resolute azure blue, reflecting the roaring, stylized flames around them. With pupils constricted and brows furrowed, they survey the perilous surroundings with unwavering determination, their intense gaze piercing the thick smoke and glowing embers that fill the air. Flames roar dramatically in the shadows of the battery, tinged with orange and yellow, licking at charred support beams and causing the air to flicker with heat, creating a tense and heroic atmosphere as the firefighters battle the overwhelming blaze in their rescue mission.", + "seed": 6179, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6179_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": -0.12959360294044017, + "vlm_logprob": -0.12959360294044017, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.12959360294044017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18482334911823273 + ], + "opposite_logprob_tokens": [ + -1.8098233938217163 + ], + "logprob_reward": -0.18482334911823273, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18482334911823273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13965967297554016, + 0.2698290944099426, + 0.032336488366127014, + 0.018589314073324203, + 0.029726503416895866, + 0.01871151104569435, + 0.17445948719978333, + 0.07025036960840225, + 0.18658307194709778, + 0.05985450744628906 + ], + "credit_map": [ + 0.011686495505273342, + 0.017256926745176315, + 0.01725924015045166, + 0.011266554705798626, + 0.026194417849183083, + 0.011440007016062737, + 0.030450057238340378, + 0.014105966314673424, + 0.0654781386256218, + 0.01663147285580635, + 0.012323017232120037, + 0.011784626170992851, + 0.01683533750474453, + 0.018487105146050453, + 0.02942919358611107, + 0.09886021167039871, + 0.004425763618201017, + 0.0034541161730885506, + 0.0065043410286307335, + 0.003001989098265767, + 0.004192765336483717, + 0.004450313746929169, + 0.00331962201744318, + 0.002987576648592949, + 0.002203066134825349, + 0.0019030246185138822, + 0.002824479481205344, + 0.0015913780080154538, + 0.002892487682402134, + 0.0031716732773929834, + 0.0020612096413969994, + 0.001941994996741414, + 0.003190679242834449, + 0.003185350215062499, + 0.004173853900283575, + 0.0032747036311775446, + 0.004351618234068155, + 0.00393645279109478, + 0.004236436914652586, + 0.0033774077892303467, + 0.002370043657720089, + 0.002988682594150305, + 0.0021133478730916977, + 0.0018582313787192106, + 0.0023956268560141325, + 0.0024906257167458534, + 0.0028233665507286787, + 0.0016715861856937408, + 0.041628796607255936, + 0.014305130578577518, + 0.060555752366781235, + 0.010622906498610973, + 0.008677809499204159, + 0.016482142731547356, + 0.010364961810410023, + 0.011821997351944447, + 0.010109640657901764, + 0.007331025321036577, + 0.011975541710853577, + 0.00734357163310051, + 0.005682130344212055, + 0.006971582770347595, + 0.01155317947268486, + 0.0092837018892169, + 0.02806730754673481, + 0.015184317715466022, + 0.05542776361107826, + 0.015445669181644917, + 0.017535721883177757, + 0.014851662330329418, + 0.018385669216513634, + 0.021684974431991577, + 0.005448981188237667, + 0.005471655633300543, + 0.006780277471989393, + 0.00638254638761282, + 0.011158779263496399, + 0.008340386673808098, + 0.00909663550555706, + 0.007175250444561243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 6 + ], + "selected_reward": -0.18482334911823273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135256767272949 + ], + "opposite_logprob_tokens": [ + -1.313525676727295 + ], + "logprob_reward": -0.3135256767272949, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135256767272949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07296571880578995, + 0.166622593998909, + 0.031689468771219254, + 0.03480106592178345, + 0.05423905327916145, + 0.12103461474180222, + 0.2063627690076828, + 0.08466225117444992, + 0.16676652431488037, + 0.060855932533741 + ], + "credit_map": [ + 0.006939942017197609, + 0.010624959133565426, + 0.008966227062046528, + 0.0066299051977694035, + 0.012190727517008781, + 0.0077139586210250854, + 0.01307560782879591, + 0.0068243942223489285, + 0.039078038185834885, + 0.010310344398021698, + 0.011601878330111504, + 0.007596408948302269, + 0.010607985779643059, + 0.012274122796952724, + 0.01764889806509018, + 0.057504910975694656, + 0.002796170534566045, + 0.0036493674851953983, + 0.003578411415219307, + 0.003733632154762745, + 0.00556905148550868, + 0.002933627925813198, + 0.005618758499622345, + 0.003810449969023466, + 0.0029471879824995995, + 0.003691084450110793, + 0.0032871421426534653, + 0.0031442600302398205, + 0.008579148910939693, + 0.003489177441224456, + 0.005245317239314318, + 0.004417745396494865, + 0.006317928433418274, + 0.007248198613524437, + 0.007528146728873253, + 0.00553763285279274, + 0.008526875637471676, + 0.006792910862714052, + 0.006696819327771664, + 0.005590539891272783, + 0.014207132160663605, + 0.021558424457907677, + 0.02015218697488308, + 0.01428674254566431, + 0.010778074152767658, + 0.018746666610240936, + 0.011432001367211342, + 0.009873390197753906, + 0.04621344059705734, + 0.017604760825634003, + 0.03768640011548996, + 0.011920305900275707, + 0.0279059000313282, + 0.02371770329773426, + 0.01584923081099987, + 0.02546503022313118, + 0.007849130779504776, + 0.013418926857411861, + 0.00916985608637333, + 0.010761785320937634, + 0.01250613946467638, + 0.008876240812242031, + 0.011539089493453503, + 0.01054108701646328, + 0.019168339669704437, + 0.020221969112753868, + 0.020045429468154907, + 0.017374815419316292, + 0.028292225673794746, + 0.017269548028707504, + 0.025246474891901016, + 0.019147710874676704, + 0.007022432517260313, + 0.007849224843084812, + 0.007424030918627977, + 0.0061484226025640965, + 0.010768230073153973, + 0.006415391340851784, + 0.00804011058062315, + 0.007188092451542616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.3135256767272949 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295673251152039 + ], + "opposite_logprob_tokens": [ + -2.2379567623138428 + ], + "logprob_reward": -0.11295673251152039, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295673251152039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06698570400476456, + 0.23923514783382416, + 0.09282186627388, + 0.1224798709154129, + 0.02520768716931343, + 0.1085708737373352, + 0.19965636730194092, + 0.059183839708566666, + 0.04543064534664154, + 0.040427982807159424 + ], + "credit_map": [ + 0.008697934448719025, + 0.008132836781442165, + 0.013645394705235958, + 0.005538315046578646, + 0.008989482186734676, + 0.0059023695066571236, + 0.010036878287792206, + 0.006042496766895056, + 0.05605273321270943, + 0.01596386916935444, + 0.014449489302933216, + 0.012321780435740948, + 0.013412848114967346, + 0.017236042767763138, + 0.02608349733054638, + 0.08371490240097046, + 0.012716904282569885, + 0.014959625899791718, + 0.009203420020639896, + 0.010522625409066677, + 0.016265759244561195, + 0.007178998086601496, + 0.013503758236765862, + 0.008470771834254265, + 0.014339029788970947, + 0.014292037114501, + 0.01849224418401718, + 0.011956791393458843, + 0.018061626702547073, + 0.018577121198177338, + 0.014470444060862064, + 0.012290576472878456, + 0.004589761141687632, + 0.0021769849117845297, + 0.003091702936217189, + 0.001441026572138071, + 0.0034697018563747406, + 0.0037625236436724663, + 0.0019591329619288445, + 0.004716851282864809, + 0.013105949386954308, + 0.016655253246426582, + 0.017778124660253525, + 0.013455923646688461, + 0.01049830112606287, + 0.015637528151273727, + 0.011660910211503506, + 0.009778880514204502, + 0.04877594858407974, + 0.01560894027352333, + 0.06122065335512161, + 0.011547509580850601, + 0.01577177830040455, + 0.019240545108914375, + 0.011846103705465794, + 0.01564488373696804, + 0.009103294461965561, + 0.0056186881847679615, + 0.010854740627110004, + 0.006701729726046324, + 0.006777802482247353, + 0.006684235762804747, + 0.0062978435307741165, + 0.007145504932850599, + 0.0058418395929038525, + 0.004740739706903696, + 0.009756787680089474, + 0.0035459266509860754, + 0.008152000606060028, + 0.004397447686642408, + 0.0049241213127970695, + 0.004071783274412155, + 0.004647849593311548, + 0.0036759835202246904, + 0.007961624301970005, + 0.004103762097656727, + 0.00477193295955658, + 0.0063056498765945435, + 0.004900282248854637, + 0.004060899373143911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.11295673251152039 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589197054505348 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589197054505348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589197054505348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0720023363828659, + 0.30347248911857605, + 0.17484541237354279, + 0.04656064510345459, + 0.0602763332426548, + 0.04067833721637726, + 0.07808377593755722, + 0.02364864945411682, + 0.10200084745883942, + 0.09843118488788605 + ], + "credit_map": [ + 0.006111668888479471, + 0.010132214985787868, + 0.007942409254610538, + 0.0050149173475801945, + 0.016915790736675262, + 0.006736093200743198, + 0.011338453739881516, + 0.007810790557414293, + 0.07202998548746109, + 0.021052634343504906, + 0.01552039198577404, + 0.013711225241422653, + 0.019456107169389725, + 0.0209598820656538, + 0.03310982510447502, + 0.10763242840766907, + 0.020289450883865356, + 0.028026731684803963, + 0.012660092674195766, + 0.02298438362777233, + 0.03408315032720566, + 0.01173650473356247, + 0.027193017303943634, + 0.01787208952009678, + 0.0045361691154539585, + 0.004662892781198025, + 0.006251601502299309, + 0.0036448596511036158, + 0.010067165829241276, + 0.005118766333907843, + 0.006905772723257542, + 0.0053734187968075275, + 0.007291427347809076, + 0.006754950154572725, + 0.008144328370690346, + 0.0052887252531945705, + 0.011341474950313568, + 0.0061948285438120365, + 0.007093869149684906, + 0.00816673506051302, + 0.003977338317781687, + 0.005335941445082426, + 0.006379822734743357, + 0.004760055337101221, + 0.005978647153824568, + 0.004659450147300959, + 0.005507748574018478, + 0.004079331643879414, + 0.013529900461435318, + 0.007097688503563404, + 0.029191600158810616, + 0.004694783594459295, + 0.004990894813090563, + 0.005181674379855394, + 0.005002468358725309, + 0.008394768461585045, + 0.0026579750701785088, + 0.002345506101846695, + 0.002843126654624939, + 0.002492315601557493, + 0.0037859356962144375, + 0.0026028184220194817, + 0.0040176669135689735, + 0.0029033059254288673, + 0.016271298751235008, + 0.007214166224002838, + 0.0300285741686821, + 0.007182809058576822, + 0.009695393033325672, + 0.009996951557695866, + 0.010397295467555523, + 0.011214354075491428, + 0.010362498462200165, + 0.009253503754734993, + 0.014920203946530819, + 0.009439115412533283, + 0.01885596103966236, + 0.011078344658017159, + 0.01355844084173441, + 0.010963118635118008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 8 + ], + "selected_reward": -0.020589197054505348 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073059290647507 + ], + "opposite_logprob_tokens": [ + -4.141073226928711 + ], + "logprob_reward": -0.016073059290647507, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073059290647507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09014768153429031, + 0.3436298966407776, + 0.05538666248321533, + 0.14031894505023956, + 0.027415823191404343, + 0.07536271214485168, + 0.05199853703379631, + 0.06683366000652313, + 0.07620292901992798, + 0.07270316034555435 + ], + "credit_map": [ + 0.007675234228372574, + 0.01374361664056778, + 0.00871992576867342, + 0.006942165549844503, + 0.022327134385704994, + 0.008133942261338234, + 0.012741697952151299, + 0.0098639614880085, + 0.06979066878557205, + 0.02944335713982582, + 0.018115820363163948, + 0.018211232498288155, + 0.04157371073961258, + 0.022704020142555237, + 0.03837248682975769, + 0.10541858524084091, + 0.008497488684952259, + 0.004495619796216488, + 0.008526427671313286, + 0.005797514691948891, + 0.009816410019993782, + 0.005025325808674097, + 0.008277488872408867, + 0.004950383212417364, + 0.01570260152220726, + 0.016059264540672302, + 0.017108123749494553, + 0.013650688342750072, + 0.016915880143642426, + 0.025885529816150665, + 0.017047280445694923, + 0.017949575558304787, + 0.003062807023525238, + 0.0030852609779685736, + 0.0034606927074491978, + 0.0029804229270666838, + 0.0038265374023467302, + 0.003657110733911395, + 0.004075252451002598, + 0.0032677401322871447, + 0.009993363171815872, + 0.007283853366971016, + 0.012942908331751823, + 0.008320972323417664, + 0.007295809220522642, + 0.014368411153554916, + 0.008305025286972523, + 0.006852364633232355, + 0.00709548220038414, + 0.004289875738322735, + 0.015224226750433445, + 0.003644751850515604, + 0.004268862307071686, + 0.0038712150417268276, + 0.00649644760414958, + 0.007107672747224569, + 0.0063482774421572685, + 0.010347597301006317, + 0.007551060523837805, + 0.006221975199878216, + 0.01253756694495678, + 0.006578348111361265, + 0.009193620644509792, + 0.00805521197617054, + 0.008149364963173866, + 0.007164576556533575, + 0.01266250666230917, + 0.005623609758913517, + 0.014204765670001507, + 0.00622553238645196, + 0.010559918358922005, + 0.01161265466362238, + 0.00811729859560728, + 0.007428107317537069, + 0.008743751794099808, + 0.005995604209601879, + 0.016978617757558823, + 0.006129967048764229, + 0.011811831966042519, + 0.007497977465391159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.016073059290647507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This vibrant, anime-style, low-angle, dynamic tracking shot follows a firefighter as they relentlessly charge into the inferno of a collapsing building. Their hulking, soot-stained shifters, with sharp lines and deep shadows, are streaked with luminous yellow reflective stripes to contrast against the flickering lights. The camera quickly captures an extreme close-up of their eyes, framed by the soot-streaked visors of their helmets. Their large, expressive anime eyes are a bright, resolute azure blue, reflecting the roaring, stylized flames around them. With pupils constricted and brows furrowed, they survey the perilous surroundings with unwavering determination, their intense gaze piercing the thick smoke and glowing embers that fill the air. Flames roar dramatically in the shadows of the battery, tinged with orange and yellow, licking at charred support beams and causing the air to flicker with heat, creating a tense and heroic atmosphere as the firefighters battle the overwhelming blaze in their rescue mission.", + "seed": 6180, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_13/index_6570_seed_6180_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": -0.0998739649541676, + "vlm_logprob": -0.0998739649541676, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.0998739649541676, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03343089297413826 + ], + "opposite_logprob_tokens": [ + -3.533430814743042 + ], + "logprob_reward": -0.03343089297413826, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03343089297413826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12748344242572784, + 0.12046955525875092, + 0.09059149026870728, + 0.1696912795305252, + 0.05004288628697395, + 0.05000588670372963, + 0.10756584256887436, + 0.18735024333000183, + 0.07866555452346802, + 0.01813381537795067 + ], + "credit_map": [ + 0.014649692922830582, + 0.022850394248962402, + 0.019925473257899284, + 0.017001166939735413, + 0.01969744637608528, + 0.009677537716925144, + 0.013203349895775318, + 0.010478388518095016, + 0.01867273636162281, + 0.010812990367412567, + 0.016021957620978355, + 0.01150039304047823, + 0.012271643616259098, + 0.01275939866900444, + 0.021369408816099167, + 0.01706102304160595, + 0.007415656466037035, + 0.008715376257896423, + 0.012037968263030052, + 0.006664137355983257, + 0.013478401117026806, + 0.005230975802987814, + 0.0209686066955328, + 0.016080370172858238, + 0.0251358300447464, + 0.016546381637454033, + 0.02333608828485012, + 0.009154538623988628, + 0.02513260953128338, + 0.015567519702017307, + 0.0288950577378273, + 0.025923265144228935, + 0.007199567276984453, + 0.004440390504896641, + 0.004405203741043806, + 0.004008918069303036, + 0.007077667396515608, + 0.005530205555260181, + 0.009342015720903873, + 0.008038914762437344, + 0.0054606287740170956, + 0.005164420232176781, + 0.004410343710333109, + 0.005343708675354719, + 0.008345439098775387, + 0.005493347533047199, + 0.008670569397509098, + 0.007117428816854954, + 0.00964352022856474, + 0.007519604172557592, + 0.008756664581596851, + 0.00913284532725811, + 0.010628314688801765, + 0.01069201435893774, + 0.029749048873782158, + 0.021443834528326988, + 0.014917139895260334, + 0.025170324370265007, + 0.01482312474399805, + 0.012745676562190056, + 0.03663624823093414, + 0.024592697620391846, + 0.03271014988422394, + 0.025754889473319054, + 0.0073840077966451645, + 0.008777324110269547, + 0.008784092962741852, + 0.009055865928530693, + 0.011051456443965435, + 0.008511082269251347, + 0.012842553667724133, + 0.012259168550372124, + 0.002515611005946994, + 0.0017999591073021293, + 0.0017532712081447244, + 0.001556196715682745, + 0.002439935691654682, + 0.0019250208279117942, + 0.0035159701947122812, + 0.00262785074301064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.03343089297413826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814108729362488 + ], + "opposite_logprob_tokens": [ + -1.4064109325408936 + ], + "logprob_reward": -0.2814108729362488, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814108729362488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2944280505180359, + 0.07989633083343506, + 0.05007438734173775, + 0.09255766123533249, + 0.07197634130716324, + 0.06723295152187347, + 0.09376280009746552, + 0.07447077333927155, + 0.1483464539051056, + 0.027254221960902214 + ], + "credit_map": [ + 0.013407367281615734, + 0.08440939337015152, + 0.015741730108857155, + 0.0177468154579401, + 0.10043171048164368, + 0.0276417825371027, + 0.021318554878234863, + 0.01373068057000637, + 0.013722146861255169, + 0.006332515738904476, + 0.009588993154466152, + 0.0075342971831560135, + 0.008173718117177486, + 0.006456849630922079, + 0.014884992502629757, + 0.013202819973230362, + 0.005182518158107996, + 0.004482609685510397, + 0.007965566590428352, + 0.00585333164781332, + 0.006284819915890694, + 0.004012314137071371, + 0.008866982534527779, + 0.007426244206726551, + 0.012471484951674938, + 0.01109364815056324, + 0.01104055903851986, + 0.006366492249071598, + 0.013143040239810944, + 0.00854475237429142, + 0.016325855627655983, + 0.01357182301580906, + 0.008661814965307713, + 0.006247416138648987, + 0.0059685492888092995, + 0.004438663832843304, + 0.011319858022034168, + 0.009288754314184189, + 0.014604036696255207, + 0.01144725177437067, + 0.009473474696278572, + 0.006237166002392769, + 0.006431466434150934, + 0.007323527708649635, + 0.010478321462869644, + 0.007610118482261896, + 0.010484830476343632, + 0.009194047190248966, + 0.016068801283836365, + 0.005650790873914957, + 0.010732431896030903, + 0.007319232914596796, + 0.008972053416073322, + 0.014114178717136383, + 0.0160311758518219, + 0.01487413328140974, + 0.01193263754248619, + 0.007125229574739933, + 0.009991100989282131, + 0.006571060977876186, + 0.008781441487371922, + 0.007479714695364237, + 0.012072429992258549, + 0.010517158545553684, + 0.01886509172618389, + 0.009516692720353603, + 0.01741827093064785, + 0.01684899628162384, + 0.013703852891921997, + 0.01626715436577797, + 0.02953796274960041, + 0.026188433170318604, + 0.003119048196822405, + 0.0027131508104503155, + 0.0021313182078301907, + 0.0029202771838754416, + 0.003981230314821005, + 0.003373251762241125, + 0.0050084637477993965, + 0.004007481504231691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.2814108729362488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038252413272858 + ], + "opposite_logprob_tokens": [ + -1.9103825092315674 + ], + "logprob_reward": -0.16038252413272858, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038252413272858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5017716288566589, + 0.06296797096729279, + 0.055200375616550446, + 0.050322484225034714, + 0.03848668560385704, + 0.04879562929272652, + 0.03236984461545944, + 0.05314742028713226, + 0.03576242923736572, + 0.12117550522089005 + ], + "credit_map": [ + 0.022746723145246506, + 0.15716159343719482, + 0.023387616500258446, + 0.024108292534947395, + 0.17299741506576538, + 0.04459014907479286, + 0.036350067704916, + 0.020429791882634163, + 0.012493680231273174, + 0.0058868215419352055, + 0.005412611179053783, + 0.004279072396457195, + 0.00798036903142929, + 0.005279890727251768, + 0.012524341233074665, + 0.009111183695495129, + 0.006810045801103115, + 0.006830898113548756, + 0.007379316724836826, + 0.0040278867818415165, + 0.007026321720331907, + 0.005273798014968634, + 0.009286544285714626, + 0.008565567433834076, + 0.007839541882276535, + 0.005385200027376413, + 0.005748173221945763, + 0.0044708289206027985, + 0.00570831261575222, + 0.00562007911503315, + 0.008344447240233421, + 0.007205904461443424, + 0.004498652648180723, + 0.004378787707537413, + 0.002997009316459298, + 0.0030476844403892756, + 0.006634231191128492, + 0.004322054795920849, + 0.007260523736476898, + 0.0053477417677640915, + 0.008438415825366974, + 0.004986317828297615, + 0.005262619815766811, + 0.004829552490264177, + 0.009678300470113754, + 0.0050475578755140305, + 0.005387518554925919, + 0.005165351089090109, + 0.004623788874596357, + 0.002225390402600169, + 0.003339494112879038, + 0.0027927227783948183, + 0.003791872411966324, + 0.0027996650896966457, + 0.007313331589102745, + 0.005483580287545919, + 0.007055394817143679, + 0.0049508484080433846, + 0.006016861647367477, + 0.003156117396429181, + 0.008995965123176575, + 0.004239133093506098, + 0.010909251868724823, + 0.00782384630292654, + 0.004659460857510567, + 0.002719811163842678, + 0.0042928908951580524, + 0.004272748250514269, + 0.0038055002223700285, + 0.003912896383553743, + 0.006416024174541235, + 0.005683098454028368, + 0.008895359933376312, + 0.011704913340508938, + 0.008059486746788025, + 0.008580481633543968, + 0.020625092089176178, + 0.014418202452361584, + 0.028297048062086105, + 0.020594922825694084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.16038252413272858 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12469259649515152, + 0.12006814777851105, + 0.18651628494262695, + 0.0791785940527916, + 0.08982814848423004, + 0.039718203246593475, + 0.09935148805379868, + 0.16414420306682587, + 0.04497622326016426, + 0.05152614787220955 + ], + "credit_map": [ + 0.01687769591808319, + 0.013465260155498981, + 0.01900838315486908, + 0.015917744487524033, + 0.015269357711076736, + 0.012136640027165413, + 0.019569532945752144, + 0.012447992339730263, + 0.0151149220764637, + 0.011628837324678898, + 0.010375323705375195, + 0.015299227088689804, + 0.01534674596041441, + 0.008926305919885635, + 0.025273695588111877, + 0.018103094771504402, + 0.017587119713425636, + 0.02370680309832096, + 0.02696928381919861, + 0.013725888915359974, + 0.03205854818224907, + 0.01153537817299366, + 0.03177197277545929, + 0.02916128747165203, + 0.010969147086143494, + 0.009328092448413372, + 0.010140728205442429, + 0.005140295252203941, + 0.0097030745819211, + 0.0072016469202935696, + 0.014920562505722046, + 0.011775046586990356, + 0.004786814097315073, + 0.016957538202404976, + 0.005047623533755541, + 0.006390019785612822, + 0.018421156331896782, + 0.006722371093928814, + 0.018353819847106934, + 0.013148799538612366, + 0.005984304007142782, + 0.0033034044317901134, + 0.0037785593885928392, + 0.003351768711581826, + 0.008630204945802689, + 0.00499696284532547, + 0.004979220684617758, + 0.004693776369094849, + 0.009281405247747898, + 0.006589014083147049, + 0.010617260821163654, + 0.00802038237452507, + 0.008808799088001251, + 0.01628252863883972, + 0.022503562271595, + 0.01724853925406933, + 0.009172498248517513, + 0.023976773023605347, + 0.00953365582972765, + 0.012343677692115307, + 0.03203315660357475, + 0.016989901661872864, + 0.03440229222178459, + 0.025692244991660118, + 0.0037319553084671497, + 0.005857387091964483, + 0.003893177257850766, + 0.0049371724016964436, + 0.007403090596199036, + 0.0041927616111934185, + 0.00842607393860817, + 0.0065346043556928635, + 0.003527540247887373, + 0.00544696394354105, + 0.0031346548348665237, + 0.003905725199729204, + 0.009958392009139061, + 0.006396066397428513, + 0.010622075758874416, + 0.008534728549420834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style anime-style with sharp lines and deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the firefighter wearing a soot-stained shift with luminous yellow reflective stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there an extreme close-up of the firefighter's eyes showing bright azure blue irises?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eyes framed by soot-streaked visors of a helmet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059529622085392475 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.0059529622085392475, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059529622085392475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31657537817955017, + 0.11723695695400238, + 0.07192446291446686, + 0.11386626958847046, + 0.06495197862386703, + 0.09735386073589325, + 0.06067943945527077, + 0.07471685856580734, + 0.05431012436747551, + 0.028384631499648094 + ], + "credit_map": [ + 0.01850471831858158, + 0.09376006573438644, + 0.023030925542116165, + 0.019596273079514503, + 0.09874127805233002, + 0.025932589545845985, + 0.022100504487752914, + 0.014909065328538418, + 0.011698766611516476, + 0.010889293625950813, + 0.008879609405994415, + 0.014442592859268188, + 0.015073689632117748, + 0.009922855533659458, + 0.028687823563814163, + 0.017642337828874588, + 0.008101734332740307, + 0.008675889112055302, + 0.010146010667085648, + 0.006403298117220402, + 0.011285723187029362, + 0.005604559555649757, + 0.010117498226463795, + 0.01158976275473833, + 0.010504551231861115, + 0.015259018167853355, + 0.009182705543935299, + 0.008696779608726501, + 0.016918223351240158, + 0.010676134377717972, + 0.025340402498841286, + 0.017288465052843094, + 0.007697733584791422, + 0.00636853976175189, + 0.005517459008842707, + 0.005929708480834961, + 0.008559801615774632, + 0.005313715431839228, + 0.014745453372597694, + 0.010819573886692524, + 0.01458407286554575, + 0.010058321058750153, + 0.009675375185906887, + 0.006806300021708012, + 0.018811123445630074, + 0.011475805193185806, + 0.014077275060117245, + 0.011865596286952496, + 0.0036603668704628944, + 0.0036885214503854513, + 0.0036909054033458233, + 0.004656759090721607, + 0.00877037551254034, + 0.009008292108774185, + 0.015842735767364502, + 0.01136149000376463, + 0.005313631612807512, + 0.00860677007585764, + 0.004697867203503847, + 0.004935824312269688, + 0.01381160318851471, + 0.008816622197628021, + 0.015537525527179241, + 0.012997030280530453, + 0.00538442749530077, + 0.005236022174358368, + 0.004936451558023691, + 0.0060528055764734745, + 0.007447334472090006, + 0.005946418270468712, + 0.010411487892270088, + 0.008895176462829113, + 0.002107288222759962, + 0.003431686433032155, + 0.0015867032343521714, + 0.002775701927021146, + 0.006113460287451744, + 0.0035429210402071476, + 0.004663272760808468, + 0.004163599573075771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0059529622085392475 + } + ], + "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/0000055/group_14/index_tensor([9513, 9513])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_tensor([9513, 9513])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e5d6bbe53e30b99c7312241e52adf1b44398c07 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_tensor([9513, 9513])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, wide-angle shot captures a scene rendered in Salvador Dalí's stark black-and-white surrealist style. The setting is a barren, dreamlike seafloor, where the sand is a fine, pale powder and the water is perfectly, unsettlingly still, like a vast expanse of solid glass. A massive, iconic, melting clock, its Roman numerals dripping like liquid tar, lies half-buried in the silent landscape. Dominating the frame is a massive crab, its body a speckled gray, armored carapace, standing on incredibly spindly legs that seem to defy physics. Intense, dramatic lighting emanating from a single, unseen source sculpts the scene into extreme contrasts of brilliant white and deep, velvety black, casting long, grotesquely distorted shadows from the crab's legs across the seafloor, making them appear even more elongated and menacing. In a slow, deliberate, and relentless display of menace, the crab raises its two massive, obsidian-black claws, mechanically rising with terrifying precision, its pincers opening and closing with measured, slow pacing, creating a palpable atmosphere of nightmarish tension.", + "seed": 7713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_9513_seed_7713_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": -0.060718349646776915, + "vlm_logprob": -0.060718349646776915, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.060718349646776915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015388447791337967 + ], + "opposite_logprob_tokens": [ + -5.015388488769531 + ], + "logprob_reward": -0.015388447791337967, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.015388447791337967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21428731083869934, + 0.16987498104572296, + 0.0709538534283638, + 0.057886719703674316, + 0.10210661590099335, + 0.05766203999519348, + 0.12446677684783936, + 0.11050494015216827, + 0.02274301089346409, + 0.0695137307047844 + ], + "credit_map": [ + 0.03451729193329811, + 0.027478115633130074, + 0.02313501015305519, + 0.02144191414117813, + 0.02864105813205242, + 0.024555159732699394, + 0.023617293685674667, + 0.03090147115290165, + 0.025599198415875435, + 0.023495279252529144, + 0.016094660386443138, + 0.012426798231899738, + 0.02022080309689045, + 0.03080962784588337, + 0.020142924040555954, + 0.021085688844323158, + 0.009911536239087582, + 0.00905842985957861, + 0.006046362221240997, + 0.005996345542371273, + 0.010063917376101017, + 0.010739337652921677, + 0.009278129786252975, + 0.009859797544777393, + 0.007863995619118214, + 0.00712354714050889, + 0.006466479040682316, + 0.004864587914198637, + 0.01004001684486866, + 0.00953797996044159, + 0.0056235953234136105, + 0.006366516463458538, + 0.011002346873283386, + 0.010993562638759613, + 0.009508609771728516, + 0.007201636675745249, + 0.012764183804392815, + 0.012035807594656944, + 0.01328076422214508, + 0.025319702923297882, + 0.007190873846411705, + 0.005815998651087284, + 0.005627247504889965, + 0.004894071724265814, + 0.006999155506491661, + 0.007128481287509203, + 0.006666314788162708, + 0.01333990041166544, + 0.01725640520453453, + 0.020325426012277603, + 0.012679859064519405, + 0.008962856605648994, + 0.013752404600381851, + 0.018544267863035202, + 0.01508548203855753, + 0.017860082909464836, + 0.017848974093794823, + 0.010560421273112297, + 0.0141572505235672, + 0.008938534185290337, + 0.018891530111432076, + 0.015309484675526619, + 0.007135650608688593, + 0.01766309142112732, + 0.0030646370723843575, + 0.00281851040199399, + 0.003140193410217762, + 0.0022815517149865627, + 0.003022710792720318, + 0.002073979005217552, + 0.0033671678975224495, + 0.002974261762574315, + 0.008874082006514072, + 0.006750860717147589, + 0.008996457792818546, + 0.007223712746053934, + 0.009007428772747517, + 0.007937939837574959, + 0.00973732303828001, + 0.01098592672497034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.015388447791337967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991703748703003 + ], + "opposite_logprob_tokens": [ + -1.8049170970916748 + ], + "logprob_reward": -0.17991703748703003, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991703748703003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18558765947818756, + 0.39765340089797974, + 0.05766734853386879, + 0.028680602088570595, + 0.09456861019134521, + 0.07109711319208145, + 0.07599418610334396, + 0.04200112447142601, + 0.021631238982081413, + 0.025118695572018623 + ], + "credit_map": [ + 0.027705827727913857, + 0.032756257802248, + 0.020015323534607887, + 0.014480525627732277, + 0.02848195843398571, + 0.022389916703104973, + 0.017967907711863518, + 0.021789954975247383, + 0.039234425872564316, + 0.04571171849966049, + 0.04120386391878128, + 0.0433732308447361, + 0.06263326853513718, + 0.06919477880001068, + 0.03843473270535469, + 0.05786741152405739, + 0.009369635954499245, + 0.007198714651167393, + 0.007798497099429369, + 0.0048497566021978855, + 0.007106286007910967, + 0.007795434445142746, + 0.005386727396398783, + 0.008162299171090126, + 0.004451381508260965, + 0.004594183061271906, + 0.0029963282868266106, + 0.002361744176596403, + 0.004495801869779825, + 0.0035199993290007114, + 0.0028975685127079487, + 0.0033635960426181555, + 0.010688913986086845, + 0.010196252726018429, + 0.008476493880152702, + 0.006657093297690153, + 0.018955454230308533, + 0.010149873793125153, + 0.011295941658318043, + 0.01814858615398407, + 0.012677092105150223, + 0.0115652559325099, + 0.008047953248023987, + 0.005582389421761036, + 0.009526860900223255, + 0.008788605220615864, + 0.0045964280143380165, + 0.010312524624168873, + 0.0095356535166502, + 0.014861779287457466, + 0.006462248042225838, + 0.006766774691641331, + 0.014296644367277622, + 0.007249647285789251, + 0.0050577460788190365, + 0.011763691902160645, + 0.005832396913319826, + 0.005849859677255154, + 0.004672463983297348, + 0.0036441851407289505, + 0.005256312899291515, + 0.005341390147805214, + 0.004126982297748327, + 0.007277532014995813, + 0.0034089814871549606, + 0.0022073425352573395, + 0.00302797369658947, + 0.0017188016790896654, + 0.002632144605740905, + 0.00213476805947721, + 0.003721600864082575, + 0.002779626753181219, + 0.0032781152985990047, + 0.0031092469580471516, + 0.0028561095241457224, + 0.0028282329440116882, + 0.0032669808715581894, + 0.00282466784119606, + 0.0032104041893035173, + 0.0037449372466653585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.17991703748703003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305129498243332 + ], + "opposite_logprob_tokens": [ + -3.168051242828369 + ], + "logprob_reward": -0.04305129498243332, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305129498243332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15834198892116547, + 0.3222155272960663, + 0.09018982201814651, + 0.023899761959910393, + 0.03545956313610077, + 0.03312629088759422, + 0.13327814638614655, + 0.11560934036970139, + 0.047729671001434326, + 0.04014987498521805 + ], + "credit_map": [ + 0.024172954261302948, + 0.022466016933321953, + 0.019691726192831993, + 0.014513115398585796, + 0.019551951438188553, + 0.02087588422000408, + 0.015737738460302353, + 0.02133261412382126, + 0.03642865642905235, + 0.03795346990227699, + 0.035002853721380234, + 0.03289486840367317, + 0.045175835490226746, + 0.05750894919037819, + 0.032788682729005814, + 0.044462233781814575, + 0.012513876892626286, + 0.016434771940112114, + 0.01077270694077015, + 0.007370040286332369, + 0.013184444047510624, + 0.010018840432167053, + 0.01026611216366291, + 0.009629033505916595, + 0.0028492878191173077, + 0.0037123362999409437, + 0.0023348708637058735, + 0.0019191819010302424, + 0.003467105096206069, + 0.0037768457550555468, + 0.00256618345156312, + 0.003273950656875968, + 0.005071740597486496, + 0.004518793895840645, + 0.003937660250812769, + 0.0032651652581989765, + 0.003461588639765978, + 0.004454322159290314, + 0.003961764741688967, + 0.006788528990000486, + 0.003299111733213067, + 0.0037143491208553314, + 0.002951551927253604, + 0.003871041350066662, + 0.005324598401784897, + 0.003579871729016304, + 0.0051171076484024525, + 0.005268660839647055, + 0.014586717821657658, + 0.02054673433303833, + 0.01500293891876936, + 0.01064439956098795, + 0.02280934900045395, + 0.015698015689849854, + 0.010876081883907318, + 0.023113913834095, + 0.01972098834812641, + 0.013709759339690208, + 0.01543019525706768, + 0.009350959211587906, + 0.015852343291044235, + 0.01479056105017662, + 0.0072103035636246204, + 0.01954422891139984, + 0.005847738590091467, + 0.005030578933656216, + 0.005677203647792339, + 0.004695272073149681, + 0.005794186610728502, + 0.00629377132281661, + 0.005667109973728657, + 0.008723813109099865, + 0.003734457539394498, + 0.005540098994970322, + 0.004681166727095842, + 0.00396276218816638, + 0.0064230794087052345, + 0.004463484976440668, + 0.0048509924672544, + 0.006493831984698772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04305129498243332 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037527561187744 + ], + "opposite_logprob_tokens": [ + -2.812037467956543 + ], + "logprob_reward": -0.062037527561187744, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037527561187744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11050525307655334, + 0.19421638548374176, + 0.12547436356544495, + 0.06446421146392822, + 0.049113597720861435, + 0.03668199107050896, + 0.034964777529239655, + 0.10294320434331894, + 0.24437913298606873, + 0.037257060408592224 + ], + "credit_map": [ + 0.015961019322276115, + 0.022244449704885483, + 0.010137873701751232, + 0.009880781173706055, + 0.018324317410588264, + 0.013629133813083172, + 0.01018852461129427, + 0.01013916078954935, + 0.021746963262557983, + 0.020746123045682907, + 0.022016702219843864, + 0.018907690420746803, + 0.027114152908325195, + 0.03696582093834877, + 0.019931718707084656, + 0.026787221431732178, + 0.015586559660732746, + 0.013625073246657848, + 0.014214626513421535, + 0.012377978302538395, + 0.014649488031864166, + 0.018333716318011284, + 0.018174223601818085, + 0.018512697890400887, + 0.0067431689240038395, + 0.007177097722887993, + 0.005693458486348391, + 0.005136349704116583, + 0.007683400996029377, + 0.00893737468868494, + 0.007516917772591114, + 0.015576447360217571, + 0.005495329387485981, + 0.005180204752832651, + 0.004714827984571457, + 0.0033232190180569887, + 0.008626461029052734, + 0.005632283166050911, + 0.0062141395173966885, + 0.009927132166922092, + 0.00534638250246644, + 0.00531783839687705, + 0.004049145150929689, + 0.003094699699431658, + 0.004361218772828579, + 0.0042551178485155106, + 0.004205799195915461, + 0.006051786243915558, + 0.004210286773741245, + 0.0041400897316634655, + 0.004282638896256685, + 0.004112853668630123, + 0.005613801069557667, + 0.0038550414610654116, + 0.004721321165561676, + 0.004028747323900461, + 0.016133328899741173, + 0.012338804081082344, + 0.014943061396479607, + 0.007416621316224337, + 0.017628002911806107, + 0.013006077148020267, + 0.010029601864516735, + 0.011447708122432232, + 0.031526852399110794, + 0.02285686694085598, + 0.027455441653728485, + 0.025824623182415962, + 0.03144875913858414, + 0.030777134001255035, + 0.0319434218108654, + 0.04254601523280144, + 0.005529344081878662, + 0.004261048510670662, + 0.0041360314935445786, + 0.004125483799725771, + 0.004861293826252222, + 0.004067892674356699, + 0.004544960800558329, + 0.005731004290282726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.062037527561187744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031974404118955135 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.0031974404118955135, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031974404118955135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1830575317144394, + 0.3355674743652344, + 0.08148607611656189, + 0.04282015934586525, + 0.04328030347824097, + 0.04603666812181473, + 0.1473340094089508, + 0.05106252804398537, + 0.03497093543410301, + 0.034384388476610184 + ], + "credit_map": [ + 0.02796580269932747, + 0.028332049027085304, + 0.0203329399228096, + 0.01490245945751667, + 0.025654718279838562, + 0.024382086470723152, + 0.01873253472149372, + 0.02275492623448372, + 0.04404078423976898, + 0.03832684084773064, + 0.03455108776688576, + 0.03749348223209381, + 0.05306655168533325, + 0.0475630983710289, + 0.03620801866054535, + 0.044317569583654404, + 0.010540787130594254, + 0.010891950689256191, + 0.008402901701629162, + 0.006936213467270136, + 0.01453833095729351, + 0.010064633563160896, + 0.008594876155257225, + 0.011516380123794079, + 0.006714853458106518, + 0.007220025174319744, + 0.00415852852165699, + 0.003425562521442771, + 0.0050923521630465984, + 0.006979785393923521, + 0.004188769031316042, + 0.005040277726948261, + 0.006116235628724098, + 0.004124615807086229, + 0.004030244424939156, + 0.0049783880822360516, + 0.004095413256436586, + 0.006638965103775263, + 0.00490735936909914, + 0.00838907714933157, + 0.006036409642547369, + 0.00532626872882247, + 0.004360484424978495, + 0.003953895531594753, + 0.009022722952067852, + 0.005454059224575758, + 0.004440591670572758, + 0.007442230824381113, + 0.01780068315565586, + 0.025193672627210617, + 0.015515168197453022, + 0.011493328958749771, + 0.020744236186146736, + 0.01918809860944748, + 0.012516823597252369, + 0.024881986901164055, + 0.00676922919228673, + 0.004887126851826906, + 0.005412594880908728, + 0.005105636548250914, + 0.008314539678394794, + 0.006542103830724955, + 0.004143967758864164, + 0.009887328371405602, + 0.004259869921952486, + 0.004025509115308523, + 0.004858153406530619, + 0.003610257524996996, + 0.0040113842114806175, + 0.0040872241370379925, + 0.004969393834471703, + 0.005149137694388628, + 0.004546635784208775, + 0.004386007320135832, + 0.0041798679158091545, + 0.003409068565815687, + 0.004305234644562006, + 0.003969392739236355, + 0.004116838797926903, + 0.00547133618965745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0031974404118955135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene rendered in Salvador Dalí's stark black-and-white surrealist style. The setting is a barren, dreamlike seafloor, where the sand is a fine, pale powder and the water is perfectly, unsettlingly still, like a vast expanse of solid glass. A massive, iconic, melting clock, its Roman numerals dripping like liquid tar, lies half-buried in the silent landscape. Dominating the frame is a massive crab, its body a speckled gray, armored carapace, standing on incredibly spindly legs that seem to defy physics. Intense, dramatic lighting emanating from a single, unseen source sculpts the scene into extreme contrasts of brilliant white and deep, velvety black, casting long, grotesquely distorted shadows from the crab's legs across the seafloor, making them appear even more elongated and menacing. In a slow, deliberate, and relentless display of menace, the crab raises its two massive, obsidian-black claws, mechanically rising with terrifying precision, its pincers opening and closing with measured, slow pacing, creating a palpable atmosphere of nightmarish tension.", + "seed": 7714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_9513_seed_7714_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": -0.3787080093985423, + "vlm_logprob": -0.3787080093985423, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.3787080093985423, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042719874531030655 + ], + "opposite_logprob_tokens": [ + -3.292719841003418 + ], + "logprob_reward": -0.042719874531030655, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.042719874531030655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13005216419696808, + 0.07714629918336868, + 0.036803796887397766, + 0.049681831151247025, + 0.10863961279392242, + 0.059615977108478546, + 0.06088394299149513, + 0.1725095957517624, + 0.22222816944122314, + 0.08243869990110397 + ], + "credit_map": [ + 0.016755560413002968, + 0.01136759389191866, + 0.009326997213065624, + 0.016183605417609215, + 0.01760552078485489, + 0.008698084391653538, + 0.03824927657842636, + 0.011865523643791676, + 0.00977222342044115, + 0.008868015371263027, + 0.006892717443406582, + 0.009721491485834122, + 0.008142639882862568, + 0.005247585009783506, + 0.020605403929948807, + 0.007896221242845058, + 0.0077699110843241215, + 0.004482480697333813, + 0.005070867948234081, + 0.003021490527316928, + 0.0038243045564740896, + 0.0036899494007229805, + 0.0052719274535775185, + 0.003672865452244878, + 0.005380196496844292, + 0.005796715617179871, + 0.005081006325781345, + 0.005394970066845417, + 0.007200388703495264, + 0.003310374217107892, + 0.01140789594501257, + 0.006110284477472305, + 0.011549764312803745, + 0.01130032166838646, + 0.012416228652000427, + 0.010313479229807854, + 0.013963822275400162, + 0.00714767025783658, + 0.030415944755077362, + 0.011532376520335674, + 0.009660489857196808, + 0.007875882089138031, + 0.0059253862127661705, + 0.007997197099030018, + 0.007665114477276802, + 0.005148470867425203, + 0.00831099133938551, + 0.00703244237229228, + 0.011101565323770046, + 0.007242667954415083, + 0.008024344220757484, + 0.007825012318789959, + 0.007588131818920374, + 0.005252128932625055, + 0.008157462812960148, + 0.005692629609256983, + 0.0351218543946743, + 0.020978037267923355, + 0.023013954982161522, + 0.016547953709959984, + 0.015481915324926376, + 0.03414987027645111, + 0.012056035920977592, + 0.015159974806010723, + 0.04183065891265869, + 0.02851799875497818, + 0.026573222130537033, + 0.03129522129893303, + 0.02993798442184925, + 0.023145798593759537, + 0.02431776188313961, + 0.01660950668156147, + 0.024127865210175514, + 0.008085149340331554, + 0.01260031946003437, + 0.0104425810277462, + 0.008383730426430702, + 0.006003869231790304, + 0.0056837075389921665, + 0.00711147440597415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.042719874531030655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.8048418760299683 + ], + "opposite_logprob_tokens": [ + -0.17984186112880707 + ], + "logprob_reward": -1.8048418760299683, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048418760299683, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12520161271095276, + 0.05941028147935867, + 0.0471869595348835, + 0.07497510313987732, + 0.24326655268669128, + 0.07845140993595123, + 0.12613724172115326, + 0.13127155601978302, + 0.06854183971881866, + 0.04555744677782059 + ], + "credit_map": [ + 0.015939293429255486, + 0.013475223444402218, + 0.014004959724843502, + 0.013621204532682896, + 0.016567159444093704, + 0.008599651977419853, + 0.02963624708354473, + 0.013357867486774921, + 0.00958898663520813, + 0.005058599635958672, + 0.005606015212833881, + 0.006315883249044418, + 0.007291421294212341, + 0.004190006759017706, + 0.017626389861106873, + 0.003732981625944376, + 0.00906990747898817, + 0.005047619808465242, + 0.006419464945793152, + 0.0037755505181849003, + 0.006050929892808199, + 0.005607631988823414, + 0.007179182022809982, + 0.0040366752073168755, + 0.0081039909273386, + 0.010480842553079128, + 0.008338000625371933, + 0.00663383211940527, + 0.013459108769893646, + 0.0057912347838282585, + 0.01275692693889141, + 0.009411166422069073, + 0.025898482650518417, + 0.028177058324217796, + 0.021715784445405006, + 0.021589018404483795, + 0.038741592317819595, + 0.03216242790222168, + 0.05433265119791031, + 0.02064952440559864, + 0.007836468517780304, + 0.010574070736765862, + 0.008203472942113876, + 0.013961267657577991, + 0.009083673357963562, + 0.009615586139261723, + 0.010280865244567394, + 0.00889600906521082, + 0.018031373620033264, + 0.018464474007487297, + 0.019427616149187088, + 0.017261052504181862, + 0.01730138249695301, + 0.012041032314300537, + 0.013887668028473854, + 0.009722643531858921, + 0.018304230645298958, + 0.019146349281072617, + 0.02323148585855961, + 0.014088987372815609, + 0.014632968232035637, + 0.018622633069753647, + 0.012596788816154003, + 0.010648111812770367, + 0.00894022174179554, + 0.009767468087375164, + 0.007519947364926338, + 0.01069993432611227, + 0.009169913828372955, + 0.008942726999521255, + 0.0082219447940588, + 0.005279681645333767, + 0.007300867699086666, + 0.004863040521740913, + 0.007664224598556757, + 0.00760577991604805, + 0.00591100612655282, + 0.0037416990380734205, + 0.003249787725508213, + 0.005221041385084391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -1.8048418760299683 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421316433697939 + ], + "opposite_logprob_tokens": [ + -4.2642130851745605 + ], + "logprob_reward": -0.01421316433697939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421316433697939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14098376035690308, + 0.08094616234302521, + 0.09766131639480591, + 0.07909492403268814, + 0.18772569298744202, + 0.07317093759775162, + 0.10208357870578766, + 0.15115119516849518, + 0.057346079498529434, + 0.02983628772199154 + ], + "credit_map": [ + 0.0156090771779418, + 0.01655779965221882, + 0.014649901539087296, + 0.012881730683147907, + 0.02666652947664261, + 0.014245246537029743, + 0.020570818334817886, + 0.01980266533792019, + 0.011044819839298725, + 0.008617406710982323, + 0.00841063354164362, + 0.007192695513367653, + 0.01651446521282196, + 0.008081475272774696, + 0.01458924263715744, + 0.006495426408946514, + 0.012098671868443489, + 0.009740076959133148, + 0.0109181497246027, + 0.009851658716797829, + 0.014050522819161415, + 0.008002492599189281, + 0.025234483182430267, + 0.0077652595937252045, + 0.007342142518609762, + 0.009662245400249958, + 0.007625785656273365, + 0.007652566768229008, + 0.012627938762307167, + 0.005046776495873928, + 0.02116287872195244, + 0.007974585518240929, + 0.018966784700751305, + 0.021061493083834648, + 0.016903793439269066, + 0.01775956153869629, + 0.029538219794631004, + 0.02104371413588524, + 0.047878198325634, + 0.01457393728196621, + 0.007241880986839533, + 0.009706339798867702, + 0.007349689025431871, + 0.01379450410604477, + 0.008900618180632591, + 0.00758999353274703, + 0.010746574960649014, + 0.007841336540877819, + 0.01658843830227852, + 0.013860732316970825, + 0.01432301476597786, + 0.01640406996011734, + 0.011379914358258247, + 0.008792301639914513, + 0.010745209641754627, + 0.009989900514483452, + 0.02100384794175625, + 0.02353092096745968, + 0.029133377596735954, + 0.018373258411884308, + 0.019180243834853172, + 0.016193371266126633, + 0.011577569879591465, + 0.012158604338765144, + 0.008430229499936104, + 0.007585892453789711, + 0.006631693337112665, + 0.008814258500933647, + 0.0076128095388412476, + 0.0070975772105157375, + 0.005930996034294367, + 0.005242621060460806, + 0.00404901709407568, + 0.0034859690349549055, + 0.003151301760226488, + 0.003943327814340591, + 0.005066453013569117, + 0.002996075199916959, + 0.0023602868895977736, + 0.004783857148140669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.01421316433697939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029814982786774635 + ], + "opposite_logprob_tokens": [ + -3.5298149585723877 + ], + "logprob_reward": -0.029814982786774635, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029814982786774635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10173920542001724, + 0.1337185651063919, + 0.05845839902758598, + 0.11307916045188904, + 0.2371932417154312, + 0.06860470026731491, + 0.08222571760416031, + 0.1220909133553505, + 0.04538290575146675, + 0.03750724717974663 + ], + "credit_map": [ + 0.013655735179781914, + 0.00991719402372837, + 0.009471596218645573, + 0.013139146380126476, + 0.0164390467107296, + 0.007387798745185137, + 0.021692294627428055, + 0.010036383755505085, + 0.01663769595324993, + 0.013578678481280804, + 0.013266700319945812, + 0.012518863193690777, + 0.024195214733481407, + 0.010396781377494335, + 0.029798757284879684, + 0.013325847685337067, + 0.008775627240538597, + 0.0053227865137159824, + 0.006962260697036982, + 0.005385383497923613, + 0.006581065244972706, + 0.004869489464908838, + 0.015873048454523087, + 0.004688732326030731, + 0.012385247275233269, + 0.013658608309924603, + 0.011166003532707691, + 0.01064052153378725, + 0.0173528753221035, + 0.009112926200032234, + 0.02470387890934944, + 0.014059092849493027, + 0.02851817011833191, + 0.025388406589627266, + 0.022742686793208122, + 0.019929436966776848, + 0.03345513716340065, + 0.04558389261364937, + 0.039287831634283066, + 0.02228764072060585, + 0.0094660185277462, + 0.009087467566132545, + 0.008086703717708588, + 0.012204527854919434, + 0.007905462756752968, + 0.004289012867957354, + 0.009535714983940125, + 0.008029784075915813, + 0.014607604593038559, + 0.01039520837366581, + 0.012439079582691193, + 0.012251150794327259, + 0.009550579823553562, + 0.0074497973546385765, + 0.009330123662948608, + 0.00620216503739357, + 0.01683310605585575, + 0.014898378401994705, + 0.015593991614878178, + 0.012574180960655212, + 0.024043582379817963, + 0.017728490754961967, + 0.011051200330257416, + 0.009367969818413258, + 0.007693609222769737, + 0.005198346916586161, + 0.006015294697135687, + 0.0064933039247989655, + 0.006326482631266117, + 0.004926957655698061, + 0.003976942040026188, + 0.004751964472234249, + 0.005009087268263102, + 0.004220583010464907, + 0.006002943962812424, + 0.006286717485636473, + 0.004887780640274286, + 0.0028931577689945698, + 0.0027206288650631905, + 0.005486346315592527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.029814982786774635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019501493079587817 + ], + "opposite_logprob_tokens": [ + -6.251950263977051 + ], + "logprob_reward": -0.0019501493079587817, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019501493079587817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0917351245880127, + 0.06467144191265106, + 0.05533555522561073, + 0.09667341411113739, + 0.19877827167510986, + 0.061825986951589584, + 0.08541971445083618, + 0.14573171734809875, + 0.138613760471344, + 0.06121507287025452 + ], + "credit_map": [ + 0.01193926576524973, + 0.010514981113374233, + 0.010480903089046478, + 0.010902433656156063, + 0.01325917523354292, + 0.009032328613102436, + 0.01369108259677887, + 0.011914952658116817, + 0.011237329803407192, + 0.005853572394698858, + 0.006671689450740814, + 0.006257487460970879, + 0.008623359724879265, + 0.0050834291614592075, + 0.016769427806138992, + 0.004175145644694567, + 0.010246631689369678, + 0.005843700375407934, + 0.0074834576807916164, + 0.004725200589746237, + 0.007044410798698664, + 0.005882661323994398, + 0.009055845439434052, + 0.0050536454655230045, + 0.00930622685700655, + 0.012464581988751888, + 0.009916380047798157, + 0.009370990097522736, + 0.015295621939003468, + 0.006896461360156536, + 0.02148929424583912, + 0.011933854781091213, + 0.019351687282323837, + 0.025081345811486244, + 0.017698587849736214, + 0.019388774409890175, + 0.03415569290518761, + 0.018380001187324524, + 0.04757242277264595, + 0.017149746417999268, + 0.008921238593757153, + 0.008007245138287544, + 0.006928206421434879, + 0.008818351663649082, + 0.008769331499934196, + 0.006623512599617243, + 0.006893928628414869, + 0.006864172872155905, + 0.00981332641094923, + 0.012621160596609116, + 0.01128388475626707, + 0.01106962375342846, + 0.014356032013893127, + 0.008214916102588177, + 0.011040549725294113, + 0.007020214572548866, + 0.024719642475247383, + 0.020214373245835304, + 0.02080347016453743, + 0.01824408397078514, + 0.017348380759358406, + 0.021424047648906708, + 0.011327303946018219, + 0.011650407686829567, + 0.021794768050312996, + 0.0183310117572546, + 0.017733782529830933, + 0.02519034966826439, + 0.017843542620539665, + 0.012555516324937344, + 0.01537203136831522, + 0.009792759083211422, + 0.014335579238831997, + 0.0066605377942323685, + 0.008692219853401184, + 0.007815244607627392, + 0.007813052274286747, + 0.005065146367996931, + 0.00462561147287488, + 0.006207682657986879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.0019501493079587817 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene rendered in Salvador Dalí's stark black-and-white surrealist style. The setting is a barren, dreamlike seafloor, where the sand is a fine, pale powder and the water is perfectly, unsettlingly still, like a vast expanse of solid glass. A massive, iconic, melting clock, its Roman numerals dripping like liquid tar, lies half-buried in the silent landscape. Dominating the frame is a massive crab, its body a speckled gray, armored carapace, standing on incredibly spindly legs that seem to defy physics. Intense, dramatic lighting emanating from a single, unseen source sculpts the scene into extreme contrasts of brilliant white and deep, velvety black, casting long, grotesquely distorted shadows from the crab's legs across the seafloor, making them appear even more elongated and menacing. In a slow, deliberate, and relentless display of menace, the crab raises its two massive, obsidian-black claws, mechanically rising with terrifying precision, its pincers opening and closing with measured, slow pacing, creating a palpable atmosphere of nightmarish tension.", + "seed": 7715, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_9513_seed_7715_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": -0.33619337221607565, + "vlm_logprob": -0.33619337221607565, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.33619337221607565, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0099091287702322 + ], + "opposite_logprob_tokens": [ + -5.259909152984619 + ], + "logprob_reward": -0.0099091287702322, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0099091287702322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11580152064561844, + 0.2225225865840912, + 0.16478295624256134, + 0.06697073578834534, + 0.07707542926073074, + 0.09818470478057861, + 0.08349079638719559, + 0.0956130102276802, + 0.023103538900613785, + 0.05245472863316536 + ], + "credit_map": [ + 0.023866036906838417, + 0.010791949927806854, + 0.006943488959223032, + 0.008730852045118809, + 0.021836746484041214, + 0.01351920422166586, + 0.013122746720910072, + 0.01699049025774002, + 0.014503475278615952, + 0.020527351647615433, + 0.009985298849642277, + 0.01258456613868475, + 0.08788041770458221, + 0.04897214099764824, + 0.014447005465626717, + 0.013622326776385307, + 0.014262771233916283, + 0.011917641386389732, + 0.01857101544737816, + 0.06665007770061493, + 0.016499008983373642, + 0.019231436774134636, + 0.009421585127711296, + 0.008229406550526619, + 0.007270199246704578, + 0.005115768406540155, + 0.00860872957855463, + 0.010119360871613026, + 0.015131305903196335, + 0.004506416153162718, + 0.008156009949743748, + 0.008062942884862423, + 0.011078407987952232, + 0.008468590676784515, + 0.008928589522838593, + 0.01154579222202301, + 0.014142115600407124, + 0.007424775045365095, + 0.00917544774711132, + 0.006311705335974693, + 0.014173963107168674, + 0.010658175684511662, + 0.00576402572914958, + 0.00938926450908184, + 0.02889281138777733, + 0.012897110544145107, + 0.006394325289875269, + 0.010015031322836876, + 0.013208682648837566, + 0.0070638637989759445, + 0.006156306713819504, + 0.007109575904905796, + 0.013061301782727242, + 0.012537417002022266, + 0.005041974596679211, + 0.01931166648864746, + 0.029388682916760445, + 0.006340078543871641, + 0.006273389793932438, + 0.0046230098232626915, + 0.006298995576798916, + 0.03079901821911335, + 0.0037090133409947157, + 0.00818081758916378, + 0.004417981021106243, + 0.002790939761325717, + 0.0023076424840837717, + 0.002051504096016288, + 0.0032802028581500053, + 0.003707312513142824, + 0.0015115413116291165, + 0.0030364140402525663, + 0.013439410366117954, + 0.006742531433701515, + 0.003566694911569357, + 0.00453433021903038, + 0.006425411440432072, + 0.00858872476965189, + 0.002940591424703598, + 0.006217030342668295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0099091287702322 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.6005443334579468 + ], + "opposite_logprob_tokens": [ + -0.2255443036556244 + ], + "logprob_reward": -1.6005443334579468, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005443334579468, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.178552508354187, + 0.22240597009658813, + 0.08211293071508408, + 0.05378647893667221, + 0.12282527983188629, + 0.022933349013328552, + 0.040716949850320816, + 0.12475140392780304, + 0.030672768130898476, + 0.12124238163232803 + ], + "credit_map": [ + 0.030110545456409454, + 0.011852113530039787, + 0.012000937946140766, + 0.012534931302070618, + 0.022939026355743408, + 0.040233392268419266, + 0.015259988605976105, + 0.03362158685922623, + 0.020253624767065048, + 0.021303264424204826, + 0.011934855952858925, + 0.014413234777748585, + 0.09182653576135635, + 0.03260553255677223, + 0.016495585441589355, + 0.013573321513831615, + 0.009013275615870953, + 0.006933887489140034, + 0.005750250071287155, + 0.016884809359908104, + 0.014647883363068104, + 0.016098428517580032, + 0.0047747171483933926, + 0.008009680546820164, + 0.005362504627555609, + 0.00525682233273983, + 0.004575829487293959, + 0.006835348904132843, + 0.008603554219007492, + 0.006957643199712038, + 0.005383008159697056, + 0.010811769403517246, + 0.024864191189408302, + 0.012504748068749905, + 0.010711326263844967, + 0.01672546938061714, + 0.019639087840914726, + 0.01184001099318266, + 0.011348004452884197, + 0.015192437916994095, + 0.0025058321189135313, + 0.002137890551239252, + 0.0026355902664363384, + 0.002760177943855524, + 0.004657662007957697, + 0.0027444656006991863, + 0.001871525659225881, + 0.003620205447077751, + 0.0050388481467962265, + 0.003061268711462617, + 0.0027607062365859747, + 0.004551473539322615, + 0.004371383227407932, + 0.011103453114628792, + 0.0025542450603097677, + 0.007275571580976248, + 0.03934003785252571, + 0.008911124430596828, + 0.008266327902674675, + 0.005452613811939955, + 0.009052146226167679, + 0.038714002817869186, + 0.004861036781221628, + 0.01015410665422678, + 0.004212264437228441, + 0.003545735962688923, + 0.0035739478189498186, + 0.004139096476137638, + 0.0033311096485704184, + 0.0054769874550402164, + 0.002748477505519986, + 0.0036451483611017466, + 0.019967159256339073, + 0.023620080202817917, + 0.009323432110249996, + 0.010593457147479057, + 0.02143472246825695, + 0.015180393122136593, + 0.008394859731197357, + 0.012728270143270493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -1.6005443334579468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005957346875220537 + ], + "opposite_logprob_tokens": [ + -5.130957126617432 + ], + "logprob_reward": -0.005957346875220537, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957346875220537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14163054525852203, + 0.16694426536560059, + 0.10160005837678909, + 0.07606291770935059, + 0.10575699806213379, + 0.06517285108566284, + 0.0645214095711708, + 0.12317480146884918, + 0.035088423639535904, + 0.12004778534173965 + ], + "credit_map": [ + 0.021725820377469063, + 0.009787343442440033, + 0.010809804312884808, + 0.010419580154120922, + 0.017077583819627762, + 0.033563222736120224, + 0.011774195358157158, + 0.02647298201918602, + 0.018603798002004623, + 0.01517411507666111, + 0.008917382918298244, + 0.00961641501635313, + 0.05344652384519577, + 0.03936018422245979, + 0.011193729005753994, + 0.010632111690938473, + 0.012896399013698101, + 0.009851254522800446, + 0.010948305949568748, + 0.01724921353161335, + 0.016895407810807228, + 0.01599280536174774, + 0.006557018030434847, + 0.011209654621779919, + 0.01252283900976181, + 0.00621924689039588, + 0.006717353593558073, + 0.012411444447934628, + 0.009649733081459999, + 0.008321532979607582, + 0.0061098989099264145, + 0.014110861346125603, + 0.01603870280086994, + 0.0092332037165761, + 0.006573684047907591, + 0.011840017512440681, + 0.015507936477661133, + 0.02208784781396389, + 0.008059942163527012, + 0.016415663063526154, + 0.008414491079747677, + 0.006725613493472338, + 0.004615800455212593, + 0.005621162708848715, + 0.018204376101493835, + 0.007680382113903761, + 0.004702598787844181, + 0.009208419360220432, + 0.008396425284445286, + 0.004498214460909367, + 0.003655742621049285, + 0.006731952540576458, + 0.007238393649458885, + 0.017542492598295212, + 0.004162147641181946, + 0.012296036817133427, + 0.03789046034216881, + 0.008672412484884262, + 0.00822977814823389, + 0.00573658337816596, + 0.010114438831806183, + 0.03773343563079834, + 0.004715621471405029, + 0.010082061402499676, + 0.0054742624051868916, + 0.004017184488475323, + 0.003438700921833515, + 0.003832260612398386, + 0.003633965039625764, + 0.007223137654364109, + 0.0026545445434749126, + 0.004814363084733486, + 0.021590298041701317, + 0.021689649671316147, + 0.009682832285761833, + 0.00980277732014656, + 0.019137080758810043, + 0.01714804768562317, + 0.007554546929895878, + 0.013442552648484707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.005957346875220537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206173077225685 + ], + "opposite_logprob_tokens": [ + -2.8120617866516113 + ], + "logprob_reward": -0.06206173077225685, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206173077225685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0834035649895668, + 0.2971174120903015, + 0.1353510469198227, + 0.06568343192338943, + 0.21845823526382446, + 0.03767724707722664, + 0.021591024473309517, + 0.017433134838938713, + 0.058707110583782196, + 0.0645778551697731 + ], + "credit_map": [ + 0.012885767966508865, + 0.007369683124125004, + 0.0057966141030192375, + 0.009021754376590252, + 0.018668580800294876, + 0.009772076271474361, + 0.00918852910399437, + 0.010700554586946964, + 0.0741010308265686, + 0.024779627099633217, + 0.01529651042073965, + 0.03478112444281578, + 0.09041354805231094, + 0.03129485994577408, + 0.01333029568195343, + 0.013120405375957489, + 0.022646965458989143, + 0.008954846300184727, + 0.016907421872019768, + 0.05295616015791893, + 0.011139296926558018, + 0.00783690344542265, + 0.007618541829288006, + 0.007290910463780165, + 0.009302964434027672, + 0.005282931961119175, + 0.008696201257407665, + 0.01596064679324627, + 0.009714809246361256, + 0.0062139397487044334, + 0.004922987427562475, + 0.005588945001363754, + 0.04619782790541649, + 0.02005077339708805, + 0.023444773629307747, + 0.025439169257879257, + 0.034747060388326645, + 0.029271449893712997, + 0.019274666905403137, + 0.02003251016139984, + 0.008462777361273766, + 0.003934413194656372, + 0.004950987175107002, + 0.006326192058622837, + 0.00458715483546257, + 0.0037980491761118174, + 0.0021977154538035393, + 0.003419957123696804, + 0.003364989999681711, + 0.0021503684110939503, + 0.002521759131923318, + 0.002402452751994133, + 0.0035239572171121836, + 0.0030738902278244495, + 0.0014948530588299036, + 0.003058752743527293, + 0.002293120138347149, + 0.0016235160874202847, + 0.0016433445271104574, + 0.00184223847463727, + 0.0040061743929982185, + 0.001957345986738801, + 0.0016058020992204547, + 0.0024615933652967215, + 0.00791085883975029, + 0.004621289670467377, + 0.005241030361503363, + 0.006105637177824974, + 0.004811485297977924, + 0.010816466994583607, + 0.004047215450555086, + 0.015153128653764725, + 0.010959724895656109, + 0.010827130638062954, + 0.005396509543061256, + 0.005830847192555666, + 0.008939669467508793, + 0.010241643525660038, + 0.0038739186711609364, + 0.008508406579494476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.06206173077225685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024943212047219276 + ], + "opposite_logprob_tokens": [ + -6.0024943351745605 + ], + "logprob_reward": -0.0024943212047219276, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024943212047219276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1710866093635559, + 0.1550675481557846, + 0.1007668673992157, + 0.05392260476946831, + 0.05004280060529709, + 0.08381999284029007, + 0.06332230567932129, + 0.18824328482151031, + 0.026130156591534615, + 0.10759773850440979 + ], + "credit_map": [ + 0.025205861777067184, + 0.012100780382752419, + 0.011233414523303509, + 0.011695017106831074, + 0.024583725258708, + 0.0363590344786644, + 0.01596016064286232, + 0.0339486300945282, + 0.017346791923046112, + 0.014755905605852604, + 0.009333203546702862, + 0.009658259339630604, + 0.06156912446022034, + 0.02040763385593891, + 0.012157673947513103, + 0.009838970378041267, + 0.012890786863863468, + 0.006412660237401724, + 0.006364261265844107, + 0.020173555240035057, + 0.015623924322426319, + 0.02381007932126522, + 0.007511353585869074, + 0.007980259135365486, + 0.006368051283061504, + 0.005713190417736769, + 0.003946144133806229, + 0.006764710880815983, + 0.0085922721773386, + 0.006199969910085201, + 0.005282040219753981, + 0.011056234128773212, + 0.006868208292871714, + 0.0051223901100456715, + 0.004796938039362431, + 0.009463528171181679, + 0.007236415985971689, + 0.0045186919160187244, + 0.006485998164862394, + 0.0055506350472569466, + 0.011175673454999924, + 0.008586629293859005, + 0.0047046416439116, + 0.006779561750590801, + 0.02313338778913021, + 0.013896280899643898, + 0.005359314382076263, + 0.010184520855545998, + 0.008217097260057926, + 0.00481549883261323, + 0.0035520729143172503, + 0.005230363924056292, + 0.010805952362716198, + 0.013942664489150047, + 0.0038887180853635073, + 0.012869940139353275, + 0.05856272950768471, + 0.013141456991434097, + 0.012326340191066265, + 0.008492019027471542, + 0.013729349710047245, + 0.05948605388402939, + 0.007241958752274513, + 0.015263412147760391, + 0.003757239319384098, + 0.002904612338170409, + 0.0032728174701333046, + 0.003357192501425743, + 0.003863709978759289, + 0.0028343030717223883, + 0.0024225441738963127, + 0.0037177391350269318, + 0.02603713609278202, + 0.009245025925338268, + 0.009449425153434277, + 0.009988204576075077, + 0.011352458968758583, + 0.01788678765296936, + 0.007224587723612785, + 0.01641412451863289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0024943212047219276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide-angle shot captures a scene rendered in Salvador Dalí's stark black-and-white surrealist style. The setting is a barren, dreamlike seafloor, where the sand is a fine, pale powder and the water is perfectly, unsettlingly still, like a vast expanse of solid glass. A massive, iconic, melting clock, its Roman numerals dripping like liquid tar, lies half-buried in the silent landscape. Dominating the frame is a massive crab, its body a speckled gray, armored carapace, standing on incredibly spindly legs that seem to defy physics. Intense, dramatic lighting emanating from a single, unseen source sculpts the scene into extreme contrasts of brilliant white and deep, velvety black, casting long, grotesquely distorted shadows from the crab's legs across the seafloor, making them appear even more elongated and menacing. In a slow, deliberate, and relentless display of menace, the crab raises its two massive, obsidian-black claws, mechanically rising with terrifying precision, its pincers opening and closing with measured, slow pacing, creating a palpable atmosphere of nightmarish tension.", + "seed": 7716, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_14/index_9513_seed_7716_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": -0.03746575448894873, + "vlm_logprob": -0.03746575448894873, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.03746575448894873, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015747476369142532 + ], + "opposite_logprob_tokens": [ + -4.640747547149658 + ], + "logprob_reward": -0.015747476369142532, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.015747476369142532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3582405149936676, + 0.14710883796215057, + 0.055954527109861374, + 0.02648576721549034, + 0.03412901610136032, + 0.15862874686717987, + 0.07226448506116867, + 0.02501508966088295, + 0.03187667578458786, + 0.09029630571603775 + ], + "credit_map": [ + 0.05013788864016533, + 0.023428622633218765, + 0.06346096843481064, + 0.024624237790703773, + 0.018280528485774994, + 0.0678257942199707, + 0.04000929370522499, + 0.07047320902347565, + 0.023706236854195595, + 0.01096008624881506, + 0.04554583877325058, + 0.013834173791110516, + 0.009754205122590065, + 0.012638687156140804, + 0.00988245103508234, + 0.020787175744771957, + 0.00866718403995037, + 0.005472199991345406, + 0.01258882973343134, + 0.005540438927710056, + 0.006895603146404028, + 0.006007497198879719, + 0.0038469049613922834, + 0.006935869809240103, + 0.005155067425221205, + 0.003051708685234189, + 0.004518663976341486, + 0.00240016495808959, + 0.0025190196465700865, + 0.002757643349468708, + 0.003113379003480077, + 0.0029701220337301493, + 0.005901094526052475, + 0.005704961717128754, + 0.00391795439645648, + 0.003114424180239439, + 0.005703364964574575, + 0.004342327825725079, + 0.0023517522495239973, + 0.003093137638643384, + 0.027095191180706024, + 0.02020719274878502, + 0.02273658476769924, + 0.014582452364265919, + 0.016347933560609818, + 0.020975211635231972, + 0.01597186177968979, + 0.02071232907474041, + 0.01345225889235735, + 0.008073666132986546, + 0.009201397188007832, + 0.006607364397495985, + 0.008670495823025703, + 0.008680012077093124, + 0.006734892260283232, + 0.010844405740499496, + 0.005887516308575869, + 0.0031672988552600145, + 0.0026992489583790302, + 0.0022616549395024776, + 0.0037053725682199, + 0.002281840890645981, + 0.0020348713733255863, + 0.002977290190756321, + 0.00431510154157877, + 0.004418991040438414, + 0.004551575053483248, + 0.005016489885747433, + 0.0026315918657928705, + 0.0039582340978085995, + 0.0033381639514118433, + 0.003646530443802476, + 0.014393821358680725, + 0.014880993403494358, + 0.008775199763476849, + 0.005997503641992807, + 0.018679289147257805, + 0.009752186946570873, + 0.00796532817184925, + 0.009851992130279541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.015747476369142532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790175348520279 + ], + "opposite_logprob_tokens": [ + -2.5790176391601562 + ], + "logprob_reward": -0.0790175348520279, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790175348520279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4599476456642151, + 0.033974263817071915, + 0.015346725471317768, + 0.02808857150375843, + 0.0193017590790987, + 0.013211572542786598, + 0.05232371762394905, + 0.1499348282814026, + 0.13824868202209473, + 0.08962225168943405 + ], + "credit_map": [ + 0.049182530492544174, + 0.020142855122685432, + 0.042978208512067795, + 0.020377924665808678, + 0.020778242498636246, + 0.10588061809539795, + 0.06362303346395493, + 0.13698425889015198, + 0.004893898963928223, + 0.0039179688319563866, + 0.006832010578364134, + 0.003018898656591773, + 0.00481001241132617, + 0.0037906821817159653, + 0.003280940931290388, + 0.003429850097745657, + 0.002246395917609334, + 0.002558185951784253, + 0.0019420965109020472, + 0.0013146003475412726, + 0.002182848984375596, + 0.0016685370355844498, + 0.0017951909685507417, + 0.0016388694057241082, + 0.005400697235018015, + 0.003226057393476367, + 0.004870258737355471, + 0.0021356737706810236, + 0.003862203098833561, + 0.0027295402251183987, + 0.0028206740971654654, + 0.0030434660147875547, + 0.003767467802390456, + 0.0028507804963737726, + 0.0016750061186030507, + 0.0016065383097156882, + 0.0023768141400069, + 0.00261539570055902, + 0.0021968886721879244, + 0.002212867373600602, + 0.0017162625445052981, + 0.002460598247125745, + 0.0015093710971996188, + 0.0017150604398921132, + 0.0022224478889256716, + 0.0010724663734436035, + 0.0011643067700788379, + 0.0013510584831237793, + 0.00747244106605649, + 0.007399111520498991, + 0.00468027126044035, + 0.00407415721565485, + 0.00898917019367218, + 0.009754367172718048, + 0.004125075414776802, + 0.005829124245792627, + 0.031991295516490936, + 0.017756691202521324, + 0.02314756065607071, + 0.016061322763562202, + 0.012339695356786251, + 0.014118978753685951, + 0.014585600234568119, + 0.019933678209781647, + 0.02163092792034149, + 0.013889085501432419, + 0.019353142008185387, + 0.020599521696567535, + 0.009037713520228863, + 0.013368913903832436, + 0.014902209863066673, + 0.025467168539762497, + 0.017239026725292206, + 0.011781987734138966, + 0.010155415162444115, + 0.007342270575463772, + 0.008925259113311768, + 0.009267928078770638, + 0.009106051176786423, + 0.015804316848516464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0790175348520279 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007194076315499842 + ], + "opposite_logprob_tokens": [ + -7.2507195472717285 + ], + "logprob_reward": -0.0007194076315499842, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007194076315499842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39032265543937683, + 0.05768967419862747, + 0.036575812846422195, + 0.02619551494717598, + 0.01962275244295597, + 0.034215670078992844, + 0.0961674302816391, + 0.07959704846143723, + 0.1915741115808487, + 0.06803937256336212 + ], + "credit_map": [ + 0.03920717537403107, + 0.02252330631017685, + 0.03533444553613663, + 0.02267633005976677, + 0.022134868428111076, + 0.07756710052490234, + 0.06293299049139023, + 0.10794641077518463, + 0.009239647537469864, + 0.008219202980399132, + 0.006292860489338636, + 0.0067567965015769005, + 0.0066063497215509415, + 0.0068319751881062984, + 0.00816077459603548, + 0.005582062527537346, + 0.006662632804363966, + 0.006719794124364853, + 0.003485912224277854, + 0.003241102211177349, + 0.0053099761717021465, + 0.0044251903891563416, + 0.0033997083082795143, + 0.003331494517624378, + 0.005582425277680159, + 0.004293191246688366, + 0.0030495328828692436, + 0.0020940860267728567, + 0.0038109745364636183, + 0.0027259723283350468, + 0.002491585910320282, + 0.0021477462723851204, + 0.003476651618257165, + 0.002145129255950451, + 0.0036539940629154444, + 0.002373016206547618, + 0.0020223648753017187, + 0.0016398727893829346, + 0.0022382549941539764, + 0.0020734653808176517, + 0.004795042797923088, + 0.004085248801857233, + 0.005192659329622984, + 0.004105037543922663, + 0.004320315085351467, + 0.0036689206026494503, + 0.0040009506046772, + 0.004047490190714598, + 0.015536571852862835, + 0.014858764596283436, + 0.006728855427354574, + 0.0069839684292674065, + 0.017099393531680107, + 0.018785053864121437, + 0.007306072860956192, + 0.00886873621493578, + 0.016216706484556198, + 0.009837459772825241, + 0.012430480681359768, + 0.008018444292247295, + 0.007330537773668766, + 0.007336469367146492, + 0.007484704256057739, + 0.010942243970930576, + 0.02674790471792221, + 0.02008999139070511, + 0.027992311865091324, + 0.03065042942762375, + 0.012481146492064, + 0.017654873430728912, + 0.02136487141251564, + 0.03459256887435913, + 0.013664898462593555, + 0.008550720289349556, + 0.008538741618394852, + 0.0060826255939900875, + 0.011621731333434582, + 0.007170972414314747, + 0.006043912842869759, + 0.006365762557834387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.0007194076315499842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902343362569809 + ], + "opposite_logprob_tokens": [ + -2.4640233516693115 + ], + "logprob_reward": -0.08902343362569809, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902343362569809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4285356402397156, + 0.09106172621250153, + 0.029139766469597816, + 0.03710068017244339, + 0.06190446764230728, + 0.019969208166003227, + 0.04385332018136978, + 0.08190375566482544, + 0.16572067141532898, + 0.04081069305539131 + ], + "credit_map": [ + 0.04539181664586067, + 0.026462657377123833, + 0.04867911338806152, + 0.025979993864893913, + 0.027811335399746895, + 0.08706022053956985, + 0.0637284591794014, + 0.10342206805944443, + 0.013458912260830402, + 0.0145381735637784, + 0.009371203370392323, + 0.010445978492498398, + 0.014738284051418304, + 0.008950596675276756, + 0.010800239630043507, + 0.008758347481489182, + 0.004978632554411888, + 0.0043675247579813, + 0.0044375271536409855, + 0.0025566499680280685, + 0.004082102794200182, + 0.0028950772248208523, + 0.0023374042939394712, + 0.0034848481882363558, + 0.006781026255339384, + 0.0041210544295609, + 0.003591582179069519, + 0.0038533988408744335, + 0.005134384613484144, + 0.0033822483383119106, + 0.005518087185919285, + 0.0047188978642225266, + 0.011951574124395847, + 0.010352285578846931, + 0.00655389903113246, + 0.005939916241914034, + 0.00702154403552413, + 0.0070538269355893135, + 0.006201793905347586, + 0.0068296282552182674, + 0.002815275453031063, + 0.0034320384729653597, + 0.0020681354217231274, + 0.002156395697966218, + 0.0038615409284830093, + 0.0021428437903523445, + 0.001738766091875732, + 0.0017542112618684769, + 0.008487403392791748, + 0.0055708917789161205, + 0.004109896253794432, + 0.0032843775115907192, + 0.00684497831389308, + 0.0072801243513822556, + 0.0033237175084650517, + 0.0049519287422299385, + 0.014005957171320915, + 0.010711437091231346, + 0.00954251829534769, + 0.012683810666203499, + 0.008317755535244942, + 0.00844084843993187, + 0.010919452644884586, + 0.007281978148967028, + 0.026136361062526703, + 0.016048330813646317, + 0.024627264589071274, + 0.02468106336891651, + 0.01049422100186348, + 0.015543047338724136, + 0.017697257921099663, + 0.030493123456835747, + 0.00647807028144598, + 0.006657699588686228, + 0.003463436383754015, + 0.0029400449711829424, + 0.0071246628649532795, + 0.005229101050645113, + 0.004023558925837278, + 0.004894123412668705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.08902343362569809 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene rendered in a stark black-and-white surrealist style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive melting clock with dripping Roman numerals visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive crab with a speckled gray armored carapace and spindly legs dominating the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the crab raise its two massive obsidian-black claws in a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting intense and dramatic, creating extreme contrasts of brilliant white and deep black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820919966325164 + ], + "opposite_logprob_tokens": [ + -5.87782096862793 + ], + "logprob_reward": -0.002820919966325164, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820919966325164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5793159604072571, + 0.07313302904367447, + 0.042637087404727936, + 0.01719340682029724, + 0.012829123064875603, + 0.05354436859488487, + 0.04037562757730484, + 0.05992702394723892, + 0.0872386172413826, + 0.03380568325519562 + ], + "credit_map": [ + 0.0673205554485321, + 0.028509991243481636, + 0.05339668318629265, + 0.027003522962331772, + 0.0250692255795002, + 0.13576972484588623, + 0.08171384036540985, + 0.16053245961666107, + 0.009883369319140911, + 0.00805788766592741, + 0.017339369282126427, + 0.006976584903895855, + 0.008546927012503147, + 0.006102608982473612, + 0.006615949794650078, + 0.009610342793166637, + 0.006603199057281017, + 0.006206749472767115, + 0.006454065442085266, + 0.003489192109555006, + 0.007628228981047869, + 0.004254295025020838, + 0.0032630625646561384, + 0.004738295450806618, + 0.003008719999343157, + 0.0018157875165343285, + 0.0027320997323840857, + 0.001399170607328415, + 0.0018019807757809758, + 0.001992851262912154, + 0.002527586417272687, + 0.0019152108579874039, + 0.002867239061743021, + 0.0022072214633226395, + 0.0012009046040475368, + 0.0013313014060258865, + 0.0015379330143332481, + 0.0013523375382646918, + 0.0013666549930348992, + 0.0009655325557105243, + 0.009730513207614422, + 0.006555141881108284, + 0.008127006702125072, + 0.004951459355652332, + 0.006002469453960657, + 0.006621526554226875, + 0.005064602009952068, + 0.006491652224212885, + 0.008621995337307453, + 0.004533447790890932, + 0.008921406231820583, + 0.0036833970807492733, + 0.003471750533208251, + 0.0039685373194515705, + 0.0023571408819407225, + 0.004817956127226353, + 0.013226429000496864, + 0.007296683266758919, + 0.009506615810096264, + 0.00604095496237278, + 0.006079073995351791, + 0.005011115688830614, + 0.0055975038558244705, + 0.007168645970523357, + 0.013058226555585861, + 0.008391734212636948, + 0.012733986601233482, + 0.014081533998250961, + 0.005684489384293556, + 0.008384528569877148, + 0.010361004620790482, + 0.014543124474585056, + 0.00594053789973259, + 0.003905325662344694, + 0.002503175986930728, + 0.0034977453760802746, + 0.004972246941179037, + 0.0035353743005543947, + 0.0033418992534279823, + 0.006109381560236216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.002820919966325164 + } + ], + "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/0000055/group_15/index_tensor([3487, 3487])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_tensor([3487, 3487])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..430b20cee7117fc699cb572429a3881de7dbe972 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_tensor([3487, 3487])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot with slight handheld camera shake, tilted slightly upward and panning slowly to the right, captures an incredibly surreal scene: a massive B-17 Flying Fortress bomber, its weathered silver fuselage reflecting a dull luster in the soft light, its rivets and minute battle scars faintly visible, its four massive propeller engines at rest. With a grace and composure that defies all physics, the behemoth glides slowly over the water, its belly almost touching the still surface as it approaches an old wooden pier, leaving a trail of faint, lingering ripples in its wake. The entire scene is bathed in the soft tones of dawn or dusk, the sky a dreamy gradation of lavender and pink-orange. This light evenly casts light across the aircraft, the water, and the pier, creating a tranquil, dreamy, and slightly unsettling atmosphere. As the plane approaches at an impossibly slow and steady pace, the camera shakes slightly, as if the observer is holding their breath at the miraculous sight before them, further enhancing the unique quality of the image, which is both unreal and extremely real.", + "seed": 14019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_3487_seed_14019_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": -0.03278535506688059, + "vlm_logprob": -0.03278535506688059, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03278535506688059, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07794628292322159 + ], + "opposite_logprob_tokens": [ + -2.827946186065674 + ], + "logprob_reward": -0.07794628292322159, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07794628292322159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12869034707546234, + 0.03569639101624489, + 0.07194951176643372, + 0.05523808300495148, + 0.07119257748126984, + 0.21256740391254425, + 0.049286674708127975, + 0.08539494127035141, + 0.17843711376190186, + 0.11154698580503464 + ], + "credit_map": [ + 0.015281188301742077, + 0.014793499372899532, + 0.014421463944017887, + 0.015774404630064964, + 0.0174973476678133, + 0.015072900801897049, + 0.013140987604856491, + 0.022708553820848465, + 0.0038078452926129103, + 0.006268542725592852, + 0.003838757984340191, + 0.003460118779912591, + 0.004549859557300806, + 0.002986677922308445, + 0.004216284491121769, + 0.006568303797394037, + 0.007982112467288971, + 0.009968060068786144, + 0.006523259449750185, + 0.006467614322900772, + 0.010371584445238113, + 0.010119878686964512, + 0.009994938038289547, + 0.010522063821554184, + 0.005526241380721331, + 0.009506621398031712, + 0.003396059386432171, + 0.005939904134720564, + 0.012025999836623669, + 0.008655029349029064, + 0.005503128282725811, + 0.004685100633651018, + 0.0062662819400429726, + 0.014105478301644325, + 0.002933658193796873, + 0.0048669795505702496, + 0.011978795751929283, + 0.020848572254180908, + 0.004410686437040567, + 0.0057821269147098064, + 0.02828880213201046, + 0.027879411354660988, + 0.003390902653336525, + 0.0040666391141712666, + 0.044719990342855453, + 0.0950300320982933, + 0.0024157529696822166, + 0.00677589001134038, + 0.008672820404171944, + 0.008827230893075466, + 0.004183327779173851, + 0.00495540164411068, + 0.00733323534950614, + 0.005150409881025553, + 0.005380564834922552, + 0.0047836871817708015, + 0.015381485223770142, + 0.018061084672808647, + 0.007501222658902407, + 0.01167747750878334, + 0.008070964366197586, + 0.009363478980958462, + 0.008619234897196293, + 0.006719992496073246, + 0.02396056428551674, + 0.01709415577352047, + 0.0057039749808609486, + 0.004551311489194632, + 0.0588526576757431, + 0.050666771829128265, + 0.006921142339706421, + 0.010686527006328106, + 0.019573112949728966, + 0.027350394055247307, + 0.00885268859565258, + 0.010202423669397831, + 0.012851405888795853, + 0.011275636032223701, + 0.010674998164176941, + 0.010766328312456608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.07794628292322159 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368139311671257 + ], + "opposite_logprob_tokens": [ + -3.6513681411743164 + ], + "logprob_reward": -0.026368139311671257, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368139311671257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0485067144036293, + 0.03656243905425072, + 0.035484280437231064, + 0.07048138976097107, + 0.03438308462500572, + 0.5316949486732483, + 0.04250122979283333, + 0.059226132929325104, + 0.08401693403720856, + 0.05714287981390953 + ], + "credit_map": [ + 0.005369146820157766, + 0.008839000016450882, + 0.007317961193621159, + 0.005052719730883837, + 0.006224407348781824, + 0.006112454459071159, + 0.004385102074593306, + 0.005205919966101646, + 0.003316373797133565, + 0.005782334133982658, + 0.0035775427240878344, + 0.003164776135236025, + 0.004207313526421785, + 0.005052512511610985, + 0.003816190641373396, + 0.007645397912710905, + 0.0038318138103932142, + 0.007346518337726593, + 0.003263107966631651, + 0.0031592831946909428, + 0.005738527979701757, + 0.004662222228944302, + 0.003078001318499446, + 0.004404804203659296, + 0.006512920372188091, + 0.009918845258653164, + 0.004636894445866346, + 0.00894935056567192, + 0.01321582030504942, + 0.010582121089100838, + 0.009460791945457458, + 0.007204640656709671, + 0.004065563436597586, + 0.006403032224625349, + 0.002350933151319623, + 0.00269595580175519, + 0.007206939626485109, + 0.005291510373353958, + 0.0023476143833249807, + 0.004021536558866501, + 0.09406597167253494, + 0.036551009863615036, + 0.011545505374670029, + 0.009765710681676865, + 0.12198285758495331, + 0.2354981154203415, + 0.005609432235360146, + 0.016676342114806175, + 0.007455227896571159, + 0.0078010899014770985, + 0.004053727723658085, + 0.00445160735398531, + 0.005512911826372147, + 0.004783520475029945, + 0.004149369429796934, + 0.004293772857636213, + 0.010022841393947601, + 0.010927172377705574, + 0.0052469875663518906, + 0.008120945654809475, + 0.00582902180030942, + 0.007925933226943016, + 0.005917376838624477, + 0.005235854536294937, + 0.010957657359540462, + 0.008355195634067059, + 0.0035264918114989996, + 0.0031769494526088238, + 0.024676326662302017, + 0.02265101671218872, + 0.0041747442446649075, + 0.006498553790152073, + 0.011808129027485847, + 0.013188228942453861, + 0.004927018191665411, + 0.004780181217938662, + 0.006306605879217386, + 0.0059251259081065655, + 0.00578186335042119, + 0.00442572683095932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.026368139311671257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009267802815884352 + ], + "opposite_logprob_tokens": [ + -7.000926971435547 + ], + "logprob_reward": -0.0009267802815884352, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009267802815884352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08095191419124603, + 0.04914585128426552, + 0.06020092964172363, + 0.094211146235466, + 0.04071245715022087, + 0.3656132221221924, + 0.1599482148885727, + 0.038822878152132034, + 0.06175768002867699, + 0.048635710030794144 + ], + "credit_map": [ + 0.00963449664413929, + 0.012375439517199993, + 0.007611168082803488, + 0.00802314467728138, + 0.01268870010972023, + 0.011308332905173302, + 0.007620467338711023, + 0.011690170504152775, + 0.005338959861546755, + 0.006577842403203249, + 0.005818344186991453, + 0.005045879166573286, + 0.007403920870274305, + 0.00555087486281991, + 0.005037551745772362, + 0.008372481912374496, + 0.00789171177893877, + 0.008140703663229942, + 0.0070482115261256695, + 0.004719721153378487, + 0.010755057446658611, + 0.009533424861729145, + 0.005728317424654961, + 0.006383779924362898, + 0.010308471508324146, + 0.009695417247712612, + 0.007839415222406387, + 0.008124985732138157, + 0.016301797702908516, + 0.014179900288581848, + 0.0164941418915987, + 0.011267012916505337, + 0.0067319911904633045, + 0.0058665163815021515, + 0.0036597575526684523, + 0.0029208229389041662, + 0.008227490819990635, + 0.006178103853017092, + 0.0030785140115767717, + 0.004049262031912804, + 0.05005403235554695, + 0.027778703719377518, + 0.008864493109285831, + 0.009185495786368847, + 0.09015686810016632, + 0.15919168293476105, + 0.007319759577512741, + 0.013062184676527977, + 0.02837703563272953, + 0.03974340483546257, + 0.012264562770724297, + 0.012677937746047974, + 0.022189775481820107, + 0.01682669296860695, + 0.013093341141939163, + 0.014775464311242104, + 0.007029831409454346, + 0.008828635327517986, + 0.0026227389462292194, + 0.004000319633632898, + 0.004796632099896669, + 0.004716291092336178, + 0.003449632553383708, + 0.003378793131560087, + 0.00947591196745634, + 0.01300672348588705, + 0.003897633869200945, + 0.003908939193934202, + 0.012784660793840885, + 0.009123336523771286, + 0.004908275790512562, + 0.004652196541428566, + 0.00799359567463398, + 0.011077317409217358, + 0.003910633735358715, + 0.0040579731576144695, + 0.006194072309881449, + 0.0038771021645516157, + 0.005529987160116434, + 0.005995024926960468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 3 + ], + "selected_reward": -0.0009267802815884352 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588846877217293 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588846877217293, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588846877217293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029779484495520592, + 0.01988101378083229, + 0.015036545693874359, + 0.02218090370297432, + 0.02399301715195179, + 0.7024213075637817, + 0.06376797705888748, + 0.028583578765392303, + 0.08260424435138702, + 0.011751888319849968 + ], + "credit_map": [ + 0.0035504335537552834, + 0.0030707514379173517, + 0.002780323615297675, + 0.0032249928917735815, + 0.0043500373139977455, + 0.0037104589864611626, + 0.0029729900415986776, + 0.006119496654719114, + 0.0021552552934736013, + 0.0024370828177779913, + 0.0021364411804825068, + 0.0016860193572938442, + 0.002635172102600336, + 0.002416855189949274, + 0.0020372308790683746, + 0.004376955796033144, + 0.0015531243989244103, + 0.0023843913804739714, + 0.0012876354157924652, + 0.0013962613884359598, + 0.0025538287591189146, + 0.0024501176085323095, + 0.0011713386047631502, + 0.0022398482542485, + 0.002494228072464466, + 0.002432507462799549, + 0.0017792783910408616, + 0.0021400086116045713, + 0.003968502394855022, + 0.003799708327278495, + 0.002727256389334798, + 0.0028394125401973724, + 0.002762844553217292, + 0.003379488131031394, + 0.0016872407868504524, + 0.0020165725145488977, + 0.004171604756265879, + 0.0036811064928770065, + 0.0022675609216094017, + 0.004026600159704685, + 0.10451287031173706, + 0.05524764582514763, + 0.012079416774213314, + 0.011834928765892982, + 0.17599020898342133, + 0.31519433856010437, + 0.006871511228382587, + 0.02069041319191456, + 0.01129752118140459, + 0.016528908163309097, + 0.004479838069528341, + 0.005012608133256435, + 0.00882875919342041, + 0.00622155424207449, + 0.005440209526568651, + 0.00595858134329319, + 0.004695456475019455, + 0.005580603610724211, + 0.002041169675067067, + 0.0025023012422025204, + 0.003408068558201194, + 0.005756264086812735, + 0.0019378848373889923, + 0.0026618288829922676, + 0.011294892057776451, + 0.012948606163263321, + 0.0032332430128008127, + 0.002939467318356037, + 0.020527057349681854, + 0.020517824217677116, + 0.004854343365877867, + 0.0062888143584132195, + 0.0019189087906852365, + 0.0027750262524932623, + 0.000689511769451201, + 0.0007817758596502244, + 0.0017056441865861416, + 0.0015543494373559952, + 0.001145816408097744, + 0.0011808553244918585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.020588846877217293 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038096725940704346 + ], + "opposite_logprob_tokens": [ + -3.2880966663360596 + ], + "logprob_reward": -0.038096725940704346, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038096725940704346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05480522662401199, + 0.01937587559223175, + 0.0262648556381464, + 0.031993668526411057, + 0.045815881341695786, + 0.5567340850830078, + 0.19437484443187714, + 0.023547038435935974, + 0.030547620728611946, + 0.016540873795747757 + ], + "credit_map": [ + 0.0067312889732420444, + 0.006976559292525053, + 0.005656159482896328, + 0.00643186317756772, + 0.008307872340083122, + 0.006420346908271313, + 0.005665356293320656, + 0.008615780621767044, + 0.0021507435012608767, + 0.002675931202247739, + 0.0021409785840660334, + 0.0017744660144671798, + 0.0024555264972150326, + 0.002101837657392025, + 0.002262260066345334, + 0.0038141319528222084, + 0.0030260931234806776, + 0.004002501256763935, + 0.0020680699963122606, + 0.0022461169864982367, + 0.004779592622071505, + 0.004228143021464348, + 0.002380102640017867, + 0.0035342348273843527, + 0.00396328279748559, + 0.0035258126445114613, + 0.0019171845633536577, + 0.0029483623802661896, + 0.005450744181871414, + 0.00532001256942749, + 0.004042956978082657, + 0.004825315438210964, + 0.006672645453363657, + 0.006811853963881731, + 0.002402418525889516, + 0.002932595321908593, + 0.008806657046079636, + 0.008942818269133568, + 0.004367835354059935, + 0.0048790546134114265, + 0.08762022107839584, + 0.04127906262874603, + 0.010597464628517628, + 0.010642722249031067, + 0.13434752821922302, + 0.24730464816093445, + 0.006516120862215757, + 0.01842629536986351, + 0.03319074213504791, + 0.04076874628663063, + 0.016878634691238403, + 0.017553986981511116, + 0.0270689707249403, + 0.020111549645662308, + 0.01950996369123459, + 0.019292257726192474, + 0.003986321855336428, + 0.005091884173452854, + 0.001870272564701736, + 0.0026707255747169256, + 0.0028056984301656485, + 0.0033885319717228413, + 0.0019182958640158176, + 0.0018153074197471142, + 0.004097784869372845, + 0.004093398340046406, + 0.0013100184733048081, + 0.0010726230684667826, + 0.008563177660107613, + 0.007864246144890785, + 0.001408047042787075, + 0.0021383247803896666, + 0.00267843552865088, + 0.0035087172873318195, + 0.0014920112444087863, + 0.0015959369484335184, + 0.002039519837126136, + 0.0014533784706145525, + 0.002064751461148262, + 0.001708123367279768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.038096725940704346 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot with slight handheld camera shake, tilted slightly upward and panning slowly to the right, captures an incredibly surreal scene: a massive B-17 Flying Fortress bomber, its weathered silver fuselage reflecting a dull luster in the soft light, its rivets and minute battle scars faintly visible, its four massive propeller engines at rest. With a grace and composure that defies all physics, the behemoth glides slowly over the water, its belly almost touching the still surface as it approaches an old wooden pier, leaving a trail of faint, lingering ripples in its wake. The entire scene is bathed in the soft tones of dawn or dusk, the sky a dreamy gradation of lavender and pink-orange. This light evenly casts light across the aircraft, the water, and the pier, creating a tranquil, dreamy, and slightly unsettling atmosphere. As the plane approaches at an impossibly slow and steady pace, the camera shakes slightly, as if the observer is holding their breath at the miraculous sight before them, further enhancing the unique quality of the image, which is both unreal and extremely real.", + "seed": 14020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_3487_seed_14020_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": -0.03764726142399013, + "vlm_logprob": -0.03764726142399013, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03764726142399013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0292247012257576 + ], + "opposite_logprob_tokens": [ + -3.9042246341705322 + ], + "logprob_reward": -0.0292247012257576, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0292247012257576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2960580885410309, + 0.210645392537117, + 0.05768429860472679, + 0.16430038213729858, + 0.08664149791002274, + 0.02821306698024273, + 0.03124511055648327, + 0.032557375729084015, + 0.054357655346393585, + 0.03829715773463249 + ], + "credit_map": [ + 0.03665480390191078, + 0.05253016576170921, + 0.021097032353281975, + 0.024435408413410187, + 0.04017051309347153, + 0.03485099598765373, + 0.028516842052340508, + 0.057802338153123856, + 0.013098814524710178, + 0.011094605550169945, + 0.010315191000699997, + 0.0122152678668499, + 0.07302021235227585, + 0.013314137235283852, + 0.014614621177315712, + 0.0629725530743599, + 0.0078955739736557, + 0.004858965519815683, + 0.00682771485298872, + 0.007408264093101025, + 0.006845436990261078, + 0.005043471697717905, + 0.010537159629166126, + 0.008267710916697979, + 0.01910872757434845, + 0.015393857844173908, + 0.01710101217031479, + 0.020869798958301544, + 0.01928519643843174, + 0.014472940936684608, + 0.02542184293270111, + 0.03264700993895531, + 0.007023037411272526, + 0.009676171466708183, + 0.011957701295614243, + 0.011562906205654144, + 0.011740132234990597, + 0.012510523200035095, + 0.011123230680823326, + 0.011047792620956898, + 0.003515685210004449, + 0.00589800113812089, + 0.003232465358451009, + 0.0023016699124127626, + 0.002861693734303117, + 0.002596552250906825, + 0.0035012238658964634, + 0.00430577527731657, + 0.003720460459589958, + 0.004045512061566114, + 0.0018119310261681676, + 0.0019098893972113729, + 0.002758193761110306, + 0.005777124781161547, + 0.002576803555712104, + 0.008645192719995975, + 0.0031125068198889494, + 0.003184250323101878, + 0.003455694764852524, + 0.0026033122558146715, + 0.005503465421497822, + 0.00387284136377275, + 0.004101112484931946, + 0.006724188569933176, + 0.002506792778149247, + 0.005376735702157021, + 0.003320031799376011, + 0.002372128190472722, + 0.015350298956036568, + 0.014735950157046318, + 0.004540672525763512, + 0.006155042443424463, + 0.0031951996497809887, + 0.0051698582246899605, + 0.0034106606617569923, + 0.004363815765827894, + 0.006327807903289795, + 0.006159198470413685, + 0.005213520023971796, + 0.004457096103578806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0292247012257576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059386856853962 + ], + "opposite_logprob_tokens": [ + -3.8955938816070557 + ], + "logprob_reward": -0.02059386856853962, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059386856853962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2142871916294098, + 0.06355318427085876, + 0.0424993522465229, + 0.03300287947058678, + 0.04434766620397568, + 0.037243884056806564, + 0.1691226065158844, + 0.2519652545452118, + 0.04877099767327309, + 0.09520695358514786 + ], + "credit_map": [ + 0.01624154858291149, + 0.032602593302726746, + 0.009461102075874805, + 0.01456502452492714, + 0.05050552636384964, + 0.017282593995332718, + 0.018263207748532295, + 0.055365607142448425, + 0.004538073670119047, + 0.005022515542805195, + 0.003908913116902113, + 0.003984789829701185, + 0.019398557022213936, + 0.006284982431679964, + 0.00435620779171586, + 0.01605914905667305, + 0.0040990267880260944, + 0.0037206867709755898, + 0.003765166038647294, + 0.005036736372858286, + 0.007156325038522482, + 0.004266898613423109, + 0.007276060990989208, + 0.007178453728556633, + 0.0035512461327016354, + 0.0021493472158908844, + 0.0034690452739596367, + 0.003342021256685257, + 0.005182599648833275, + 0.002930589020252228, + 0.00420755660161376, + 0.008170478977262974, + 0.003039297182112932, + 0.0054732454009354115, + 0.005014454480260611, + 0.0037904800847172737, + 0.0069955200888216496, + 0.0050925156101584435, + 0.007864906452596188, + 0.0070772492326796055, + 0.0028912308625876904, + 0.0036001447588205338, + 0.004208442289382219, + 0.004644094966351986, + 0.004144671838730574, + 0.005858740769326687, + 0.005237926729023457, + 0.006658635567873716, + 0.01738348975777626, + 0.02352960966527462, + 0.013017484918236732, + 0.01654091663658619, + 0.01553492434322834, + 0.02387882024049759, + 0.01740337908267975, + 0.041833993047475815, + 0.030778734013438225, + 0.02744642272591591, + 0.026537707075476646, + 0.017159823328256607, + 0.026782816275954247, + 0.03892473131418228, + 0.028877941891551018, + 0.05545710027217865, + 0.0031908471137285233, + 0.005996381398290396, + 0.0031404937617480755, + 0.004538073670119047, + 0.012668212875723839, + 0.007311119232326746, + 0.005398387089371681, + 0.00652748579159379, + 0.010106345638632774, + 0.019697554409503937, + 0.011124643497169018, + 0.012649753130972385, + 0.009234244003891945, + 0.006785113364458084, + 0.01522243581712246, + 0.010386873036623001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.02059386856853962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022033003624528646 + ], + "opposite_logprob_tokens": [ + -6.127203464508057 + ], + "logprob_reward": -0.0022033003624528646, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022033003624528646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14882409572601318, + 0.08050563931465149, + 0.10752982646226883, + 0.16000394523143768, + 0.06038670241832733, + 0.0441683791577816, + 0.11197581887245178, + 0.09666513651609421, + 0.08281190693378448, + 0.1071285530924797 + ], + "credit_map": [ + 0.015300026163458824, + 0.0240870900452137, + 0.006185782141983509, + 0.009555156342685223, + 0.030283771455287933, + 0.014494504779577255, + 0.013538530096411705, + 0.035379234701395035, + 0.00506079476326704, + 0.005255918018519878, + 0.00473086116835475, + 0.005800584331154823, + 0.02287253923714161, + 0.009146532975137234, + 0.007177440449595451, + 0.020460957661271095, + 0.010858912020921707, + 0.012878364883363247, + 0.006910915020853281, + 0.008175772614777088, + 0.02155013009905815, + 0.009579477831721306, + 0.010755001567304134, + 0.02682124637067318, + 0.017672698944807053, + 0.014875609427690506, + 0.01761452667415142, + 0.020867284387350082, + 0.018241625279188156, + 0.01259070448577404, + 0.024601811543107033, + 0.033539675176143646, + 0.005020532291382551, + 0.006906257942318916, + 0.006582970265299082, + 0.0059440284967422485, + 0.011663543060421944, + 0.00689446646720171, + 0.006578132510185242, + 0.010796766728162766, + 0.003188190283253789, + 0.006228197831660509, + 0.003681646194308996, + 0.0038252228405326605, + 0.008421966806054115, + 0.00668200571089983, + 0.004777516238391399, + 0.0073636313900351524, + 0.011420046910643578, + 0.015041641891002655, + 0.008147135376930237, + 0.010198485106229782, + 0.010117622092366219, + 0.01706441305577755, + 0.011243723332881927, + 0.02874274179339409, + 0.010304387658834457, + 0.010211052373051643, + 0.009403221309185028, + 0.007510903291404247, + 0.010856433771550655, + 0.016296228393912315, + 0.010450495406985283, + 0.02163241058588028, + 0.007205723784863949, + 0.009216484613716602, + 0.004160879645496607, + 0.004912927746772766, + 0.022030742838978767, + 0.012922336347401142, + 0.00685703894123435, + 0.015505765564739704, + 0.01142390538007021, + 0.02373191900551319, + 0.011230087839066982, + 0.011917145922780037, + 0.013345377519726753, + 0.008979365229606628, + 0.014993767254054546, + 0.011506977491080761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.0022033003624528646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291567340493202 + ], + "opposite_logprob_tokens": [ + -3.77329158782959 + ], + "logprob_reward": -0.023291567340493202, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291567340493202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12135085463523865, + 0.2787800431251526, + 0.07535817474126816, + 0.11612856388092041, + 0.03637997433543205, + 0.034995995461940765, + 0.08838546276092529, + 0.05706274136900902, + 0.1186872124671936, + 0.07287092506885529 + ], + "credit_map": [ + 0.01860932819545269, + 0.020212914794683456, + 0.008572731167078018, + 0.008496473543345928, + 0.017863096669316292, + 0.013183262199163437, + 0.010216793976724148, + 0.024196265265345573, + 0.013730347156524658, + 0.013909927569329739, + 0.008877593092620373, + 0.010505077429115772, + 0.1052325889468193, + 0.01910705864429474, + 0.01487745437771082, + 0.09254001826047897, + 0.007182201836258173, + 0.008918870240449905, + 0.008075255900621414, + 0.00932434480637312, + 0.011403486132621765, + 0.007442060858011246, + 0.010052190162241459, + 0.012959767132997513, + 0.014887067489326, + 0.010984295979142189, + 0.012572824023663998, + 0.014130362309515476, + 0.012688669376075268, + 0.008907968178391457, + 0.018392065539956093, + 0.02356531098484993, + 0.00323260435834527, + 0.0038400960620492697, + 0.006179300136864185, + 0.003804936073720455, + 0.005324368830770254, + 0.004656380508095026, + 0.004672599490731955, + 0.0046696895733475685, + 0.0030153514817357063, + 0.004856681916862726, + 0.003271237015724182, + 0.0033030512277036905, + 0.004927393049001694, + 0.005477333907037973, + 0.00491036893799901, + 0.005234580021351576, + 0.008312714286148548, + 0.012435286305844784, + 0.007143442984670401, + 0.008287062868475914, + 0.010578920133411884, + 0.011461959220468998, + 0.009483277797698975, + 0.020682798698544502, + 0.0053273215889930725, + 0.00607706792652607, + 0.007057045120745897, + 0.00671985000371933, + 0.009416768327355385, + 0.0058748614974319935, + 0.007252804934978485, + 0.00933702103793621, + 0.006901979446411133, + 0.013262809254229069, + 0.0037510786205530167, + 0.003998969681560993, + 0.039089009165763855, + 0.02896242029964924, + 0.008863170631229877, + 0.013857776299118996, + 0.005803798325359821, + 0.011948591098189354, + 0.00969388522207737, + 0.01115332543849945, + 0.007360499352216721, + 0.0053258673287928104, + 0.01366156805306673, + 0.007923386991024017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.023291567340493202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292286962270737 + ], + "opposite_logprob_tokens": [ + -2.2379229068756104 + ], + "logprob_reward": -0.11292286962270737, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292286962270737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08013337105512619, + 0.06564478576183319, + 0.07417795807123184, + 0.17175589501857758, + 0.033290114253759384, + 0.024490797892212868, + 0.20591594278812408, + 0.1775856614112854, + 0.0294792540371418, + 0.1375262588262558 + ], + "credit_map": [ + 0.009122059680521488, + 0.0060615441761910915, + 0.005215814802795649, + 0.005570647772401571, + 0.022972555831074715, + 0.007292288355529308, + 0.006637754384428263, + 0.017260706052184105, + 0.006329003721475601, + 0.006031711585819721, + 0.004567364696413279, + 0.007894099690020084, + 0.01291665155440569, + 0.0077875275164842606, + 0.006849568337202072, + 0.013268858194351196, + 0.00837712548673153, + 0.005909898784011602, + 0.008753953501582146, + 0.01131056621670723, + 0.008189625106751919, + 0.007358760107308626, + 0.012611768208444118, + 0.011666256003081799, + 0.02032107301056385, + 0.016669098287820816, + 0.019154906272888184, + 0.02341703325510025, + 0.01803719252347946, + 0.012577999383211136, + 0.027600642293691635, + 0.03397795557975769, + 0.0024745918344706297, + 0.003837918397039175, + 0.004697096534073353, + 0.004669124726206064, + 0.003796475240960717, + 0.003164719557389617, + 0.004557326436042786, + 0.006092861760407686, + 0.0015350095927715302, + 0.0033944433089345694, + 0.004038146696984768, + 0.002373937051743269, + 0.0028645950369536877, + 0.003619497874751687, + 0.002923323307186365, + 0.0037418457213789225, + 0.020961446687579155, + 0.027111902832984924, + 0.016190698370337486, + 0.018248623237013817, + 0.020239854231476784, + 0.029394883662462234, + 0.022071631625294685, + 0.05169690027832985, + 0.02121342532336712, + 0.019814716652035713, + 0.020200781524181366, + 0.014449352398514748, + 0.020553570240736008, + 0.023227520287036896, + 0.0207320936024189, + 0.0373942032456398, + 0.0024373240303248167, + 0.0033988950308412313, + 0.004512149374932051, + 0.004725163336843252, + 0.0025297915562987328, + 0.003013175679370761, + 0.004880430642515421, + 0.00398232601583004, + 0.012977642007172108, + 0.02397574484348297, + 0.018155008554458618, + 0.018402326852083206, + 0.014806739054620266, + 0.009895900264382362, + 0.024670807644724846, + 0.014642099849879742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.11292286962270737 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot with slight handheld camera shake, tilted slightly upward and panning slowly to the right, captures an incredibly surreal scene: a massive B-17 Flying Fortress bomber, its weathered silver fuselage reflecting a dull luster in the soft light, its rivets and minute battle scars faintly visible, its four massive propeller engines at rest. With a grace and composure that defies all physics, the behemoth glides slowly over the water, its belly almost touching the still surface as it approaches an old wooden pier, leaving a trail of faint, lingering ripples in its wake. The entire scene is bathed in the soft tones of dawn or dusk, the sky a dreamy gradation of lavender and pink-orange. This light evenly casts light across the aircraft, the water, and the pier, creating a tranquil, dreamy, and slightly unsettling atmosphere. As the plane approaches at an impossibly slow and steady pace, the camera shakes slightly, as if the observer is holding their breath at the miraculous sight before them, further enhancing the unique quality of the image, which is both unreal and extremely real.", + "seed": 14021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_3487_seed_14021_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": -0.021007495280355216, + "vlm_logprob": -0.021007495280355216, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.021007495280355216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00729279313236475 + ], + "opposite_logprob_tokens": [ + -5.757292747497559 + ], + "logprob_reward": -0.00729279313236475, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00729279313236475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553764045238495, + 0.10753133893013, + 0.16339609026908875, + 0.08532669395208359, + 0.0881306529045105, + 0.07097561657428741, + 0.06519807875156403, + 0.06882720440626144, + 0.06574104726314545, + 0.12949688732624054 + ], + "credit_map": [ + 0.028229542076587677, + 0.020073922351002693, + 0.01957852765917778, + 0.014227787032723427, + 0.02037244290113449, + 0.021867040544748306, + 0.011345149017870426, + 0.01968199945986271, + 0.011009477078914642, + 0.013451031409204006, + 0.007723980117589235, + 0.018411710858345032, + 0.01398832444101572, + 0.012047250755131245, + 0.01967335306107998, + 0.011226197704672813, + 0.015411613509058952, + 0.02660209685564041, + 0.00975976325571537, + 0.022588249295949936, + 0.024099057540297508, + 0.02084006369113922, + 0.02312660589814186, + 0.020968634635210037, + 0.007316259667277336, + 0.01041246484965086, + 0.006862550042569637, + 0.013665112666785717, + 0.010453996248543262, + 0.009695683605968952, + 0.01727270521223545, + 0.009647922590374947, + 0.009346507489681244, + 0.011475341394543648, + 0.008887771517038345, + 0.012745749205350876, + 0.00804924126714468, + 0.00966203399002552, + 0.012753188610076904, + 0.015210820361971855, + 0.0069353096187114716, + 0.006569983903318644, + 0.009500070475041866, + 0.011516866274178028, + 0.009341527707874775, + 0.008923827670514584, + 0.00953206978738308, + 0.00865596067160368, + 0.007188904564827681, + 0.007388758473098278, + 0.008131119422614574, + 0.009903932921588421, + 0.007760416716337204, + 0.008820331655442715, + 0.009241164661943913, + 0.006763446610420942, + 0.009798000566661358, + 0.01177577581256628, + 0.00823722779750824, + 0.007947150617837906, + 0.007937083952128887, + 0.008894005790352821, + 0.008759569376707077, + 0.0054783932864665985, + 0.006406455300748348, + 0.010055972263216972, + 0.007695366628468037, + 0.012049911543726921, + 0.0063440087251365185, + 0.008584152907133102, + 0.0074126748368144035, + 0.007192507851868868, + 0.016535978764295578, + 0.0136784752830863, + 0.014668616466224194, + 0.012913553975522518, + 0.017553700134158134, + 0.014102396555244923, + 0.022654753178358078, + 0.017389418557286263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.00729279313236475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420564204454422 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.01420564204454422, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420564204454422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1987191140651703, + 0.08577381819486618, + 0.055651694536209106, + 0.07094623148441315, + 0.0352892205119133, + 0.12220565229654312, + 0.026275843381881714, + 0.06673093140125275, + 0.237567737698555, + 0.10083978623151779 + ], + "credit_map": [ + 0.01969200186431408, + 0.020767368376255035, + 0.018493512645363808, + 0.025216488167643547, + 0.035769034177064896, + 0.025458330288529396, + 0.024015216156840324, + 0.029307162389159203, + 0.008387415669858456, + 0.010478651151061058, + 0.006494584027677774, + 0.014530126936733723, + 0.013756374828517437, + 0.010076357051730156, + 0.014174984768033028, + 0.007875326089560986, + 0.004581240005791187, + 0.006027798168361187, + 0.005845627747476101, + 0.009974319487810135, + 0.008020222187042236, + 0.005989001132547855, + 0.008256575092673302, + 0.006956905126571655, + 0.0071828351356089115, + 0.009876817464828491, + 0.005555328913033009, + 0.012579375877976418, + 0.006334323901683092, + 0.008603204041719437, + 0.01200234517455101, + 0.00881199724972248, + 0.0033186287619173527, + 0.003934844862669706, + 0.0038903490640223026, + 0.006211928091943264, + 0.0031808787025511265, + 0.003549061017110944, + 0.0063191005028784275, + 0.004884432069957256, + 0.015521968714892864, + 0.012117577716708183, + 0.01762276329100132, + 0.02569364570081234, + 0.01406943891197443, + 0.015270519070327282, + 0.013948422856628895, + 0.007961311377584934, + 0.0031101962085813284, + 0.0029284965712577105, + 0.0035641025751829147, + 0.004056830890476704, + 0.0031427983194589615, + 0.003229657420888543, + 0.003868737490847707, + 0.002375025302171707, + 0.008996818214654922, + 0.009148919954895973, + 0.009526785463094711, + 0.008673429489135742, + 0.00692092115059495, + 0.008344489149749279, + 0.008622255176305771, + 0.006497311405837536, + 0.02824922278523445, + 0.03204090893268585, + 0.03323567286133766, + 0.0421573705971241, + 0.023554295301437378, + 0.021612344309687614, + 0.02380702644586563, + 0.03291088342666626, + 0.010626065544784069, + 0.01061991136521101, + 0.012111243791878223, + 0.00890384428203106, + 0.013331543654203415, + 0.01180325262248516, + 0.01919536478817463, + 0.014248562976717949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.01420564204454422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000922969076782465 + ], + "opposite_logprob_tokens": [ + -7.000923156738281 + ], + "logprob_reward": -0.000922969076782465, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000922969076782465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22644682228565216, + 0.07982978969812393, + 0.09149392694234848, + 0.0792553573846817, + 0.037211306393146515, + 0.0931820347905159, + 0.06871869415044785, + 0.07432349771261215, + 0.11520203948020935, + 0.13433648645877838 + ], + "credit_map": [ + 0.040066223591566086, + 0.02833828330039978, + 0.01840479113161564, + 0.03328704088926315, + 0.03441798686981201, + 0.028394894674420357, + 0.023755764588713646, + 0.019781850278377533, + 0.011669948697090149, + 0.012437170371413231, + 0.006573250517249107, + 0.010296070016920567, + 0.010125464759767056, + 0.010047978721559048, + 0.00972254853695631, + 0.008957360871136189, + 0.009809820912778378, + 0.01144428551197052, + 0.007615396287292242, + 0.015390963293612003, + 0.014067638665437698, + 0.010601901449263096, + 0.011931370943784714, + 0.010632556863129139, + 0.011337963864207268, + 0.01055155135691166, + 0.00733593525364995, + 0.01158037967979908, + 0.007578625809401274, + 0.010702691040933132, + 0.009809155017137527, + 0.01035905722528696, + 0.007056051399558783, + 0.0041953204199671745, + 0.004643172491341829, + 0.0043815611861646175, + 0.0037824444007128477, + 0.004792529158294201, + 0.005100626964122057, + 0.003259604098275304, + 0.014516309835016727, + 0.009245118126273155, + 0.012536276131868362, + 0.011990846134722233, + 0.008923675864934921, + 0.011077199131250381, + 0.011249871924519539, + 0.013642744161188602, + 0.015463901683688164, + 0.008281505666673183, + 0.008381422609090805, + 0.006112037226557732, + 0.009698507376015186, + 0.010464940220117569, + 0.006422269158065319, + 0.003894116962328553, + 0.007548623718321323, + 0.0069370330311357975, + 0.00936649926006794, + 0.013762111775577068, + 0.00799410231411457, + 0.010493610985577106, + 0.010288495570421219, + 0.00793302059173584, + 0.012569437734782696, + 0.013599585741758347, + 0.015984077006578445, + 0.018825802952051163, + 0.013653047382831573, + 0.016375310719013214, + 0.010874390602111816, + 0.013320397585630417, + 0.017335372045636177, + 0.018511662259697914, + 0.015259992331266403, + 0.011957275681197643, + 0.01933293603360653, + 0.01533246599137783, + 0.020233571529388428, + 0.016373220831155777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.000922969076782465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580671727657318 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.020580671727657318, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580671727657318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10724450647830963, + 0.055133286863565445, + 0.08873002976179123, + 0.03938937932252884, + 0.08387155830860138, + 0.11267944425344467, + 0.12133724242448807, + 0.0540083572268486, + 0.26565825939178467, + 0.0719478651881218 + ], + "credit_map": [ + 0.020953860133886337, + 0.01448921486735344, + 0.011362002231180668, + 0.012322425842285156, + 0.011342453770339489, + 0.01644325815141201, + 0.010112603195011616, + 0.010218697600066662, + 0.008880485780537128, + 0.008400505408644676, + 0.006254552863538265, + 0.006374634802341461, + 0.006388540379703045, + 0.007537477649748325, + 0.005935106426477432, + 0.0053619868122041225, + 0.008216728456318378, + 0.012106138281524181, + 0.0055280583910644054, + 0.011477038264274597, + 0.014560154639184475, + 0.012006695382297039, + 0.010893182829022408, + 0.013942040503025055, + 0.004632858093827963, + 0.004398129880428314, + 0.003986418712884188, + 0.0067588696256279945, + 0.0038749135565012693, + 0.004585498012602329, + 0.005227097310125828, + 0.005925594829022884, + 0.008889826014637947, + 0.012292549014091492, + 0.008732336573302746, + 0.01195276714861393, + 0.010515121743083, + 0.00904141180217266, + 0.01214158907532692, + 0.010305963456630707, + 0.014634164050221443, + 0.009753232821822166, + 0.016612974926829338, + 0.022791016846895218, + 0.012243212200701237, + 0.013898728415369987, + 0.011593928560614586, + 0.011152204126119614, + 0.017151273787021637, + 0.01315244846045971, + 0.01816382259130478, + 0.01592269167304039, + 0.013723970390856266, + 0.014854952692985535, + 0.01646089181303978, + 0.011907191947102547, + 0.00574808893725276, + 0.008176540955901146, + 0.005971450824290514, + 0.009539189748466015, + 0.006147961597889662, + 0.007821934297680855, + 0.006836194545030594, + 0.0037669993471354246, + 0.03287183865904808, + 0.03247876837849617, + 0.03782617673277855, + 0.03877709433436394, + 0.030724722892045975, + 0.03535118326544762, + 0.02511727623641491, + 0.032511234283447266, + 0.0077256555669009686, + 0.007546493783593178, + 0.010343746282160282, + 0.00801777932792902, + 0.009466130286455154, + 0.008042716421186924, + 0.011417973786592484, + 0.009387376718223095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.020580671727657318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203540042042732 + ], + "opposite_logprob_tokens": [ + -2.812035322189331 + ], + "logprob_reward": -0.06203540042042732, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203540042042732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2629333734512329, + 0.05383777990937233, + 0.10998047888278961, + 0.02940344624221325, + 0.13905951380729675, + 0.08606534451246262, + 0.052540816366672516, + 0.027874905616044998, + 0.06893592327833176, + 0.1693684309720993 + ], + "credit_map": [ + 0.03394930064678192, + 0.039842743426561356, + 0.017757827416062355, + 0.031232981011271477, + 0.030006496235728264, + 0.03704698756337166, + 0.02421197108924389, + 0.048885077238082886, + 0.005447837058454752, + 0.005972850136458874, + 0.00337642221711576, + 0.010442584753036499, + 0.005156851373612881, + 0.006121325772255659, + 0.010809722356498241, + 0.006510185543447733, + 0.011855288408696651, + 0.010776538401842117, + 0.009632442146539688, + 0.015472260303795338, + 0.014871896244585514, + 0.014920287765562534, + 0.01666153222322464, + 0.01579023152589798, + 0.0033169726375490427, + 0.004781667143106461, + 0.002489238977432251, + 0.0036583070177584887, + 0.0033869771286845207, + 0.0036616225261241198, + 0.003859330900013447, + 0.004249330144375563, + 0.013999490067362785, + 0.020168405026197433, + 0.015156049281358719, + 0.02248154580593109, + 0.01569988578557968, + 0.014962776564061642, + 0.01719551347196102, + 0.019395848736166954, + 0.00849281158298254, + 0.011034490540623665, + 0.012362241744995117, + 0.010186242870986462, + 0.008916011080145836, + 0.009413321502506733, + 0.012029275298118591, + 0.013630948960781097, + 0.007375809829682112, + 0.006478906609117985, + 0.007502928841859102, + 0.005926409270614386, + 0.006956683471798897, + 0.006890190299600363, + 0.007068971637636423, + 0.004340916406363249, + 0.0028807297348976135, + 0.002985095838084817, + 0.004005084745585918, + 0.0033628474920988083, + 0.003411809215322137, + 0.0043117632158100605, + 0.004052674863487482, + 0.002864900976419449, + 0.007156963925808668, + 0.007994893938302994, + 0.008907873183488846, + 0.012335425242781639, + 0.00791782233864069, + 0.009361069649457932, + 0.006783457472920418, + 0.008478416129946709, + 0.016974296420812607, + 0.01618737354874611, + 0.020523814484477043, + 0.017647694796323776, + 0.0215052030980587, + 0.020118527114391327, + 0.034150879830121994, + 0.022260643541812897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.06203540042042732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot with slight handheld camera shake, tilted slightly upward and panning slowly to the right, captures an incredibly surreal scene: a massive B-17 Flying Fortress bomber, its weathered silver fuselage reflecting a dull luster in the soft light, its rivets and minute battle scars faintly visible, its four massive propeller engines at rest. With a grace and composure that defies all physics, the behemoth glides slowly over the water, its belly almost touching the still surface as it approaches an old wooden pier, leaving a trail of faint, lingering ripples in its wake. The entire scene is bathed in the soft tones of dawn or dusk, the sky a dreamy gradation of lavender and pink-orange. This light evenly casts light across the aircraft, the water, and the pier, creating a tranquil, dreamy, and slightly unsettling atmosphere. As the plane approaches at an impossibly slow and steady pace, the camera shakes slightly, as if the observer is holding their breath at the miraculous sight before them, further enhancing the unique quality of the image, which is both unreal and extremely real.", + "seed": 14022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_15/index_3487_seed_14022_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": -0.07934700772166252, + "vlm_logprob": -0.07934700772166252, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07934700772166252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2899720370769501 + ], + "opposite_logprob_tokens": [ + -1.4149720668792725 + ], + "logprob_reward": -0.2899720370769501, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2899720370769501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04875022917985916, + 0.3536053001880646, + 0.04422171786427498, + 0.017493685707449913, + 0.0646100640296936, + 0.0565776489675045, + 0.02326134592294693, + 0.023587625473737717, + 0.27874308824539185, + 0.08914922177791595 + ], + "credit_map": [ + 0.005808237940073013, + 0.010656281374394894, + 0.004635530058294535, + 0.0038330587558448315, + 0.007279477082192898, + 0.008426007814705372, + 0.004332172684371471, + 0.0037794632371515036, + 0.04731190204620361, + 0.04936593398451805, + 0.036173637956380844, + 0.038028933107852936, + 0.06035697087645531, + 0.0497274175286293, + 0.03217027336359024, + 0.04047023877501488, + 0.004360671155154705, + 0.005719385575503111, + 0.003771758871152997, + 0.00469921063631773, + 0.005484745837748051, + 0.006376868579536676, + 0.005358182359486818, + 0.008450896479189396, + 0.0015811704797670245, + 0.0025207209400832653, + 0.0009858947014436126, + 0.0010194758651778102, + 0.0040106927044689655, + 0.0036212990526109934, + 0.0010009686229750514, + 0.002753464039415121, + 0.008260502479970455, + 0.004727090243250132, + 0.023494411259889603, + 0.00503480713814497, + 0.00661073625087738, + 0.00624393206089735, + 0.004810801707208157, + 0.005427785217761993, + 0.007936823181807995, + 0.00812965165823698, + 0.006046933121979237, + 0.008773477748036385, + 0.007271775044500828, + 0.006586155854165554, + 0.00806179828941822, + 0.0037710340693593025, + 0.0033608274534344673, + 0.0036251393612474203, + 0.0028397426940500736, + 0.0040056137368083, + 0.00244430685415864, + 0.002310367301106453, + 0.002482699230313301, + 0.0021926481276750565, + 0.0031112930737435818, + 0.0026536702644079924, + 0.002326453337445855, + 0.004142923280596733, + 0.0029851049184799194, + 0.0027189154643565416, + 0.002673003589734435, + 0.0029762601479887962, + 0.03818875551223755, + 0.03856542706489563, + 0.035410117357969284, + 0.021076997742056847, + 0.03828590735793114, + 0.03424540162086487, + 0.03282753378152847, + 0.04014299064874649, + 0.009775572456419468, + 0.014923732727766037, + 0.007527051959186792, + 0.005139314103871584, + 0.023384517058730125, + 0.012066280469298363, + 0.009374580346047878, + 0.00695817032828927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.2899720370769501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981209009885788 + ], + "opposite_logprob_tokens": [ + -3.5298120975494385 + ], + "logprob_reward": -0.02981209009885788, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981209009885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1542927771806717, + 0.10809284448623657, + 0.21659885346889496, + 0.027997799217700958, + 0.022380318492650986, + 0.04622632637619972, + 0.05225913226604462, + 0.06552788615226746, + 0.1849460005760193, + 0.12167809903621674 + ], + "credit_map": [ + 0.018329713493585587, + 0.02566174976527691, + 0.01312981452792883, + 0.011069404892623425, + 0.019519563764333725, + 0.025301238521933556, + 0.01861480437219143, + 0.02266647480428219, + 0.019390588626265526, + 0.016591180115938187, + 0.009400843642652035, + 0.008923784829676151, + 0.017187798395752907, + 0.01895156502723694, + 0.007151476573199034, + 0.010495590046048164, + 0.0353078730404377, + 0.028665846213698387, + 0.019321512430906296, + 0.025552373379468918, + 0.026465000584721565, + 0.02629510499536991, + 0.021991707384586334, + 0.0329994261264801, + 0.0042719594202935696, + 0.0041130767203867435, + 0.001762215862981975, + 0.002750593237578869, + 0.005324922502040863, + 0.0045000831596553326, + 0.0023667425848543644, + 0.002908205147832632, + 0.003372439183294773, + 0.0020604580640792847, + 0.0028779993299394846, + 0.0029981492552906275, + 0.003403154667466879, + 0.002934122923761606, + 0.0021810869220644236, + 0.0025529039558023214, + 0.006453430745750666, + 0.00504496181383729, + 0.005639487411826849, + 0.006835622247308493, + 0.0056242384016513824, + 0.004455315414816141, + 0.008604656904935837, + 0.003568608546629548, + 0.007485507056117058, + 0.01147784385830164, + 0.0074861194007098675, + 0.006438684184104204, + 0.006767718121409416, + 0.003640577895566821, + 0.004234691150486469, + 0.004727987572550774, + 0.009488176554441452, + 0.007350219879299402, + 0.005733015015721321, + 0.010080879554152489, + 0.007910771295428276, + 0.00742965517565608, + 0.006238395813852549, + 0.01129677053540945, + 0.02582990936934948, + 0.025212116539478302, + 0.02088349312543869, + 0.013282754458487034, + 0.03161890059709549, + 0.021924931555986404, + 0.021819159388542175, + 0.024374723434448242, + 0.018450431525707245, + 0.017077894881367683, + 0.015354683622717857, + 0.009458580985665321, + 0.015483088791370392, + 0.015431306324899197, + 0.014358802698552608, + 0.016063310205936432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.02981209009885788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004100605845451355 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.004100605845451355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004100605845451355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11519594490528107, + 0.16971611976623535, + 0.04913901910185814, + 0.03788338974118233, + 0.1766166388988495, + 0.0796702578663826, + 0.08578754216432571, + 0.041388820856809616, + 0.2157551497220993, + 0.028847061097621918 + ], + "credit_map": [ + 0.022227942943572998, + 0.012897529639303684, + 0.011686839163303375, + 0.009803923778235912, + 0.015296587720513344, + 0.019464189186692238, + 0.009928564541041851, + 0.013890380971133709, + 0.023040467873215675, + 0.024135051295161247, + 0.019185764715075493, + 0.01811152696609497, + 0.029402034357190132, + 0.023227611556649208, + 0.014974195510149002, + 0.01763947494328022, + 0.0054286178201437, + 0.006183437537401915, + 0.004234454594552517, + 0.006051023956388235, + 0.006812776438891888, + 0.0072938380762934685, + 0.005395492073148489, + 0.0077393813990056515, + 0.00532031012699008, + 0.007051276043057442, + 0.0037536059971898794, + 0.003404551651328802, + 0.007319864351302385, + 0.004126419313251972, + 0.00252728583291173, + 0.004380079451948404, + 0.02170472964644432, + 0.012752127833664417, + 0.0664411336183548, + 0.012948382645845413, + 0.016184397041797638, + 0.017125507816672325, + 0.014462336897850037, + 0.01499803364276886, + 0.012315538711845875, + 0.01892939582467079, + 0.01264415867626667, + 0.009139708243310452, + 0.007183133624494076, + 0.006767802406102419, + 0.007513829041272402, + 0.005176699720323086, + 0.013819036073982716, + 0.021007241681218147, + 0.008652635850012302, + 0.008902005851268768, + 0.011436725035309792, + 0.00610854709520936, + 0.006382943596690893, + 0.009478416293859482, + 0.007224630564451218, + 0.007912102155387402, + 0.0035963342525064945, + 0.004602414555847645, + 0.00524912029504776, + 0.004094207193702459, + 0.00374384643509984, + 0.004966169595718384, + 0.032197006046772, + 0.02972559444606304, + 0.02939458191394806, + 0.016420774161815643, + 0.023483967408537865, + 0.026997094973921776, + 0.02408578060567379, + 0.03345036879181862, + 0.0040589431300759315, + 0.0040776170790195465, + 0.0032111413311213255, + 0.002352255629375577, + 0.005009711720049381, + 0.003370382357388735, + 0.0032270560041069984, + 0.0035399580374360085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.004100605845451355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979855239391327 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.02979855239391327, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979855239391327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07637054473161697, + 0.28880271315574646, + 0.06966280937194824, + 0.0568414106965065, + 0.04655931144952774, + 0.09293282777070999, + 0.04961651191115379, + 0.05153105780482292, + 0.12390632927417755, + 0.14377646148204803 + ], + "credit_map": [ + 0.00917926337569952, + 0.013620409183204174, + 0.012236818671226501, + 0.008451346307992935, + 0.009830428287386894, + 0.010986324399709702, + 0.004735130351036787, + 0.007330821827054024, + 0.036281127482652664, + 0.04188496619462967, + 0.03002907522022724, + 0.031257808208465576, + 0.04908108338713646, + 0.038613349199295044, + 0.026300687342882156, + 0.03535458445549011, + 0.009181203320622444, + 0.009484865702688694, + 0.005566607695072889, + 0.00910266861319542, + 0.009917564690113068, + 0.00881263054907322, + 0.007376380264759064, + 0.010220893658697605, + 0.004068656824529171, + 0.009319962933659554, + 0.0029307196382433176, + 0.002754951361566782, + 0.010158945806324482, + 0.012481622397899628, + 0.004081223625689745, + 0.011045332066714764, + 0.006273763719946146, + 0.003626844845712185, + 0.013671044260263443, + 0.004197839647531509, + 0.005156642757356167, + 0.005820174235850573, + 0.0032616546377539635, + 0.004551350604742765, + 0.02399112842977047, + 0.015460059978067875, + 0.006522699724882841, + 0.006563851609826088, + 0.007394799962639809, + 0.004278109408915043, + 0.023149756714701653, + 0.005572428461164236, + 0.006875191815197468, + 0.008582571521401405, + 0.006395735777914524, + 0.005919229239225388, + 0.007077504880726337, + 0.005002697464078665, + 0.004762015305459499, + 0.005001567304134369, + 0.008021912537515163, + 0.00723860552534461, + 0.003808210603892803, + 0.005783436354249716, + 0.0075287409126758575, + 0.005501470994204283, + 0.006303419359028339, + 0.007345262449234724, + 0.017106642946600914, + 0.0158950537443161, + 0.017817385494709015, + 0.009983658790588379, + 0.013784144073724747, + 0.015936540439724922, + 0.014232455752789974, + 0.019150445237755775, + 0.008806806057691574, + 0.028511537238955498, + 0.007708799093961716, + 0.005416436120867729, + 0.04645808786153793, + 0.02529943734407425, + 0.012144387699663639, + 0.009430964477360249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.02979855239391327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 B-17 Flying Fortress bomber with a weathered silver fuselage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bomber gliding slowly over water with its belly almost touching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set at dawn or dusk with a sky showing lavender and pink-orange tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shot a long shot with slight handheld shake and a tilt slightly upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera panning slowly to the right as the bomber approaches an old wooden pier?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305175319314003 + ], + "opposite_logprob_tokens": [ + -3.1680517196655273 + ], + "logprob_reward": -0.04305175319314003, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305175319314003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10447598993778229, + 0.19516399502754211, + 0.07424945384263992, + 0.07513262331485748, + 0.19879905879497528, + 0.06551926583051682, + 0.11236274242401123, + 0.05862284079194069, + 0.05583729222416878, + 0.05983668193221092 + ], + "credit_map": [ + 0.013734288513660431, + 0.01963295042514801, + 0.009297206066548824, + 0.008104565553367138, + 0.014758729375898838, + 0.015859194099903107, + 0.010922838933765888, + 0.012166226282715797, + 0.02376878447830677, + 0.02685379981994629, + 0.016831325367093086, + 0.022965963929891586, + 0.030423646792769432, + 0.024878939613699913, + 0.022166728973388672, + 0.027274826541543007, + 0.009168668650090694, + 0.010952025651931763, + 0.006259739398956299, + 0.009326448664069176, + 0.010062281973659992, + 0.009641166776418686, + 0.008335810154676437, + 0.010503318160772324, + 0.005734991282224655, + 0.012864155694842339, + 0.0036042272113263607, + 0.004054039251059294, + 0.019785400480031967, + 0.013567289337515831, + 0.004499437287449837, + 0.011023083701729774, + 0.02454979717731476, + 0.015710875391960144, + 0.07122654467821121, + 0.014413152821362019, + 0.018399715423583984, + 0.020965825766324997, + 0.015166699886322021, + 0.01836645044386387, + 0.010013447143137455, + 0.014031215570867062, + 0.007636138238012791, + 0.008739291690289974, + 0.006720075849443674, + 0.005605708807706833, + 0.0071443901397287846, + 0.005629002116620541, + 0.016468826681375504, + 0.022543443366885185, + 0.014532094821333885, + 0.015232588164508343, + 0.01035301387310028, + 0.011297916993498802, + 0.010878317058086395, + 0.011056546121835709, + 0.008612096309661865, + 0.006722199730575085, + 0.005351649131625891, + 0.00995938666164875, + 0.006861050147563219, + 0.00804729014635086, + 0.006322687491774559, + 0.006746486760675907, + 0.011404345743358135, + 0.010325183160603046, + 0.0052162338979542255, + 0.00398362847045064, + 0.0060747358947992325, + 0.006226514466106892, + 0.0055390214547514915, + 0.007067630533128977, + 0.00896276906132698, + 0.009188736788928509, + 0.005145590286701918, + 0.005245429463684559, + 0.009826690889894962, + 0.009311666712164879, + 0.006175802554935217, + 0.0059800040908157825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.04305175319314003 + } + ], + "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/0000055/group_2/index_tensor([4222, 4222])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_tensor([4222, 4222])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..59fbf91a31667a5f76e34d1337c269a97e9bedc8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_tensor([4222, 4222])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle shot slowly tilts down across the vast Amazon rainforest, the entire scene animated in a vibrant, nostalgic style inspired by Studio Ghibli. As the camera pans, the forest floor suddenly shakes and cracks, and a subterranean river erupts to the surface with tremendous force. Its churning, foaming, turquoise water swiftly and decisively carves a new, winding channel through dense, verdant undergrowth, sweeping over oversized, waxy leaves and tumbling, moss-covered rocks. Diffuse golden sidelight penetrates the dense jungle canopy in ethereal beams, illuminating swirling mist rising from the swift flow and capturing airborne droplets in a shimmering dance. This magical light punctuates the rich flora—towering ancient trees with massive buttressing roots, dangling vines, and exotic flowers—creating a profound sense of awe and natural wonder as the newborn surface river carves its way through the living, breathing jungle.", + "seed": 11610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11610_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": -0.26427178573794663, + "vlm_logprob": -0.26427178573794663, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.26427178573794663, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8291791081428528 + ], + "opposite_logprob_tokens": [ + -0.5791791081428528 + ], + "logprob_reward": -0.8291791081428528, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8291791081428528, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19043894112110138, + 0.17603422701358795, + 0.26680827140808105, + 0.07509607076644897, + 0.03435699641704559, + 0.057256005704402924, + 0.043947476893663406, + 0.02870308607816696, + 0.04929868504405022, + 0.07806017994880676 + ], + "credit_map": [ + 0.02355256676673889, + 0.01264620665460825, + 0.034451764076948166, + 0.023655887693166733, + 0.017634741961956024, + 0.02611210010945797, + 0.02379351109266281, + 0.028592173010110855, + 0.021087918430566788, + 0.010427439585328102, + 0.021135931834578514, + 0.03126993402838707, + 0.021350666880607605, + 0.01807480864226818, + 0.030129533261060715, + 0.022557996213436127, + 0.029686370864510536, + 0.057331718504428864, + 0.02242973819375038, + 0.013246376067399979, + 0.024205351248383522, + 0.08515463769435883, + 0.011421116068959236, + 0.023332972079515457, + 0.008952844887971878, + 0.010274067521095276, + 0.014117871411144733, + 0.005616903305053711, + 0.009630504064261913, + 0.013068563304841518, + 0.0052322326228022575, + 0.008203087374567986, + 0.00679011968895793, + 0.004807685501873493, + 0.005717525724321604, + 0.0027238759212195873, + 0.005263811443001032, + 0.0033084710594266653, + 0.003245527856051922, + 0.002499975264072418, + 0.0077120037749409676, + 0.011683201417326927, + 0.008321227505803108, + 0.004041897598654032, + 0.008529096841812134, + 0.0049058906733989716, + 0.007288583088666201, + 0.004774102941155434, + 0.005987531039863825, + 0.0039564138278365135, + 0.005842246580868959, + 0.004875463433563709, + 0.005952761974185705, + 0.006353185977786779, + 0.005331974942237139, + 0.005647899582982063, + 0.003714289516210556, + 0.0029533524066209793, + 0.003444785950705409, + 0.0036403872072696686, + 0.003716337261721492, + 0.0028942860662937164, + 0.005544228944927454, + 0.002795417793095112, + 0.007219626102596521, + 0.0036502759903669357, + 0.005757470615208149, + 0.01123104803264141, + 0.003568943589925766, + 0.004586632829159498, + 0.006447512656450272, + 0.006837175693362951, + 0.008342701941728592, + 0.006302988156676292, + 0.009746861644089222, + 0.011607807129621506, + 0.008426863700151443, + 0.011918765492737293, + 0.011384318582713604, + 0.010329870507121086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.8291791081428528 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.474311888217926 + ], + "opposite_logprob_tokens": [ + -0.974311888217926 + ], + "logprob_reward": -0.474311888217926, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.474311888217926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14457036554813385, + 0.10704503208398819, + 0.299621045589447, + 0.05110734701156616, + 0.052309535443782806, + 0.10010252147912979, + 0.05510243773460388, + 0.055567409843206406, + 0.0757906436920166, + 0.05878366157412529 + ], + "credit_map": [ + 0.013027656823396683, + 0.009907853789627552, + 0.022176504135131836, + 0.020595107227563858, + 0.017163202166557312, + 0.01899722032248974, + 0.02116192691028118, + 0.021540895104408264, + 0.009432157501578331, + 0.006702004466205835, + 0.015792759135365486, + 0.01567482016980648, + 0.011958997696638107, + 0.01658458635210991, + 0.015775490552186966, + 0.015124218538403511, + 0.031697552651166916, + 0.061921294778585434, + 0.031023401767015457, + 0.01638849638402462, + 0.025635041296482086, + 0.09564077854156494, + 0.015328429639339447, + 0.02198607847094536, + 0.004537390079349279, + 0.004732648842036724, + 0.0071990881115198135, + 0.0075319078750908375, + 0.005109128076583147, + 0.0072121890261769295, + 0.00890597328543663, + 0.005879023112356663, + 0.008370795287191868, + 0.005024676211178303, + 0.010939315892755985, + 0.004475092515349388, + 0.007572820410132408, + 0.0053003462962806225, + 0.006342421751469374, + 0.004284064285457134, + 0.00960839819163084, + 0.04041536524891853, + 0.016757000237703323, + 0.0046973214484751225, + 0.011173970997333527, + 0.003982356283813715, + 0.008958939462900162, + 0.0045091682113707066, + 0.013185804709792137, + 0.004941283725202084, + 0.009128621779382229, + 0.005568704567849636, + 0.004088455345481634, + 0.005669074598699808, + 0.005866886116564274, + 0.006653606425970793, + 0.0075289481319487095, + 0.007507206406444311, + 0.006323733367025852, + 0.005784747190773487, + 0.005883585195988417, + 0.006728373933583498, + 0.00978870689868927, + 0.006022108253091574, + 0.013369091786444187, + 0.004702843725681305, + 0.007899700663983822, + 0.016983512789011, + 0.005510595161467791, + 0.009175173006951809, + 0.009943773970007896, + 0.008205958642065525, + 0.004811310209333897, + 0.00536208413541317, + 0.007256895769387484, + 0.013194468803703785, + 0.0062958188354969025, + 0.0079502547159791, + 0.007429702673107386, + 0.006483124569058418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.474311888217926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790970012545586 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790970012545586, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790970012545586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07716484367847443, + 0.0940459594130516, + 0.25884830951690674, + 0.14518383145332336, + 0.0486011803150177, + 0.06038984656333923, + 0.04276880994439125, + 0.15179762244224548, + 0.06434722989797592, + 0.05685243010520935 + ], + "credit_map": [ + 0.009487926959991455, + 0.006495000794529915, + 0.012975736521184444, + 0.009875502437353134, + 0.0075906021520495415, + 0.008897327817976475, + 0.01041332632303238, + 0.01142941415309906, + 0.010211056098341942, + 0.006235105916857719, + 0.012184841558337212, + 0.016889424994587898, + 0.011449015699326992, + 0.009235864505171776, + 0.01699313521385193, + 0.010847522877156734, + 0.03158934786915779, + 0.04891052842140198, + 0.025031693279743195, + 0.013666173443198204, + 0.020028937608003616, + 0.08640895038843155, + 0.013014894910156727, + 0.02019777148962021, + 0.018163777887821198, + 0.018616192042827606, + 0.022333601489663124, + 0.014444000087678432, + 0.019612828269600868, + 0.01966778188943863, + 0.013148529455065727, + 0.019197113811969757, + 0.008604636415839195, + 0.005265002604573965, + 0.008163582533597946, + 0.004426855593919754, + 0.007200257852673531, + 0.0042927381582558155, + 0.006571056786924601, + 0.004077052231878042, + 0.007376726251095533, + 0.01617235317826271, + 0.008884096518158913, + 0.0040341089479625225, + 0.007256029639393091, + 0.004907527472823858, + 0.006676465272903442, + 0.00508253974840045, + 0.006475050933659077, + 0.003933379892259836, + 0.007114149164408445, + 0.004541613627225161, + 0.006282794754952192, + 0.0041283173486590385, + 0.005116707179695368, + 0.0051767961122095585, + 0.0275284256786108, + 0.009917357936501503, + 0.007569123525172472, + 0.019954774528741837, + 0.01408106554299593, + 0.0068837436847388744, + 0.05578558146953583, + 0.01007754635065794, + 0.010747993364930153, + 0.003944729454815388, + 0.005735961254686117, + 0.01665755920112133, + 0.005624011158943176, + 0.006803826428949833, + 0.008354543708264828, + 0.006478607188910246, + 0.007799869868904352, + 0.004883537534624338, + 0.00906208623200655, + 0.007223249413073063, + 0.005389698781073093, + 0.007390229031443596, + 0.0071486528031528, + 0.007955103181302547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.009790970012545586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258300807327032 + ], + "opposite_logprob_tokens": [ + -5.255258083343506 + ], + "logprob_reward": -0.005258300807327032, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258300807327032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0389106385409832, + 0.04935983940958977, + 0.4013844132423401, + 0.021836761385202408, + 0.039671748876571655, + 0.12797145545482635, + 0.05939139425754547, + 0.21184249222278595, + 0.03123944066464901, + 0.018391789868474007 + ], + "credit_map": [ + 0.0054015954956412315, + 0.0023737261071801186, + 0.009411919862031937, + 0.0037252341862767935, + 0.003494956297799945, + 0.0027782495599240065, + 0.005062548443675041, + 0.006662409752607346, + 0.00527526531368494, + 0.004020910244435072, + 0.007051265798509121, + 0.005627624690532684, + 0.005110790953040123, + 0.00630402285605669, + 0.007449085358530283, + 0.008520879782736301, + 0.04384584724903107, + 0.08691681176424026, + 0.035715073347091675, + 0.019634412601590157, + 0.03545098379254341, + 0.1241280809044838, + 0.021387431770563126, + 0.034305792301893234, + 0.002900803927332163, + 0.0020231676753610373, + 0.004632549826055765, + 0.002260582521557808, + 0.00186055654194206, + 0.003787311026826501, + 0.002288089133799076, + 0.0020837027113884687, + 0.007622388191521168, + 0.004657537676393986, + 0.0067039006389677525, + 0.00375636899843812, + 0.0061944532208144665, + 0.003504935186356306, + 0.004313728306442499, + 0.0029184387531131506, + 0.012257262133061886, + 0.0513039194047451, + 0.022122912108898163, + 0.005990076810121536, + 0.01409679464995861, + 0.005434352904558182, + 0.011812832206487656, + 0.00495331734418869, + 0.009085233323276043, + 0.006059417966753244, + 0.011203168891370296, + 0.007920508272945881, + 0.005420978181064129, + 0.00674642575904727, + 0.007330290507525206, + 0.005625376012176275, + 0.03851952403783798, + 0.01435503177344799, + 0.011186002753674984, + 0.027563441544771194, + 0.019130242988467216, + 0.00978931039571762, + 0.07681242376565933, + 0.01448653731495142, + 0.00393892265856266, + 0.0021891763899475336, + 0.002807896351441741, + 0.0072078038938343525, + 0.0036416337825357914, + 0.0025332937948405743, + 0.005924977362155914, + 0.002995740156620741, + 0.002704811282455921, + 0.0013314664829522371, + 0.0028501092456281185, + 0.0028072234708815813, + 0.0015055120456963778, + 0.0024398218374699354, + 0.0023583373986184597, + 0.002394510433077812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.005258300807327032 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028186615090817213 + ], + "opposite_logprob_tokens": [ + -5.877818584442139 + ], + "logprob_reward": -0.0028186615090817213, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028186615090817213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062495384365320206, + 0.06488459557294846, + 0.1844489723443985, + 0.044570304453372955, + 0.05206914618611336, + 0.10552062094211578, + 0.10672074556350708, + 0.12594103813171387, + 0.08061769604682922, + 0.1727314591407776 + ], + "credit_map": [ + 0.006952547933906317, + 0.004432603716850281, + 0.007907580584287643, + 0.006127261556684971, + 0.008932949975132942, + 0.010983399115502834, + 0.008060618303716183, + 0.009098422713577747, + 0.007372291758656502, + 0.005301869008690119, + 0.007580813020467758, + 0.009303386323153973, + 0.008144949562847614, + 0.010545267723500729, + 0.008131125941872597, + 0.008504889905452728, + 0.01744951121509075, + 0.043455395847558975, + 0.01387234777212143, + 0.008358637802302837, + 0.018570244312286377, + 0.03541591018438339, + 0.030629368498921394, + 0.01669756881892681, + 0.005523284897208214, + 0.0042420937679708, + 0.007771901786327362, + 0.003765127155929804, + 0.00624497327953577, + 0.006540738511830568, + 0.005728411488234997, + 0.00475377356633544, + 0.010880980640649796, + 0.007058931514620781, + 0.007601962890475988, + 0.004321437794715166, + 0.008265631273388863, + 0.0051487223245203495, + 0.005512180272489786, + 0.0032792999409139156, + 0.011890288442373276, + 0.02912917174398899, + 0.01591051183640957, + 0.0072253188118338585, + 0.012851964682340622, + 0.00871777068823576, + 0.013118849135935307, + 0.0066767409443855286, + 0.015263780020177364, + 0.01509360782802105, + 0.01043317373842001, + 0.01494364719837904, + 0.015547812916338444, + 0.009898943826556206, + 0.011401473544538021, + 0.014138307422399521, + 0.02008821815252304, + 0.014093099161982536, + 0.007379728835076094, + 0.012985050678253174, + 0.014913298189640045, + 0.012383157387375832, + 0.03202354162931442, + 0.012074952945113182, + 0.014038996770977974, + 0.004222707822918892, + 0.006606915965676308, + 0.023189499974250793, + 0.005409278441220522, + 0.0058102086186409, + 0.013831065967679024, + 0.0075090192258358, + 0.018310820683836937, + 0.01712597720324993, + 0.019851839169859886, + 0.03225616365671158, + 0.017942359670996666, + 0.018520666286349297, + 0.028344567865133286, + 0.020379062741994858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.0028186615090817213 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot slowly tilts down across the vast Amazon rainforest, the entire scene animated in a vibrant, nostalgic style inspired by Studio Ghibli. As the camera pans, the forest floor suddenly shakes and cracks, and a subterranean river erupts to the surface with tremendous force. Its churning, foaming, turquoise water swiftly and decisively carves a new, winding channel through dense, verdant undergrowth, sweeping over oversized, waxy leaves and tumbling, moss-covered rocks. Diffuse golden sidelight penetrates the dense jungle canopy in ethereal beams, illuminating swirling mist rising from the swift flow and capturing airborne droplets in a shimmering dance. This magical light punctuates the rich flora—towering ancient trees with massive buttressing roots, dangling vines, and exotic flowers—creating a profound sense of awe and natural wonder as the newborn surface river carves its way through the living, breathing jungle.", + "seed": 11611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11611_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": -0.16001918939873577, + "vlm_logprob": -0.16001918939873577, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.16001918939873577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5780538320541382 + ], + "opposite_logprob_tokens": [ + -0.8280538320541382 + ], + "logprob_reward": -0.5780538320541382, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5780538320541382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2034989893436432, + 0.26946353912353516, + 0.09750352799892426, + 0.12503521144390106, + 0.07825411111116409, + 0.038454439491033554, + 0.06096472591161728, + 0.07649481296539307, + 0.024854587391018867, + 0.025476044043898582 + ], + "credit_map": [ + 0.030278511345386505, + 0.017267446964979172, + 0.03272952511906624, + 0.02027471363544464, + 0.030789192765951157, + 0.018816202878952026, + 0.023568997159600258, + 0.029774414375424385, + 0.022745752707123756, + 0.064058318734169, + 0.06962468475103378, + 0.0173902940005064, + 0.026251304894685745, + 0.022523466497659683, + 0.01604129932820797, + 0.030828434973955154, + 0.011008084751665592, + 0.0111845713108778, + 0.022208496928215027, + 0.010641953907907009, + 0.00959840789437294, + 0.01365739107131958, + 0.008731109090149403, + 0.010473517701029778, + 0.015155277214944363, + 0.015817604959011078, + 0.01635506935417652, + 0.018085574731230736, + 0.011639867909252644, + 0.020111199468374252, + 0.00974934920668602, + 0.018121276050806046, + 0.013956231065094471, + 0.010076847858726978, + 0.009449033997952938, + 0.00873294472694397, + 0.01582479104399681, + 0.011343511752784252, + 0.0034760807175189257, + 0.00539466692134738, + 0.0071877907030284405, + 0.004607720300555229, + 0.004303057212382555, + 0.0038435915485024452, + 0.003900040639564395, + 0.006439139135181904, + 0.003986954689025879, + 0.0041861445643007755, + 0.011045055463910103, + 0.007408548146486282, + 0.0067284852266311646, + 0.007235157769173384, + 0.006263719405978918, + 0.009166907519102097, + 0.006191083695739508, + 0.006925764959305525, + 0.013635770417749882, + 0.0070003317669034, + 0.009748001582920551, + 0.008238093927502632, + 0.008101353421807289, + 0.009771514683961868, + 0.007407079450786114, + 0.012592664919793606, + 0.003221050137653947, + 0.0019218034576624632, + 0.0031599008943885565, + 0.00530320405960083, + 0.002409284235909581, + 0.002738518873229623, + 0.002784567419439554, + 0.0033162587787956, + 0.003665890311822295, + 0.003322456032037735, + 0.002847791649401188, + 0.003184030530974269, + 0.002507063327357173, + 0.004006096161901951, + 0.00277885259129107, + 0.0031638620421290398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.5780538320541382 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015760838985443 + ], + "opposite_logprob_tokens": [ + -1.7015761137008667 + ], + "logprob_reward": -0.2015760838985443, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015760838985443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12243673950433731, + 0.28937625885009766, + 0.08716916292905807, + 0.25108209252357483, + 0.05037730559706688, + 0.0263321865350008, + 0.031185179948806763, + 0.09211795777082443, + 0.0400598905980587, + 0.009863193146884441 + ], + "credit_map": [ + 0.017906352877616882, + 0.012510848231613636, + 0.019300712272524834, + 0.013203982263803482, + 0.014120145700871944, + 0.011095535941421986, + 0.01551690511405468, + 0.018782254308462143, + 0.021931564435362816, + 0.0765308141708374, + 0.0884200781583786, + 0.0147366588935256, + 0.026980487629771233, + 0.019443757832050323, + 0.007378872483968735, + 0.03395402058959007, + 0.00930153951048851, + 0.008083117194473743, + 0.015521909110248089, + 0.008263492956757545, + 0.009727934375405312, + 0.012758043594658375, + 0.007679726462811232, + 0.015833396464586258, + 0.012250404804944992, + 0.007391929626464844, + 0.0859169065952301, + 0.01128090638667345, + 0.013254438526928425, + 0.10330080986022949, + 0.005304492078721523, + 0.012382208369672298, + 0.008414947427809238, + 0.004703220911324024, + 0.007609938271343708, + 0.005801178514957428, + 0.007710369769483805, + 0.007651936262845993, + 0.0037930237594991922, + 0.004692692309617996, + 0.004607698414474726, + 0.00194198964163661, + 0.0021432240027934313, + 0.0018159886822104454, + 0.004377749748528004, + 0.005335412919521332, + 0.001490382943302393, + 0.004619738552719355, + 0.0034772385843098164, + 0.00438776146620512, + 0.0024633179418742657, + 0.0034976929891854525, + 0.006623918656259775, + 0.004382636863738298, + 0.0027955106925219297, + 0.003557103918865323, + 0.011536840349435806, + 0.010366310365498066, + 0.01133687887340784, + 0.008837677538394928, + 0.013710090890526772, + 0.013929132372140884, + 0.010250567458570004, + 0.012150457128882408, + 0.005091083236038685, + 0.003991162870079279, + 0.004742407239973545, + 0.006497385911643505, + 0.0043179066851735115, + 0.005602748598903418, + 0.005913225933909416, + 0.0039039740804582834, + 0.0010340417502447963, + 0.0010800737654790282, + 0.0010402783518657088, + 0.0018049911595880985, + 0.0007278869161382318, + 0.001621793257072568, + 0.0013578657526522875, + 0.001196262426674366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.2015760838985443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10395614802837372, + 0.21423302590847015, + 0.0441015288233757, + 0.36460453271865845, + 0.02175205387175083, + 0.042150091379880905, + 0.042690712958574295, + 0.06439124047756195, + 0.08398471772670746, + 0.01813601702451706 + ], + "credit_map": [ + 0.01668856479227543, + 0.007275749929249287, + 0.018871944397687912, + 0.011954860761761665, + 0.010050873272120953, + 0.009464110247790813, + 0.012465042993426323, + 0.017184991389513016, + 0.016397390514612198, + 0.054050642997026443, + 0.062395721673965454, + 0.011897776275873184, + 0.019712205976247787, + 0.0166323184967041, + 0.008426474407315254, + 0.02472047135233879, + 0.005106586497277021, + 0.0038156104274094105, + 0.008407306857407093, + 0.004929097834974527, + 0.003990774042904377, + 0.0063026901334524155, + 0.0030600479803979397, + 0.008489411324262619, + 0.02172589674592018, + 0.015027379617094994, + 0.11907105147838593, + 0.016914641484618187, + 0.02135240100324154, + 0.13680393993854523, + 0.010438797995448112, + 0.023270409554243088, + 0.0031398606952279806, + 0.0026246269699186087, + 0.003157268976792693, + 0.0030955204274505377, + 0.0027476814575493336, + 0.003735023783519864, + 0.0015706777339801192, + 0.0016813940601423383, + 0.00814096163958311, + 0.0033964805770665407, + 0.005050108768045902, + 0.00500104483217001, + 0.00486470153555274, + 0.0068590473383665085, + 0.0031084949150681496, + 0.005729249678552151, + 0.0076329936273396015, + 0.0051021515391767025, + 0.00482872175052762, + 0.005983025301247835, + 0.005208603106439114, + 0.005094352178275585, + 0.005165713839232922, + 0.003675151150673628, + 0.009430870413780212, + 0.006657262332737446, + 0.008039583452045918, + 0.006950439885258675, + 0.007203135173767805, + 0.008796854875981808, + 0.007180177140980959, + 0.010132912546396255, + 0.010765623301267624, + 0.007603071630001068, + 0.011346319690346718, + 0.01468442752957344, + 0.009385821409523487, + 0.011909046210348606, + 0.009507833980023861, + 0.008782565593719482, + 0.002302079927176237, + 0.0026968782767653465, + 0.00229782541282475, + 0.002448700601235032, + 0.0015714076580479741, + 0.002388606546446681, + 0.001902817515656352, + 0.002527701435610652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.18343648314476013, + 0.35549527406692505, + 0.17804750800132751, + 0.09106836467981339, + 0.031190022826194763, + 0.02748473361134529, + 0.03470876067876816, + 0.028347214683890343, + 0.023576002568006516, + 0.046645648777484894 + ], + "credit_map": [ + 0.03339656442403793, + 0.014472438022494316, + 0.028890198096632957, + 0.015747657045722008, + 0.019841106608510017, + 0.017328200861811638, + 0.019790785387158394, + 0.033969536423683167, + 0.02672812156379223, + 0.09251635521650314, + 0.11184592545032501, + 0.016577735543251038, + 0.03303080052137375, + 0.02278391644358635, + 0.008522320538759232, + 0.043490130454301834, + 0.018937740474939346, + 0.01950913853943348, + 0.02735438384115696, + 0.016421662643551826, + 0.02300058864057064, + 0.02664816565811634, + 0.010585504584014416, + 0.03559032455086708, + 0.008904042653739452, + 0.008820047602057457, + 0.021189238876104355, + 0.00830073095858097, + 0.006418302189558744, + 0.027055351063609123, + 0.0036311994772404432, + 0.00674945255741477, + 0.003021273296326399, + 0.004049409180879593, + 0.005474457982927561, + 0.003998449072241783, + 0.003339389804750681, + 0.005399840418249369, + 0.0025410030502825975, + 0.0033661997877061367, + 0.0033672163262963295, + 0.002794325817376375, + 0.0034384597092866898, + 0.003026732476428151, + 0.004456676542758942, + 0.0038726176135241985, + 0.0029809039551764727, + 0.0035478018689900637, + 0.005699748173356056, + 0.003035755129531026, + 0.003977976739406586, + 0.0042496914975345135, + 0.006080991122871637, + 0.005550791509449482, + 0.002945179818198085, + 0.0031686259899288416, + 0.004511232487857342, + 0.0032062395475804806, + 0.0024018725380301476, + 0.0023404164239764214, + 0.00633924501016736, + 0.0035431948490440845, + 0.0022909047547727823, + 0.0037141083739697933, + 0.002622053027153015, + 0.002152158645913005, + 0.00414538849145174, + 0.0034254358615726233, + 0.0029420731589198112, + 0.003121882677078247, + 0.0027164085768163204, + 0.0024505998007953167, + 0.006569976918399334, + 0.0064413840882480145, + 0.004615442361682653, + 0.005387972574681044, + 0.004256554879248142, + 0.007073007989674807, + 0.004559666384011507, + 0.007741642650216818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.002199375070631504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.21666137874126434, + 0.22601082921028137, + 0.09987514466047287, + 0.2312489002943039, + 0.02505512349307537, + 0.042805567383766174, + 0.0520598404109478, + 0.03314415365457535, + 0.05716706067323685, + 0.015971999615430832 + ], + "credit_map": [ + 0.03781653940677643, + 0.017159322276711464, + 0.0384528711438179, + 0.021253788843750954, + 0.025217320770025253, + 0.019262639805674553, + 0.02140929363667965, + 0.03608958050608635, + 0.017268000170588493, + 0.05533573403954506, + 0.06960020214319229, + 0.014605128206312656, + 0.019885338842868805, + 0.015743497759103775, + 0.007071244064718485, + 0.02650168165564537, + 0.008953711949288845, + 0.008731554262340069, + 0.021365376189351082, + 0.009704281575977802, + 0.007945962250232697, + 0.015990884974598885, + 0.008870727382600307, + 0.018312642350792885, + 0.0156471636146307, + 0.012243933975696564, + 0.06990570574998856, + 0.012262499891221523, + 0.01598144881427288, + 0.07930409908294678, + 0.008937367238104343, + 0.016966672614216805, + 0.003234229050576687, + 0.0025792268570512533, + 0.004931064322590828, + 0.0033567105419933796, + 0.00294788321480155, + 0.003853376256301999, + 0.0019795566331595182, + 0.0021730747539550066, + 0.008398429490625858, + 0.002946858759969473, + 0.004807335324585438, + 0.004811167251318693, + 0.006477372255176306, + 0.006912223994731903, + 0.0035430649295449257, + 0.004909113049507141, + 0.006344515364617109, + 0.007182113826274872, + 0.006386404391378164, + 0.007250976283103228, + 0.006294350139796734, + 0.0059401788748800755, + 0.006556410808116198, + 0.006104889325797558, + 0.006268181372433901, + 0.004193705040961504, + 0.0038334329146891832, + 0.0036289768759161234, + 0.0035032760351896286, + 0.0030245627276599407, + 0.004326437599956989, + 0.004365579225122929, + 0.007653173990547657, + 0.005288965068757534, + 0.006950609385967255, + 0.009717348963022232, + 0.007402933668345213, + 0.008686116896569729, + 0.006264830473810434, + 0.005203074775636196, + 0.002328254049643874, + 0.001986525021493435, + 0.0023846745025366545, + 0.0018309600418433547, + 0.0014734143624082208, + 0.0017208077479153872, + 0.0021343729458749294, + 0.0021129909437149763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0021971152164041996 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot slowly tilts down across the vast Amazon rainforest, the entire scene animated in a vibrant, nostalgic style inspired by Studio Ghibli. As the camera pans, the forest floor suddenly shakes and cracks, and a subterranean river erupts to the surface with tremendous force. Its churning, foaming, turquoise water swiftly and decisively carves a new, winding channel through dense, verdant undergrowth, sweeping over oversized, waxy leaves and tumbling, moss-covered rocks. Diffuse golden sidelight penetrates the dense jungle canopy in ethereal beams, illuminating swirling mist rising from the swift flow and capturing airborne droplets in a shimmering dance. This magical light punctuates the rich flora—towering ancient trees with massive buttressing roots, dangling vines, and exotic flowers—creating a profound sense of awe and natural wonder as the newborn surface river carves its way through the living, breathing jungle.", + "seed": 11612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11612_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": -0.14024652529042214, + "vlm_logprob": -0.14024652529042214, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.14024652529042214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5264899730682373 + ], + "opposite_logprob_tokens": [ + -0.9014899730682373 + ], + "logprob_reward": -0.5264899730682373, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5264899730682373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2431289255619049, + 0.025193391367793083, + 0.049372877925634384, + 0.08126869797706604, + 0.17310209572315216, + 0.15940219163894653, + 0.030814455822110176, + 0.03402075543999672, + 0.13657405972480774, + 0.06712251901626587 + ], + "credit_map": [ + 0.02045065350830555, + 0.019958507269620895, + 0.044674888253211975, + 0.07109375298023224, + 0.011769946664571762, + 0.034967489540576935, + 0.02287359908223152, + 0.01734009198844433, + 0.0028179713990539312, + 0.002245329087600112, + 0.002900882391259074, + 0.0035335025750100613, + 0.002389672677963972, + 0.003350555896759033, + 0.004498372320085764, + 0.003457106649875641, + 0.003039717674255371, + 0.0032853581942617893, + 0.007603155914694071, + 0.0056314715184271336, + 0.00684778019785881, + 0.004018711857497692, + 0.010899996384978294, + 0.008046685717999935, + 0.0074121844954788685, + 0.005608377046883106, + 0.00988904107362032, + 0.008780570700764656, + 0.012335346080362797, + 0.00893508642911911, + 0.01695830374956131, + 0.011349791660904884, + 0.013099619187414646, + 0.01281778234988451, + 0.02126595377922058, + 0.015219457447528839, + 0.02440732531249523, + 0.030943313613533974, + 0.031685665249824524, + 0.023662971332669258, + 0.01934574730694294, + 0.024093130603432655, + 0.02170824445784092, + 0.018885942175984383, + 0.021697038784623146, + 0.023168671876192093, + 0.017290610820055008, + 0.013212824240326881, + 0.0029889035504311323, + 0.004534255713224411, + 0.0040265279822051525, + 0.003878446063026786, + 0.006039011292159557, + 0.004480837844312191, + 0.002492669504135847, + 0.0023738034069538116, + 0.003486393950879574, + 0.004231010563671589, + 0.004394607152789831, + 0.003284227568656206, + 0.006606409326195717, + 0.004852376878261566, + 0.0037845424376428127, + 0.003381188493221998, + 0.010019849054515362, + 0.007271071895956993, + 0.0038452043663710356, + 0.01169348880648613, + 0.010476704686880112, + 0.03606267273426056, + 0.01060912199318409, + 0.04659593850374222, + 0.005225940141826868, + 0.007488898001611233, + 0.005021965596824884, + 0.007290964480489492, + 0.014766507782042027, + 0.007673016283661127, + 0.010690517723560333, + 0.00896470621228218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.5264899730682373 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283011853694916 + ], + "opposite_logprob_tokens": [ + -2.0178301334381104 + ], + "logprob_reward": -0.14283011853694916, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283011853694916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7149890065193176, + 0.035338737070560455, + 0.02167849987745285, + 0.03125057741999626, + 0.034906912595033646, + 0.0457393154501915, + 0.009562952443957329, + 0.02716664783656597, + 0.048964519053697586, + 0.0304027758538723 + ], + "credit_map": [ + 0.02381836622953415, + 0.0376654751598835, + 0.0321449339389801, + 0.28742995858192444, + 0.02221500873565674, + 0.22075992822647095, + 0.06432580947875977, + 0.02662956342101097, + 0.0026236651465296745, + 0.002890876494348049, + 0.004354453179985285, + 0.0036339026410132647, + 0.0026606174651533365, + 0.0018816309748217463, + 0.012089017778635025, + 0.005204574204981327, + 0.001889416598714888, + 0.002279791282489896, + 0.0027624214999377728, + 0.002259005792438984, + 0.002713578986003995, + 0.001810024376027286, + 0.00425027497112751, + 0.0037139886990189552, + 0.0029366386588662863, + 0.0024292480666190386, + 0.003682664129883051, + 0.004173821769654751, + 0.004218015354126692, + 0.002587422728538513, + 0.00680044200271368, + 0.00442232470959425, + 0.0033432699274271727, + 0.004635418765246868, + 0.004132888745516539, + 0.003694885643199086, + 0.0050549996085464954, + 0.004581166431307793, + 0.005192848388105631, + 0.004271436482667923, + 0.0047337887808680534, + 0.006198266055434942, + 0.006614241749048233, + 0.005903400480747223, + 0.00674033910036087, + 0.0062823835760355, + 0.004907880909740925, + 0.004359019920229912, + 0.0015943943290039897, + 0.0013157827779650688, + 0.0009473511017858982, + 0.0012646510731428862, + 0.0013763905735686421, + 0.0012308681616559625, + 0.0008292790153063834, + 0.001004236051812768, + 0.003094595856964588, + 0.0032132002525031567, + 0.0032219658605754375, + 0.0024268634151667356, + 0.005465231370180845, + 0.003998952452093363, + 0.0027340291999280453, + 0.003011810127645731, + 0.0032636141404509544, + 0.003079100977629423, + 0.0019247131422162056, + 0.0041977642104029655, + 0.004267701413482428, + 0.01182718575000763, + 0.004297721199691296, + 0.01610672101378441, + 0.0038815313018858433, + 0.0034505026414990425, + 0.0027533655520528555, + 0.003574993461370468, + 0.004378609824925661, + 0.004962227772921324, + 0.003679528832435608, + 0.0037220180965960026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.14283011853694916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635246515274048 + ], + "opposite_logprob_tokens": [ + -3.6513524055480957 + ], + "logprob_reward": -0.02635246515274048, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635246515274048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6501006484031677, + 0.07838174700737, + 0.027471566572785378, + 0.017537960782647133, + 0.0518072135746479, + 0.07065323740243912, + 0.032534122467041016, + 0.021642671898007393, + 0.018007144331932068, + 0.03186364471912384 + ], + "credit_map": [ + 0.025755448266863823, + 0.036387208849191666, + 0.042323775589466095, + 0.2760958671569824, + 0.021151946857571602, + 0.15528100728988647, + 0.07510357350111008, + 0.01800180971622467, + 0.004135613329708576, + 0.004401763901114464, + 0.007809779141098261, + 0.005749771371483803, + 0.004231249913573265, + 0.003665398806333542, + 0.02817201241850853, + 0.020216163247823715, + 0.002478167647495866, + 0.0020656718406826258, + 0.005323934368789196, + 0.0023882193490862846, + 0.0027048473712056875, + 0.0027774672489613295, + 0.004235144704580307, + 0.005498114041984081, + 0.001228352077305317, + 0.0013469918631017208, + 0.0023168660700321198, + 0.002214935142546892, + 0.002313227392733097, + 0.0015140267787501216, + 0.00396158779039979, + 0.002641974715515971, + 0.0038881637156009674, + 0.005945817567408085, + 0.005810888484120369, + 0.006795786786824465, + 0.006219396833330393, + 0.0072627803310751915, + 0.006173335947096348, + 0.009711041115224361, + 0.0068666147999465466, + 0.012407162226736546, + 0.006847628857940435, + 0.007926588878035545, + 0.011428683996200562, + 0.009951728396117687, + 0.008424188941717148, + 0.0068006436340510845, + 0.0037299427203834057, + 0.0034673060290515423, + 0.0031924324575811625, + 0.004860719665884972, + 0.004947949666529894, + 0.004824741743505001, + 0.003865231527015567, + 0.003645798657089472, + 0.0014759328914806247, + 0.002934901975095272, + 0.0017634877003729343, + 0.0025548546109348536, + 0.003341681556776166, + 0.0027462923899292946, + 0.0038794255815446377, + 0.002946096006780863, + 0.001846199855208397, + 0.0030885678716003895, + 0.0022830930538475513, + 0.0022628693841397762, + 0.002267354167997837, + 0.002100497018545866, + 0.0022374058607965708, + 0.0019211579347029328, + 0.003199713071808219, + 0.0036062838044017553, + 0.0027535089757293463, + 0.002859934465959668, + 0.0060310871340334415, + 0.0047771017998456955, + 0.004630456678569317, + 0.0040055569261312485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02635246515274048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019452712731435895 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019452712731435895, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019452712731435895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35104966163635254, + 0.121710866689682, + 0.03285052999854088, + 0.04859881103038788, + 0.09272444248199463, + 0.16567672789096832, + 0.03172071650624275, + 0.025302814319729805, + 0.08894059807062149, + 0.04142488166689873 + ], + "credit_map": [ + 0.014336091466248035, + 0.019840411841869354, + 0.028174065053462982, + 0.09392021596431732, + 0.019095130264759064, + 0.14248055219650269, + 0.015214895829558372, + 0.017988277599215508, + 0.006242359057068825, + 0.006280741188675165, + 0.011126122437417507, + 0.008425763808190823, + 0.005839391145855188, + 0.00619098823517561, + 0.04172394797205925, + 0.03588155284523964, + 0.0032927123829722404, + 0.002625692170113325, + 0.0055812811478972435, + 0.0030766271520406008, + 0.004262796137481928, + 0.00455918163061142, + 0.0035923935938626528, + 0.005859841592609882, + 0.0045731752179563046, + 0.004862611647695303, + 0.006628990173339844, + 0.006835735868662596, + 0.0054665119387209415, + 0.004460633732378483, + 0.008026809431612492, + 0.007744339294731617, + 0.008787035010755062, + 0.011874164454638958, + 0.01170025859028101, + 0.009298542514443398, + 0.012742615304887295, + 0.011522838845849037, + 0.01497801672667265, + 0.011820974759757519, + 0.01635093428194523, + 0.03171677142381668, + 0.018517978489398956, + 0.022022748365998268, + 0.023770203813910484, + 0.023117801174521446, + 0.017153093591332436, + 0.013027198612689972, + 0.004232077859342098, + 0.004564467817544937, + 0.0033558551222085953, + 0.003845482598990202, + 0.005653813946992159, + 0.0038560249377042055, + 0.0032358646858483553, + 0.002977127907797694, + 0.0030226889066398144, + 0.003141281194984913, + 0.002220524474978447, + 0.0036902304273098707, + 0.0032281738240271807, + 0.003107690252363682, + 0.004133203532546759, + 0.002759021008387208, + 0.006560713518410921, + 0.0051620458252727985, + 0.003146485658362508, + 0.007218346931040287, + 0.008105723187327385, + 0.022329067811369896, + 0.007151825353503227, + 0.02926638349890709, + 0.005376572720706463, + 0.00430763466283679, + 0.0034359232522547245, + 0.004589590709656477, + 0.007076841779053211, + 0.00681085092946887, + 0.004768203478306532, + 0.005059265531599522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0019452712731435895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036147984210401773 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036147984210401773, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036147984210401773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.511114776134491, + 0.03361557424068451, + 0.05806202068924904, + 0.024553494527935982, + 0.053688663989305496, + 0.08841291069984436, + 0.03839145228266716, + 0.024900145828723907, + 0.13661466538906097, + 0.030646277591586113 + ], + "credit_map": [ + 0.01996154524385929, + 0.026551198214292526, + 0.03956843167543411, + 0.20981676876544952, + 0.015316134318709373, + 0.14575447142124176, + 0.03919820114970207, + 0.014948086813092232, + 0.002453272230923176, + 0.0018621172057464719, + 0.0028818235732614994, + 0.002219263231381774, + 0.0019381087040528655, + 0.0023462919052690268, + 0.010874928906559944, + 0.009039771743118763, + 0.005193402990698814, + 0.0033378032967448235, + 0.010648464784026146, + 0.005320383235812187, + 0.005968455225229263, + 0.00449910806491971, + 0.013034399598836899, + 0.010060012340545654, + 0.0019157076021656394, + 0.001884216326288879, + 0.004370071459561586, + 0.00238750665448606, + 0.0025271845515817404, + 0.0020003444515168667, + 0.005705277435481548, + 0.00376318721100688, + 0.004484562203288078, + 0.006498919799923897, + 0.004754677880555391, + 0.005903230048716068, + 0.0077345590107142925, + 0.009202372282743454, + 0.007471436634659767, + 0.007638911716639996, + 0.007313308771699667, + 0.01696396991610527, + 0.008893284015357494, + 0.01186902355402708, + 0.01229272224009037, + 0.011742903850972652, + 0.011235257610678673, + 0.008102447725832462, + 0.002797971712425351, + 0.004633410833775997, + 0.004624334629625082, + 0.006678787060081959, + 0.004883029032498598, + 0.00510422233492136, + 0.005715757608413696, + 0.003953944426029921, + 0.0025765381287783384, + 0.003228840883821249, + 0.0027189648244529963, + 0.0028365179896354675, + 0.004004457499831915, + 0.0031425021588802338, + 0.003682774258777499, + 0.0027095505502074957, + 0.011295422911643982, + 0.005854333750903606, + 0.004943624138832092, + 0.011569366790354252, + 0.009430494159460068, + 0.037070758640766144, + 0.010380091145634651, + 0.04607057198882103, + 0.0034166108816862106, + 0.005774148274213076, + 0.003214500145986676, + 0.004099138546735048, + 0.004185941070318222, + 0.0032404623925685883, + 0.0031074583530426025, + 0.003608020953834057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0036147984210401773 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot slowly tilts down across the vast Amazon rainforest, the entire scene animated in a vibrant, nostalgic style inspired by Studio Ghibli. As the camera pans, the forest floor suddenly shakes and cracks, and a subterranean river erupts to the surface with tremendous force. Its churning, foaming, turquoise water swiftly and decisively carves a new, winding channel through dense, verdant undergrowth, sweeping over oversized, waxy leaves and tumbling, moss-covered rocks. Diffuse golden sidelight penetrates the dense jungle canopy in ethereal beams, illuminating swirling mist rising from the swift flow and capturing airborne droplets in a shimmering dance. This magical light punctuates the rich flora—towering ancient trees with massive buttressing roots, dangling vines, and exotic flowers—creating a profound sense of awe and natural wonder as the newborn surface river carves its way through the living, breathing jungle.", + "seed": 11613, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_2/index_4222_seed_11613_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": -0.08585851155221462, + "vlm_logprob": -0.08585851155221462, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.08585851155221462, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1833152174949646 + ], + "opposite_logprob_tokens": [ + -1.8083152770996094 + ], + "logprob_reward": -0.1833152174949646, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1833152174949646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08415812999010086, + 0.01555568166077137, + 0.20133037865161896, + 0.28826937079429626, + 0.24808786809444427, + 0.03516421839594841, + 0.04765242338180542, + 0.04489992558956146, + 0.014408349059522152, + 0.020473681390285492 + ], + "credit_map": [ + 0.008633526042103767, + 0.009663787670433521, + 0.011314332485198975, + 0.010908471420407295, + 0.008644411340355873, + 0.009632142260670662, + 0.015439331531524658, + 0.009922118857502937, + 0.002508132718503475, + 0.0018487113993614912, + 0.002040569670498371, + 0.0017094856593757868, + 0.0016990554286167026, + 0.0020247504580765963, + 0.0017271984834223986, + 0.001997776562348008, + 0.015353716909885406, + 0.009112701751291752, + 0.012515785172581673, + 0.011793556623160839, + 0.016373664140701294, + 0.060500554740428925, + 0.06419362127780914, + 0.011486779898405075, + 0.020157229155302048, + 0.10026261955499649, + 0.015192503109574318, + 0.01357966661453247, + 0.01795145682990551, + 0.018199995160102844, + 0.08467642217874527, + 0.018249494954943657, + 0.02662736549973488, + 0.08408085256814957, + 0.013195017352700233, + 0.009135134518146515, + 0.08243413269519806, + 0.018355390056967735, + 0.005721455905586481, + 0.008538522757589817, + 0.004368287045508623, + 0.007577138487249613, + 0.005471505224704742, + 0.0027836421504616737, + 0.00427623838186264, + 0.0048543233424425125, + 0.002391797723248601, + 0.0034412865061312914, + 0.0035592808853834867, + 0.012103225104510784, + 0.005821079947054386, + 0.003282802412286401, + 0.010684429667890072, + 0.004521016031503677, + 0.003075989196076989, + 0.004604598507285118, + 0.0036340239457786083, + 0.004181471653282642, + 0.004407551605254412, + 0.004628720227628946, + 0.005484061315655708, + 0.007406012620776892, + 0.005925322882831097, + 0.009232761338353157, + 0.0017027204157784581, + 0.001763549749739468, + 0.002007139381021261, + 0.0012479970464482903, + 0.0020342404022812843, + 0.0024247956462204456, + 0.0015233804006129503, + 0.0017045261338353157, + 0.002546996343880892, + 0.002085549756884575, + 0.0045037344098091125, + 0.0021527360659092665, + 0.0015776653308421373, + 0.003581409342586994, + 0.0015433098888024688, + 0.0024822803679853678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.1833152174949646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225591778755188 + ], + "opposite_logprob_tokens": [ + -1.600591778755188 + ], + "logprob_reward": -0.225591778755188, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225591778755188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06266571581363678, + 0.07841166108846664, + 0.12835387885570526, + 0.43342307209968567, + 0.18339434266090393, + 0.03529301658272743, + 0.03418535366654396, + 0.013263542205095291, + 0.02091156132519245, + 0.010097815655171871 + ], + "credit_map": [ + 0.007548044435679913, + 0.007170545868575573, + 0.007373467553406954, + 0.006841805763542652, + 0.006890390068292618, + 0.00949558150023222, + 0.008602824062108994, + 0.008743060752749443, + 0.008587982505559921, + 0.007986161857843399, + 0.0113611351698637, + 0.008481145836412907, + 0.007779958192259073, + 0.010010626167058945, + 0.009588677436113358, + 0.01461597066372633, + 0.012110062874853611, + 0.012096251361072063, + 0.014523575082421303, + 0.011778048239648342, + 0.012862385250627995, + 0.028844760730862617, + 0.027519118040800095, + 0.008619667962193489, + 0.029027877375483513, + 0.12534117698669434, + 0.014403462409973145, + 0.017264660447835922, + 0.04747326299548149, + 0.026919908821582794, + 0.14866726100444794, + 0.024325452744960785, + 0.017689790576696396, + 0.06139269098639488, + 0.013574209064245224, + 0.007550511974841356, + 0.05413316935300827, + 0.014318354427814484, + 0.006762746721506119, + 0.007972878403961658, + 0.005108622368425131, + 0.0055304341949522495, + 0.004606132861226797, + 0.0028680607210844755, + 0.003986085765063763, + 0.003964852076023817, + 0.004511033184826374, + 0.0047177947126328945, + 0.0032563975546509027, + 0.00425766920670867, + 0.003627016907557845, + 0.004322366788983345, + 0.005099112167954445, + 0.004951873794198036, + 0.004246457479894161, + 0.004424458835273981, + 0.001982672605663538, + 0.000979159027338028, + 0.0015960655873641372, + 0.001339782029390335, + 0.0015819086693227291, + 0.0028283586725592613, + 0.0009534878772683442, + 0.0020021076779812574, + 0.0033953231759369373, + 0.004004430491477251, + 0.002478077309206128, + 0.0016509346896782517, + 0.0027389791794121265, + 0.003317364724352956, + 0.0014595503453165293, + 0.0018669012933969498, + 0.0010136407800018787, + 0.0011116745881736279, + 0.0018559468444436789, + 0.0017195288091897964, + 0.0009144363575614989, + 0.0014617065899074078, + 0.0008570998325012624, + 0.0011637820862233639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.225591778755188 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546890415251255 + ], + "opposite_logprob_tokens": [ + -4.387547016143799 + ], + "logprob_reward": -0.012546890415251255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546890415251255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16141989827156067, + 0.060987602919340134, + 0.14502359926700592, + 0.23546145856380463, + 0.12962321937084198, + 0.09060674905776978, + 0.07006356865167618, + 0.04460003972053528, + 0.01936863735318184, + 0.0428452230989933 + ], + "credit_map": [ + 0.01979789510369301, + 0.018979864194989204, + 0.022194696590304375, + 0.020708627998828888, + 0.017581317573785782, + 0.0184895321726799, + 0.025997379794716835, + 0.017670590430498123, + 0.006143887527287006, + 0.006334446836262941, + 0.007447297219187021, + 0.008270020596683025, + 0.006534364074468613, + 0.008019372820854187, + 0.010044178925454617, + 0.008194035850465298, + 0.011145961470901966, + 0.006978777702897787, + 0.009708866477012634, + 0.009517339989542961, + 0.00979878380894661, + 0.04623544216156006, + 0.0428721122443676, + 0.00876631774008274, + 0.019448360428214073, + 0.03814636915922165, + 0.020589983090758324, + 0.01750364527106285, + 0.031402625143527985, + 0.02287600189447403, + 0.059933632612228394, + 0.025560837239027023, + 0.025248974561691284, + 0.01300788763910532, + 0.022873785346746445, + 0.010996305383741856, + 0.014596214517951012, + 0.011411161161959171, + 0.018038783222436905, + 0.013450099155306816, + 0.008898385800421238, + 0.0148880984634161, + 0.013330958783626556, + 0.008572753518819809, + 0.01352557260543108, + 0.01017192006111145, + 0.010173381306231022, + 0.011045681312680244, + 0.006467620376497507, + 0.01173535268753767, + 0.01319374330341816, + 0.007780909538269043, + 0.011234458535909653, + 0.009241745807230473, + 0.004545898642390966, + 0.005863837897777557, + 0.005116345826536417, + 0.004185441881418228, + 0.004060531500726938, + 0.005578666925430298, + 0.006638672668486834, + 0.007307412568479776, + 0.004317787941545248, + 0.007395181804895401, + 0.0025321166031062603, + 0.004494915250688791, + 0.0023451419547200203, + 0.0013302136212587357, + 0.0034889508970081806, + 0.002300810068845749, + 0.0016301253344863653, + 0.0012463644379749894, + 0.004474150948226452, + 0.003135176608338952, + 0.008647858165204525, + 0.007673847489058971, + 0.0034099454060196877, + 0.00687021529302001, + 0.0043451362289488316, + 0.0042888931930065155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.012546890415251255 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004645033273845911 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004645033273845911, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004645033273845911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14797918498516083, + 0.05381149798631668, + 0.38535061478614807, + 0.0593109093606472, + 0.14674502611160278, + 0.06717661768198013, + 0.03009595349431038, + 0.07938460260629654, + 0.010041882283985615, + 0.02010369673371315 + ], + "credit_map": [ + 0.015054071322083473, + 0.01508259866386652, + 0.021497992798686028, + 0.017772015184164047, + 0.017223602160811424, + 0.019019033759832382, + 0.024682143703103065, + 0.01764771156013012, + 0.007960814982652664, + 0.006561594549566507, + 0.007754793390631676, + 0.005653600674122572, + 0.005812367424368858, + 0.00818279292434454, + 0.0064718155190348625, + 0.005413717124611139, + 0.028218857944011688, + 0.01476222276687622, + 0.02689039334654808, + 0.02504383958876133, + 0.023277398198843002, + 0.1280864179134369, + 0.1170918270945549, + 0.021979643031954765, + 0.005335574503988028, + 0.009774685837328434, + 0.007541191298514605, + 0.005731860175728798, + 0.011111123487353325, + 0.006180193740874529, + 0.006590353790670633, + 0.007045926991850138, + 0.024951180443167686, + 0.032237257808446884, + 0.010041782632470131, + 0.0073686386458575726, + 0.045162323862314224, + 0.012761258520185947, + 0.00750798499211669, + 0.006714594084769487, + 0.007364982273429632, + 0.01310963835567236, + 0.010547636076807976, + 0.005745467729866505, + 0.009533239528536797, + 0.007196038030087948, + 0.00712563656270504, + 0.006553977262228727, + 0.0025302020367234945, + 0.007020651362836361, + 0.0035889886785298586, + 0.0019578381907194853, + 0.006415216717869043, + 0.003607317339628935, + 0.002008218551054597, + 0.0029675199184566736, + 0.006516218185424805, + 0.00873890146613121, + 0.008686200715601444, + 0.006255360320210457, + 0.010859156027436256, + 0.011624511331319809, + 0.00935706589370966, + 0.0173471849411726, + 0.0016396878054365516, + 0.0011329629924148321, + 0.0013117833295837045, + 0.0011443243129178882, + 0.0010814345441758633, + 0.0015002903528511524, + 0.0009362218552269042, + 0.0012951769167557359, + 0.002126330276951194, + 0.0021866364404559135, + 0.003504684427753091, + 0.0025213095359504223, + 0.00174108543433249, + 0.004018873907625675, + 0.002135242335498333, + 0.0018695351900532842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.004645033273845911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a vast Amazon rainforest featuring towering ancient trees with massive buttressing roots and dense undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera start with a high-angle wide shot and slowly tilt down while panning across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the forest floor shake and crack to reveal a subterranean river erupting with turquoise, foaming water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style vibrant and nostalgic, inspired by Studio Ghibli, with diffuse golden sidelight creating ethereal beams through the canopy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the erupting river swiftly carve a new winding channel through oversized waxy leaves and moss-covered rocks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031936378218233585 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.0031936378218233585, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031936378218233585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21893861889839172, + 0.05096302181482315, + 0.06046353653073311, + 0.18880988657474518, + 0.12867753207683563, + 0.0683562383055687, + 0.1118403747677803, + 0.04542876407504082, + 0.06232380494475365, + 0.06419827789068222 + ], + "credit_map": [ + 0.026624996215105057, + 0.02477487549185753, + 0.029152430593967438, + 0.030017312616109848, + 0.022379910573363304, + 0.026899462565779686, + 0.03158525377511978, + 0.02750435471534729, + 0.0055255345068871975, + 0.005349279381334782, + 0.007039587013423443, + 0.0061316038481891155, + 0.005185738205909729, + 0.007320014294236898, + 0.007499342784285545, + 0.006911918055266142, + 0.005356768146157265, + 0.004698595497757196, + 0.017433417961001396, + 0.005134192761033773, + 0.004794420674443245, + 0.01596817933022976, + 0.003652130253612995, + 0.0034258305095136166, + 0.0140201011672616, + 0.04028507322072983, + 0.008855745196342468, + 0.008867169730365276, + 0.02694915048778057, + 0.015177847817540169, + 0.06237822771072388, + 0.012276568450033665, + 0.03788655996322632, + 0.009046530351042747, + 0.008919961750507355, + 0.00575272785499692, + 0.03843488171696663, + 0.01376009825617075, + 0.008117410354316235, + 0.006759356241673231, + 0.008147862739861012, + 0.009402171708643436, + 0.011219829320907593, + 0.005039457231760025, + 0.0093916030600667, + 0.007888393476605415, + 0.009988101199269295, + 0.007278815843164921, + 0.008604873903095722, + 0.01421411894261837, + 0.012104920111596584, + 0.011679363436996937, + 0.027777843177318573, + 0.015239715576171875, + 0.011407414451241493, + 0.010812120512127876, + 0.007337718736380339, + 0.0037239897064864635, + 0.004696888383477926, + 0.004366764798760414, + 0.008537499234080315, + 0.00704698683694005, + 0.005447086412459612, + 0.004271827172487974, + 0.008438341319561005, + 0.008120544254779816, + 0.007583520840853453, + 0.0056071835570037365, + 0.010211632587015629, + 0.010055621154606342, + 0.005254852585494518, + 0.007052104454487562, + 0.005853365641087294, + 0.00710322754457593, + 0.010237790644168854, + 0.013918937183916569, + 0.00638194102793932, + 0.007646976970136166, + 0.007045642007142305, + 0.006010395009070635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0031936378218233585 + } + ], + "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/0000055/group_3/index_tensor([7301, 7301])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_tensor([7301, 7301])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f208497a91cb2b4268398c69eab825610fad1d0c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_tensor([7301, 7301])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, wide shot, the camera panning in a smooth, 360-degree orbit, captures an epic, surreal spectacle from every angle. Amidst a velvety, inky night sky dotted with distant stars, countless fiery meteors streaked like a torrential rain, trailing brilliant white and gold trails and tearing through the night from every angle. Their light briefly and dramatically illuminated a vast parametric holographic projection of a storm below, composed of countless shimmering blue and white light particles. Within this complex, translucent three-dimensional structure, simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors were clearly visible, the light particles undulating subtly as the meteors passed. At the very core of the holographic projection, an ominous crimson glow pulsed steadily and menacingly, its light piercing through layers of data and creating a stark, unsettling contrast against the calm blue and white tones of the surroundings. As the camera continues to move, each angle reveals new details about this meteorological spectacle where the universe and technology intertwine, creating an atmosphere that is both magnificent and awe-inspiring, yet also full of potential threats.", + "seed": 8749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8749_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": -0.4435706580989063, + "vlm_logprob": -0.4435706580989063, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.4435706580989063, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.6087383031845093 + ], + "opposite_logprob_tokens": [ + -0.23373833298683167 + ], + "logprob_reward": -1.6087383031845093, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6087383031845093, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.35159969329833984, + 0.07572274655103683, + 0.03841567784547806, + 0.14961187541484833, + 0.04262319207191467, + 0.09258171170949936, + 0.048892322927713394, + 0.09731042385101318, + 0.05634792894124985, + 0.04689444601535797 + ], + "credit_map": [ + 0.027791457250714302, + 0.020309776067733765, + 0.03880530968308449, + 0.10375973582267761, + 0.028112340718507767, + 0.08284421265125275, + 0.029327020049095154, + 0.020649824291467667, + 0.005408580880612135, + 0.002365399617701769, + 0.02753535471856594, + 0.010882972739636898, + 0.0029969625174999237, + 0.006579466629773378, + 0.0038465256802737713, + 0.016107479110360146, + 0.004239406902343035, + 0.005487704183906317, + 0.004162564408034086, + 0.00314398854970932, + 0.008304859511554241, + 0.003531051566824317, + 0.0036050276830792427, + 0.005941078066825867, + 0.020125219598412514, + 0.01377047784626484, + 0.024554742500185966, + 0.015438276343047619, + 0.012138296850025654, + 0.018394123762845993, + 0.01182346511632204, + 0.03336727246642113, + 0.0060163154266774654, + 0.003817177377641201, + 0.005973448511213064, + 0.0056685940362513065, + 0.005398693960160017, + 0.005425366573035717, + 0.004844202660024166, + 0.005479391664266586, + 0.006226748693734407, + 0.0058584753423929214, + 0.005134539678692818, + 0.00693519739434123, + 0.015146621502935886, + 0.026645874604582787, + 0.007811396382749081, + 0.018822865560650826, + 0.005130748730152845, + 0.008403487503528595, + 0.005944413132965565, + 0.003611261723563075, + 0.006821648683398962, + 0.01012854091823101, + 0.004729974549263716, + 0.004122249782085419, + 0.008829663507640362, + 0.032155390828847885, + 0.0075115030631423, + 0.006025841925293207, + 0.01472074817866087, + 0.008544559590518475, + 0.0051046740263700485, + 0.014418044127523899, + 0.005323521327227354, + 0.009850913658738136, + 0.00987535621970892, + 0.010060805827379227, + 0.004384838044643402, + 0.0072824168018996716, + 0.0051102349534630775, + 0.004459845367819071, + 0.002461872762069106, + 0.0029086696449667215, + 0.013039675541222095, + 0.0033390559256076813, + 0.003855495946481824, + 0.015901807695627213, + 0.0017274355050176382, + 0.0036604367196559906 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -1.6087383031845093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233689546585083 + ], + "opposite_logprob_tokens": [ + -0.8983689546585083 + ], + "logprob_reward": -0.5233689546585083, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233689546585083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7202505469322205, + 0.03993279114365578, + 0.02905571460723877, + 0.022977329790592194, + 0.023556889966130257, + 0.07276567816734314, + 0.016151035204529762, + 0.0374414324760437, + 0.020230889320373535, + 0.01763763092458248 + ], + "credit_map": [ + 0.05805118381977081, + 0.04524274170398712, + 0.06117217615246773, + 0.20920775830745697, + 0.06109717860817909, + 0.1843540072441101, + 0.05894619971513748, + 0.04217938706278801, + 0.004578284919261932, + 0.0025095923338085413, + 0.009352867491543293, + 0.0038660159334540367, + 0.0016780018340796232, + 0.0038700802251696587, + 0.0032407965045422316, + 0.01083715632557869, + 0.0033449670299887657, + 0.00246814894489944, + 0.0017513282364234328, + 0.002064112573862076, + 0.0025830059312283993, + 0.008649335242807865, + 0.0016978754429146647, + 0.00649694399908185, + 0.0030514481477439404, + 0.0023008163552731276, + 0.003531115362420678, + 0.00212047784589231, + 0.0024958839640021324, + 0.0024812568444758654, + 0.001883364049717784, + 0.005112968385219574, + 0.0026210483629256487, + 0.0024204414803534746, + 0.002576397266238928, + 0.0022585252299904823, + 0.003053183201700449, + 0.004862232133746147, + 0.002188029233366251, + 0.003577037015929818, + 0.004823926370590925, + 0.0034539487678557634, + 0.004004275426268578, + 0.006020085886120796, + 0.011714915744960308, + 0.022409526631236076, + 0.005496528930962086, + 0.01484247762709856, + 0.0015481915324926376, + 0.0023319500032812357, + 0.0016862807096913457, + 0.0014108101604506373, + 0.0028734758961945772, + 0.00347072328440845, + 0.0015609198017045856, + 0.001268685795366764, + 0.0034363071899861097, + 0.010220557451248169, + 0.004073913674801588, + 0.003065443132072687, + 0.0051595065742731094, + 0.004164878744632006, + 0.0024096921551972628, + 0.004911137279123068, + 0.0017193816602230072, + 0.0019222345435991883, + 0.0018581327749416232, + 0.0011960490373894572, + 0.0026884048711508512, + 0.0068566519767045975, + 0.0014219853328540921, + 0.0025680505204945803, + 0.0009347950108349323, + 0.0011702594347298145, + 0.004702860955148935, + 0.0014019114896655083, + 0.0013685089070349932, + 0.005453150719404221, + 0.0007899846532382071, + 0.001816161093302071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.5233689546585083 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635337766259909 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.007635337766259909, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635337766259909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2540222108364105, + 0.15277601778507233, + 0.033378664404153824, + 0.33348503708839417, + 0.03354577720165253, + 0.06297081708908081, + 0.04254010692238808, + 0.03515484929084778, + 0.025677742436528206, + 0.02644873410463333 + ], + "credit_map": [ + 0.020199788734316826, + 0.03123031184077263, + 0.019222455099225044, + 0.06129397079348564, + 0.025162745267152786, + 0.06033630296587944, + 0.02099737338721752, + 0.015579272992908955, + 0.012774529866874218, + 0.006081447470933199, + 0.05387392267584801, + 0.0205960925668478, + 0.005933957640081644, + 0.012952576391398907, + 0.007409095764160156, + 0.03315439820289612, + 0.004045998677611351, + 0.003795074997469783, + 0.003916819114238024, + 0.0033062591683119535, + 0.004078838508576155, + 0.004534680396318436, + 0.0036778482608497143, + 0.006023143418133259, + 0.04971810802817345, + 0.02728930674493313, + 0.035930078476667404, + 0.02954822964966297, + 0.036297351121902466, + 0.02929062768816948, + 0.024653909727931023, + 0.1007574200630188, + 0.004026358015835285, + 0.003321784781292081, + 0.004345076624304056, + 0.0037563468795269728, + 0.0045926617458462715, + 0.00565707590430975, + 0.003405864117667079, + 0.0044406079687178135, + 0.004872787743806839, + 0.003916386980563402, + 0.004160318057984114, + 0.005255907773971558, + 0.00964813120663166, + 0.01577853038907051, + 0.006072402931749821, + 0.013266358524560928, + 0.0042367493733763695, + 0.006029004231095314, + 0.004908706992864609, + 0.0034497168380767107, + 0.006223509553819895, + 0.008973232470452785, + 0.004747774917632341, + 0.0039714123122394085, + 0.003164248540997505, + 0.011938413605093956, + 0.002866561757400632, + 0.002244185423478484, + 0.0064531597308814526, + 0.0025469078682363033, + 0.0022517831530421972, + 0.003689591307193041, + 0.0024558783043175936, + 0.004658639896661043, + 0.00413980009034276, + 0.004269921686500311, + 0.002212540013715625, + 0.002747409278526902, + 0.0025799665600061417, + 0.002613587537780404, + 0.0012948967050760984, + 0.0011823431123048067, + 0.00763337267562747, + 0.0020678199362009764, + 0.0022829656954854727, + 0.009194966405630112, + 0.0010708959307521582, + 0.0017214753897860646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.007635337766259909 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062971204519272 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062971204519272, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062971204519272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2449088990688324, + 0.1349000632762909, + 0.05924335494637489, + 0.5181403756141663, + 0.005992033518850803, + 0.010960729792714119, + 0.010073311626911163, + 0.004603795241564512, + 0.0037384447641670704, + 0.007438975851982832 + ], + "credit_map": [ + 0.020506201311945915, + 0.01623813807964325, + 0.014611062593758106, + 0.0657293051481247, + 0.02334325946867466, + 0.06993968784809113, + 0.019724447280168533, + 0.014816784299910069, + 0.009176556952297688, + 0.004412316717207432, + 0.048174068331718445, + 0.01849951036274433, + 0.00441175140440464, + 0.008198591880500317, + 0.00711739482358098, + 0.03490987420082092, + 0.006096994504332542, + 0.009233023039996624, + 0.00674801878631115, + 0.004985202103853226, + 0.01312087383121252, + 0.006546081975102425, + 0.005606289952993393, + 0.006906876806169748, + 0.0750231146812439, + 0.04270908236503601, + 0.061540041118860245, + 0.04777727276086807, + 0.054811276495456696, + 0.04809359833598137, + 0.03911144286394119, + 0.149074524641037, + 0.0007452311692759395, + 0.0005949240876361728, + 0.0009699269430711865, + 0.0007982639945112169, + 0.0007910417625680566, + 0.000529140408616513, + 0.0008021811372600496, + 0.0007613236666657031, + 0.0006755812210030854, + 0.0007450977573171258, + 0.0006825659656897187, + 0.001260804245248437, + 0.002157326554879546, + 0.0029488177970051765, + 0.0009597609750926495, + 0.001530775218270719, + 0.0007482210057787597, + 0.0023205862380564213, + 0.0008201716700568795, + 0.0007642711279913783, + 0.001174911274574697, + 0.0024153469130396843, + 0.0009415499516762793, + 0.0008882533293217421, + 0.0004816145228687674, + 0.001187547342851758, + 0.0006789902690798044, + 0.0004073143354617059, + 0.0005971475620754063, + 0.0003479112056083977, + 0.0005684366915374994, + 0.00033483351580798626, + 0.0003721747489180416, + 0.0006741302786394954, + 0.0005934266373515129, + 0.0006086707580834627, + 0.0003036921552848071, + 0.0004083559033460915, + 0.00040947660454548895, + 0.0003685177071020007, + 0.0003791704075410962, + 0.00027282870723865926, + 0.002188407815992832, + 0.0005279936012811959, + 0.0005676496075466275, + 0.0024681438226252794, + 0.000313277036184445, + 0.0007215053192339838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.016062971204519272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062047723680734634 + ], + "opposite_logprob_tokens": [ + -2.8120477199554443 + ], + "logprob_reward": -0.062047723680734634, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062047723680734634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5614505410194397, + 0.1354924589395523, + 0.03210887685418129, + 0.0696692019701004, + 0.01124296709895134, + 0.07480917870998383, + 0.057577911764383316, + 0.0241643488407135, + 0.0117470882833004, + 0.021737389266490936 + ], + "credit_map": [ + 0.04730110242962837, + 0.04229016602039337, + 0.034423306584358215, + 0.14977937936782837, + 0.05340543016791344, + 0.15713506937026978, + 0.04506421461701393, + 0.03205188736319542, + 0.00975827220827341, + 0.004131109919399023, + 0.04884932190179825, + 0.018703985959291458, + 0.004921513609588146, + 0.009840492159128189, + 0.0068720243871212006, + 0.03241575136780739, + 0.003311456646770239, + 0.0029405970126390457, + 0.002732930239289999, + 0.00278054759837687, + 0.004718858283013105, + 0.005727312061935663, + 0.003013711655512452, + 0.0068834638223052025, + 0.008995125070214272, + 0.006582694593816996, + 0.009411392733454704, + 0.006313051097095013, + 0.009162292815744877, + 0.007051401771605015, + 0.005928691476583481, + 0.016224553808569908, + 0.0015663385856896639, + 0.0011876511853188276, + 0.0019550470169633627, + 0.001228124019689858, + 0.0007541087688878179, + 0.0011960527626797557, + 0.0011504413560032845, + 0.002205203752964735, + 0.005914239678531885, + 0.005413682200014591, + 0.004323252942413092, + 0.006811985280364752, + 0.01409895159304142, + 0.018347658216953278, + 0.006978617515414953, + 0.012920799665153027, + 0.004325860179960728, + 0.007614622358232737, + 0.005099165719002485, + 0.004799754824489355, + 0.010479806922376156, + 0.013375837355852127, + 0.0061825597658753395, + 0.005700312554836273, + 0.0020982902497053146, + 0.004021462053060532, + 0.0033890698105096817, + 0.002711227396503091, + 0.0030928393825888634, + 0.0030491945799440145, + 0.002811075421050191, + 0.002991190878674388, + 0.0008091242052614689, + 0.0010440672049298882, + 0.0019389657536521554, + 0.0018183079082518816, + 0.0012155738659203053, + 0.0019015897996723652, + 0.0017969172913581133, + 0.0012225424870848656, + 0.0011281527113169432, + 0.0018276595510542393, + 0.0037405085749924183, + 0.003658441361039877, + 0.0018046614713966846, + 0.0021314334589987993, + 0.002554969396442175, + 0.004891565069556236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.062047723680734634 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, wide shot, the camera panning in a smooth, 360-degree orbit, captures an epic, surreal spectacle from every angle. Amidst a velvety, inky night sky dotted with distant stars, countless fiery meteors streaked like a torrential rain, trailing brilliant white and gold trails and tearing through the night from every angle. Their light briefly and dramatically illuminated a vast parametric holographic projection of a storm below, composed of countless shimmering blue and white light particles. Within this complex, translucent three-dimensional structure, simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors were clearly visible, the light particles undulating subtly as the meteors passed. At the very core of the holographic projection, an ominous crimson glow pulsed steadily and menacingly, its light piercing through layers of data and creating a stark, unsettling contrast against the calm blue and white tones of the surroundings. As the camera continues to move, each angle reveals new details about this meteorological spectacle where the universe and technology intertwine, creating an atmosphere that is both magnificent and awe-inspiring, yet also full of potential threats.", + "seed": 8750, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8750_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": -0.25654800748452544, + "vlm_logprob": -0.25654800748452544, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.25654800748452544, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.0650638341903687 + ], + "opposite_logprob_tokens": [ + -0.44006386399269104 + ], + "logprob_reward": -1.0650638341903687, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0650638341903687, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16629816591739655, + 0.09021653234958649, + 0.09607808291912079, + 0.0708516538143158, + 0.22245080769062042, + 0.08260627090930939, + 0.04517534747719765, + 0.06907419860363007, + 0.05400329828262329, + 0.10324563086032867 + ], + "credit_map": [ + 0.02075832337141037, + 0.020756429061293602, + 0.016638565808534622, + 0.017344951629638672, + 0.020026953890919685, + 0.01877530850470066, + 0.029884088784456253, + 0.02211354300379753, + 0.012276622466742992, + 0.013715892098844051, + 0.008586566895246506, + 0.009305920451879501, + 0.013644509948790073, + 0.017612308263778687, + 0.007108967751264572, + 0.00796574354171753, + 0.0064722690731287, + 0.004765304271131754, + 0.030975688248872757, + 0.009490570053458214, + 0.009114355780184269, + 0.006306936033070087, + 0.005680643487721682, + 0.0232723169028759, + 0.008687656372785568, + 0.0066739232279360294, + 0.006596268154680729, + 0.00691330898553133, + 0.012172267772257328, + 0.010503334924578667, + 0.009274890646338463, + 0.010030001401901245, + 0.05021708831191063, + 0.029825136065483093, + 0.0148786511272192, + 0.01536556426435709, + 0.03996609151363373, + 0.04263732209801674, + 0.01382419839501381, + 0.0157367791980505, + 0.014085948467254639, + 0.010946748778223991, + 0.0077581158839166164, + 0.005946679040789604, + 0.017384659498929977, + 0.010841473937034607, + 0.007422717288136482, + 0.008219925686717033, + 0.004845575429499149, + 0.007142611779272556, + 0.003692764090374112, + 0.004055233206599951, + 0.008640069514513016, + 0.007641082163900137, + 0.004217792768031359, + 0.004940220154821873, + 0.022206183522939682, + 0.009167307056486607, + 0.005895921029150486, + 0.004187832586467266, + 0.006347202695906162, + 0.010630481876432896, + 0.0032229446806013584, + 0.00741632329300046, + 0.0037516753654927015, + 0.004184415098279715, + 0.01513917837291956, + 0.0034773240331560373, + 0.002542466390877962, + 0.004704141058027744, + 0.003806757042184472, + 0.016397343948483467, + 0.008903225883841515, + 0.02137814089655876, + 0.019424034282565117, + 0.007530588656663895, + 0.008066977374255657, + 0.00911164190620184, + 0.009068879298865795, + 0.019762149080634117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -1.0650638341903687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034903138875961 + ], + "opposite_logprob_tokens": [ + -2.350348949432373 + ], + "logprob_reward": -0.10034903138875961, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034903138875961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07312405854463577, + 0.06102003902196884, + 0.09293273836374283, + 0.05199582874774933, + 0.2655803859233856, + 0.07842739671468735, + 0.03974132612347603, + 0.024438438937067986, + 0.0893225371837616, + 0.2234172374010086 + ], + "credit_map": [ + 0.009677463211119175, + 0.010259927250444889, + 0.00852169655263424, + 0.0072817495092749596, + 0.007589197717607021, + 0.006110784597694874, + 0.013209010474383831, + 0.010474227368831635, + 0.006228960584849119, + 0.007563531398773193, + 0.008136015385389328, + 0.009310023859143257, + 0.010067836381494999, + 0.006851592101156712, + 0.007197671569883823, + 0.005664404947310686, + 0.00495715020224452, + 0.0035949854645878077, + 0.03222358226776123, + 0.00959597434848547, + 0.00406185258179903, + 0.007433723192662001, + 0.004462736193090677, + 0.02660273388028145, + 0.005760029889643192, + 0.006177036557346582, + 0.005755948834121227, + 0.00532684987410903, + 0.008648508228361607, + 0.0079197371378541, + 0.005396230611950159, + 0.007011487148702145, + 0.01588679850101471, + 0.01728600263595581, + 0.014695007354021072, + 0.021478869020938873, + 0.09311588108539581, + 0.07812689244747162, + 0.01283194962888956, + 0.012159005738794804, + 0.023325329646468163, + 0.008973836898803711, + 0.00580732524394989, + 0.004491626285016537, + 0.012172014452517033, + 0.010758565738797188, + 0.004746933002024889, + 0.008151763118803501, + 0.0035943863913416862, + 0.005217276979237795, + 0.001607335638254881, + 0.0023174083326011896, + 0.007149787619709969, + 0.009055056609213352, + 0.0025726077146828175, + 0.008227468468248844, + 0.005051981192082167, + 0.0032252527307718992, + 0.0024578089360147715, + 0.0013783456524834037, + 0.002175429603084922, + 0.004690113011747599, + 0.0012654815800487995, + 0.004194026812911034, + 0.005473159719258547, + 0.009403734467923641, + 0.028859999030828476, + 0.003312914865091443, + 0.0036528566852211952, + 0.006758220959454775, + 0.002981689525768161, + 0.028879960998892784, + 0.022858185693621635, + 0.08470717072486877, + 0.045854147523641586, + 0.009396512992680073, + 0.01683378964662552, + 0.021945271641016006, + 0.0065862443298101425, + 0.015235929749906063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 2 + ], + "selected_reward": -0.10034903138875961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077099479734898 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011077099479734898, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077099479734898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11401855945587158, + 0.040527716279029846, + 0.04931650310754776, + 0.03962016478180885, + 0.18552100658416748, + 0.23140111565589905, + 0.07198607176542282, + 0.0522678904235363, + 0.06620743870735168, + 0.1491335779428482 + ], + "credit_map": [ + 0.01962919533252716, + 0.021732989698648453, + 0.010862931609153748, + 0.009973778389394283, + 0.01376643218100071, + 0.0138772651553154, + 0.011427507735788822, + 0.012748458422720432, + 0.0040256185457110405, + 0.004406939260661602, + 0.007412769831717014, + 0.0048120585270226, + 0.007417534943670034, + 0.005471391137689352, + 0.003907297737896442, + 0.0030741028022021055, + 0.003299908246845007, + 0.0028613079339265823, + 0.01637602411210537, + 0.005065917037427425, + 0.003093837294727564, + 0.008427930064499378, + 0.0028728449251502752, + 0.0073187281377613544, + 0.005088514648377895, + 0.005686751566827297, + 0.003338057082146406, + 0.003960923757404089, + 0.008932986296713352, + 0.004783048294484615, + 0.004570829216390848, + 0.0032590499613434076, + 0.010681070387363434, + 0.011721323244273663, + 0.006019169930368662, + 0.006988044828176498, + 0.07366807758808136, + 0.061406295746564865, + 0.007538932841271162, + 0.007498068269342184, + 0.05445593222975731, + 0.02974056638777256, + 0.018542854115366936, + 0.015287897549569607, + 0.04139673709869385, + 0.039018575102090836, + 0.010824386030435562, + 0.022134143859148026, + 0.014174503274261951, + 0.007095312234014273, + 0.0038626983296126127, + 0.0043311226181685925, + 0.013610739260911942, + 0.014183463528752327, + 0.004923322703689337, + 0.009804901666939259, + 0.007081764750182629, + 0.0053758262656629086, + 0.010213715024292469, + 0.003449233714491129, + 0.005581758450716734, + 0.010054262354969978, + 0.0033911673817783594, + 0.00712015712633729, + 0.003266408573836088, + 0.0063741253688931465, + 0.02290000207722187, + 0.0025041287299245596, + 0.0025329601485282183, + 0.003431602381169796, + 0.0024603698402643204, + 0.022737830877304077, + 0.013999342918395996, + 0.05155588313937187, + 0.027800412848591805, + 0.007637114264070988, + 0.011082513257861137, + 0.014815160073339939, + 0.006298054940998554, + 0.0159450750797987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.011077099479734898 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943956319242716 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943956319242716, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943956319242716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10871028900146484, + 0.04422168806195259, + 0.16647912561893463, + 0.0539851039648056, + 0.2028712034225464, + 0.1559445559978485, + 0.05580611526966095, + 0.11615996062755585, + 0.07357737421989441, + 0.022244639694690704 + ], + "credit_map": [ + 0.014661413617432117, + 0.015538640320301056, + 0.011291436851024628, + 0.010417507030069828, + 0.011321178637444973, + 0.015125876292586327, + 0.013746262528002262, + 0.016607973724603653, + 0.005918784067034721, + 0.006441617850214243, + 0.005184810608625412, + 0.006377779878675938, + 0.006434772163629532, + 0.0060422890819609165, + 0.003440510481595993, + 0.004381123464554548, + 0.010480308905243874, + 0.008745057508349419, + 0.0577310211956501, + 0.017828909680247307, + 0.00831917580217123, + 0.026090752333402634, + 0.008323661983013153, + 0.028960227966308594, + 0.006145814899355173, + 0.007480674423277378, + 0.005191037431359291, + 0.004989833105355501, + 0.011206437833607197, + 0.007786449044942856, + 0.0061088986694812775, + 0.00507596367970109, + 0.012964933179318905, + 0.013381099328398705, + 0.007439630571752787, + 0.008155124261975288, + 0.0767831951379776, + 0.06710260361433029, + 0.008754235692322254, + 0.008290386758744717, + 0.02032541297376156, + 0.020025014877319336, + 0.016664505004882812, + 0.013043402694165707, + 0.02409462444484234, + 0.028911611065268517, + 0.017751220613718033, + 0.015128770843148232, + 0.009659984149038792, + 0.005339133553206921, + 0.00470385979861021, + 0.005487141199409962, + 0.008755595423281193, + 0.009194779209792614, + 0.006243139039725065, + 0.0064224847592413425, + 0.02033914625644684, + 0.01144246757030487, + 0.017762115225195885, + 0.009491562843322754, + 0.01169453002512455, + 0.017176778987050056, + 0.012113571166992188, + 0.016139790415763855, + 0.005471324548125267, + 0.00870074238628149, + 0.015773965045809746, + 0.006415923126041889, + 0.0052761356346309185, + 0.010454380884766579, + 0.008032064884901047, + 0.013452840968966484, + 0.0024600003380328417, + 0.0026099795941263437, + 0.0025023738853633404, + 0.0026538630481809378, + 0.003085905918851495, + 0.0029054570477455854, + 0.002912314375862479, + 0.0031147454865276814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.005943956319242716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030611604452133 + ], + "opposite_logprob_tokens": [ + -2.3503060340881348 + ], + "logprob_reward": -0.10030611604452133, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030611604452133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12366610765457153, + 0.09627029299736023, + 0.07531677186489105, + 0.04532134160399437, + 0.12201477587223053, + 0.10974764078855515, + 0.16953527927398682, + 0.10676206648349762, + 0.09133464843034744, + 0.06003108248114586 + ], + "credit_map": [ + 0.013722111470997334, + 0.01555875688791275, + 0.013189643621444702, + 0.015007480047643185, + 0.011752398684620857, + 0.01879505254328251, + 0.01717982068657875, + 0.018460847437381744, + 0.008772607892751694, + 0.012115326710045338, + 0.009423406794667244, + 0.007503034081310034, + 0.02235044725239277, + 0.013635246083140373, + 0.0072808535769581795, + 0.015189366415143013, + 0.005896128248423338, + 0.003928127698600292, + 0.026006262749433517, + 0.007807786576449871, + 0.004077306482940912, + 0.008412804454565048, + 0.004659957718104124, + 0.014528400264680386, + 0.005340557545423508, + 0.006060527637600899, + 0.004376493394374847, + 0.005555835086852312, + 0.006641075015068054, + 0.004807017743587494, + 0.007728135213255882, + 0.004811701364815235, + 0.039058152586221695, + 0.018460910767316818, + 0.00868191011250019, + 0.008328392170369625, + 0.014673524536192417, + 0.014216158539056778, + 0.008869529701769352, + 0.009726198390126228, + 0.030110610648989677, + 0.011913232505321503, + 0.009134185500442982, + 0.0071290466003119946, + 0.017301097512245178, + 0.017548872157931328, + 0.006195825524628162, + 0.010414764285087585, + 0.02516975998878479, + 0.023214612156152725, + 0.007375805173069239, + 0.00903049111366272, + 0.028680160641670227, + 0.04156527668237686, + 0.008320840075612068, + 0.02617833763360977, + 0.019284099340438843, + 0.010072438046336174, + 0.020801637321710587, + 0.0076094986870884895, + 0.012754647992551327, + 0.014388221316039562, + 0.0067514944821596146, + 0.015100021846592426, + 0.00856713205575943, + 0.0152444988489151, + 0.018542859703302383, + 0.005693843122571707, + 0.0052574509754776955, + 0.012428054586052895, + 0.006408216897398233, + 0.019192595034837723, + 0.006341628264635801, + 0.013141597621142864, + 0.006981892045587301, + 0.005632536485791206, + 0.005081108771264553, + 0.0067131007090210915, + 0.005865512881428003, + 0.010273708961904049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.10030611604452133 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, wide shot, the camera panning in a smooth, 360-degree orbit, captures an epic, surreal spectacle from every angle. Amidst a velvety, inky night sky dotted with distant stars, countless fiery meteors streaked like a torrential rain, trailing brilliant white and gold trails and tearing through the night from every angle. Their light briefly and dramatically illuminated a vast parametric holographic projection of a storm below, composed of countless shimmering blue and white light particles. Within this complex, translucent three-dimensional structure, simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors were clearly visible, the light particles undulating subtly as the meteors passed. At the very core of the holographic projection, an ominous crimson glow pulsed steadily and menacingly, its light piercing through layers of data and creating a stark, unsettling contrast against the calm blue and white tones of the surroundings. As the camera continues to move, each angle reveals new details about this meteorological spectacle where the universe and technology intertwine, creating an atmosphere that is both magnificent and awe-inspiring, yet also full of potential threats.", + "seed": 8751, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8751_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": -0.4802787813358009, + "vlm_logprob": -0.4802787813358009, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.4802787813358009, + "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 camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.2343841791152954 + ], + "opposite_logprob_tokens": [ + -0.359384149312973 + ], + "logprob_reward": -1.2343841791152954, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2343841791152954, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11392902582883835, + 0.0770881325006485, + 0.061788227409124374, + 0.05212639272212982, + 0.12017766386270523, + 0.11760172247886658, + 0.0636947900056839, + 0.16218246519565582, + 0.13779817521572113, + 0.09361341595649719 + ], + "credit_map": [ + 0.012034835293889046, + 0.015378560870885849, + 0.023525279015302658, + 0.016486143693327904, + 0.008999488316476345, + 0.011515718884766102, + 0.012666227295994759, + 0.01332276314496994, + 0.007753490004688501, + 0.00848784763365984, + 0.00855270866304636, + 0.010788890533149242, + 0.009958994574844837, + 0.008887460455298424, + 0.010580155998468399, + 0.01207857858389616, + 0.0070352498441934586, + 0.006748397834599018, + 0.007704794406890869, + 0.012331712059676647, + 0.005158758722245693, + 0.007167375180870295, + 0.00763380341231823, + 0.008008130826056004, + 0.004792246967554092, + 0.009133215062320232, + 0.006325351539999247, + 0.008565884083509445, + 0.004029640927910805, + 0.00562412990257144, + 0.005884544923901558, + 0.007771375123411417, + 0.011804773472249508, + 0.0062790485098958015, + 0.01939620077610016, + 0.009236941114068031, + 0.010753735899925232, + 0.015052040107548237, + 0.02233554981648922, + 0.02531936764717102, + 0.013143697753548622, + 0.03923449292778969, + 0.007059111725538969, + 0.006048085633665323, + 0.008254115469753742, + 0.014265802688896656, + 0.005926699843257666, + 0.023669710382819176, + 0.007275649346411228, + 0.01966099441051483, + 0.009055220521986485, + 0.006218205206096172, + 0.004541885107755661, + 0.008898396044969559, + 0.0037820960860699415, + 0.0042623430490493774, + 0.02395646832883358, + 0.016802409663796425, + 0.017874691635370255, + 0.028065020218491554, + 0.016496464610099792, + 0.017078403383493423, + 0.020634418353438377, + 0.021274587139487267, + 0.009068094193935394, + 0.013375427573919296, + 0.021581944078207016, + 0.012319419533014297, + 0.012065394781529903, + 0.03713885322213173, + 0.015613454394042492, + 0.016635578125715256, + 0.008781843818724155, + 0.016622940078377724, + 0.013186296448111534, + 0.010559462942183018, + 0.00936330296099186, + 0.008180362172424793, + 0.00993092730641365, + 0.016988275572657585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -1.2343841791152954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.1370728015899658 + ], + "opposite_logprob_tokens": [ + -0.38707277178764343 + ], + "logprob_reward": -1.1370728015899658, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370728015899658, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13077771663665771, + 0.10897950083017349, + 0.1164715439081192, + 0.09855453670024872, + 0.11929672956466675, + 0.20981474220752716, + 0.08004866540431976, + 0.045481499284505844, + 0.03702843561768532, + 0.053546614944934845 + ], + "credit_map": [ + 0.01579936407506466, + 0.013462553732097149, + 0.018505193293094635, + 0.010300658643245697, + 0.009955979883670807, + 0.03372495248913765, + 0.010147304274141788, + 0.01888171024620533, + 0.01010223664343357, + 0.01330519001930952, + 0.012762567959725857, + 0.015249809250235558, + 0.016310935840010643, + 0.018639231100678444, + 0.012440052814781666, + 0.010169475339353085, + 0.016678927466273308, + 0.012196039780974388, + 0.017500339075922966, + 0.013923768885433674, + 0.01785259321331978, + 0.012750387191772461, + 0.014814491383731365, + 0.010755001567304134, + 0.008155884221196175, + 0.011350870132446289, + 0.013546197675168514, + 0.0172567181289196, + 0.009624287486076355, + 0.010736271739006042, + 0.012307346798479557, + 0.015576960518956184, + 0.011554593220353127, + 0.01054589543491602, + 0.013290200382471085, + 0.011758177541196346, + 0.018741577863693237, + 0.02201448194682598, + 0.01507338136434555, + 0.016318408772349358, + 0.02554173767566681, + 0.05932366102933884, + 0.01613951288163662, + 0.01461803913116455, + 0.018404247239232063, + 0.020332347601652145, + 0.014679823070764542, + 0.0407753549516201, + 0.007093445397913456, + 0.0196617990732193, + 0.01017958577722311, + 0.008927919901907444, + 0.011073277331888676, + 0.0090249702334404, + 0.0064138080924749374, + 0.007673858664929867, + 0.00442065205425024, + 0.00488225556910038, + 0.005274063441902399, + 0.006053300108760595, + 0.005966453813016415, + 0.005907321348786354, + 0.007826660759747028, + 0.005150794982910156, + 0.0034707935992628336, + 0.005646601319313049, + 0.0049383570440113544, + 0.004455823916941881, + 0.004142193589359522, + 0.004672089125961065, + 0.004802285693585873, + 0.004900289233773947, + 0.00554592115804553, + 0.01148434542119503, + 0.009041307494044304, + 0.005462519358843565, + 0.005598259158432484, + 0.004042330197989941, + 0.005366567522287369, + 0.007005364168435335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -1.1370728015899658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260079633444548 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005260079633444548, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260079633444548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12607285380363464, + 0.07338481396436691, + 0.09475718438625336, + 0.09687142074108124, + 0.12197937816381454, + 0.186935156583786, + 0.07767391204833984, + 0.07559701800346375, + 0.08616979420185089, + 0.060558535158634186 + ], + "credit_map": [ + 0.013843805529177189, + 0.012695088982582092, + 0.024924656376242638, + 0.013118806295096874, + 0.007790760602802038, + 0.024440664798021317, + 0.013465875759720802, + 0.015793167054653168, + 0.007020844146609306, + 0.00893023144453764, + 0.009812594391405582, + 0.008893955498933792, + 0.011901844292879105, + 0.012902267277240753, + 0.008321833796799183, + 0.005601236131042242, + 0.01590893417596817, + 0.00953163392841816, + 0.017172137275338173, + 0.011313769035041332, + 0.011373920366168022, + 0.009319974109530449, + 0.011184217408299446, + 0.008952590636909008, + 0.00618444150313735, + 0.009481742046773434, + 0.014201801270246506, + 0.016570108011364937, + 0.00876309722661972, + 0.010758679360151291, + 0.014143274165689945, + 0.016768263652920723, + 0.009878283366560936, + 0.009821618907153606, + 0.016231246292591095, + 0.011408726684749126, + 0.01678513176739216, + 0.01992190070450306, + 0.017737392336130142, + 0.020195061340928078, + 0.022945143282413483, + 0.05543747916817665, + 0.013968192040920258, + 0.012168393470346928, + 0.015520159155130386, + 0.01879756897687912, + 0.012338357046246529, + 0.03575984388589859, + 0.008228609338402748, + 0.02291630581021309, + 0.011054537259042263, + 0.007101729046553373, + 0.007731197401881218, + 0.009214290417730808, + 0.005058125592768192, + 0.006369114387780428, + 0.010127970948815346, + 0.007946661673486233, + 0.008418613113462925, + 0.012003874406218529, + 0.00860195979475975, + 0.008083359338343143, + 0.010524235665798187, + 0.00989033654332161, + 0.0065904827788472176, + 0.009020938538014889, + 0.014562641270458698, + 0.008040276356041431, + 0.006306310184299946, + 0.0215633362531662, + 0.007533998694270849, + 0.012551799416542053, + 0.00745709054172039, + 0.013061796315014362, + 0.009488234296441078, + 0.005511285737156868, + 0.005516829434782267, + 0.005280451383441687, + 0.006552475970238447, + 0.0076903668232262135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.005260079633444548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093838855624199 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093838855624199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093838855624199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13986919820308685, + 0.1350853145122528, + 0.08771622180938721, + 0.07188134640455246, + 0.06935499608516693, + 0.24010170996189117, + 0.0956527441740036, + 0.030481208115816116, + 0.07966785877943039, + 0.05018946900963783 + ], + "credit_map": [ + 0.015518464148044586, + 0.014834223315119743, + 0.027073903009295464, + 0.015309851616621017, + 0.00903630442917347, + 0.02481658011674881, + 0.014874495565891266, + 0.018405362963676453, + 0.017064154148101807, + 0.017132310196757317, + 0.013200944289565086, + 0.019306326285004616, + 0.01739218831062317, + 0.014538963325321674, + 0.017264196649193764, + 0.019186222925782204, + 0.011340363882482052, + 0.00889099482446909, + 0.00979627575725317, + 0.010054678656160831, + 0.013647735118865967, + 0.011819143779575825, + 0.013061138801276684, + 0.009105878882110119, + 0.004736713133752346, + 0.0066845351830124855, + 0.01094602607190609, + 0.012715362943708897, + 0.006659670267254114, + 0.007573103066533804, + 0.010030614212155342, + 0.01253531314432621, + 0.006049132905900478, + 0.006173004396259785, + 0.008409399539232254, + 0.006561206188052893, + 0.009551120921969414, + 0.01336216926574707, + 0.009056705050170422, + 0.01019225176423788, + 0.029825666919350624, + 0.07039918750524521, + 0.018918810412287712, + 0.015391841530799866, + 0.020918944850564003, + 0.02456405572593212, + 0.015636827796697617, + 0.04444634169340134, + 0.010306235402822495, + 0.031109297648072243, + 0.015728700906038284, + 0.007724539376795292, + 0.007506709545850754, + 0.011450284160673618, + 0.005242404993623495, + 0.006584562826901674, + 0.003917591646313667, + 0.003238516394048929, + 0.0035215963143855333, + 0.004885764792561531, + 0.003388256300240755, + 0.0032547335140407085, + 0.004378459881991148, + 0.0038962876424193382, + 0.005794336553663015, + 0.008517300710082054, + 0.009191285818815231, + 0.006775632966309786, + 0.0063899424858391285, + 0.02176920883357525, + 0.00638630660250783, + 0.014843838289380074, + 0.006413565017282963, + 0.009502132423222065, + 0.008305207826197147, + 0.00414971262216568, + 0.004748405423015356, + 0.0040087890811264515, + 0.005273974500596523, + 0.007787677925080061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.004093838855624199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583007484674454 + ], + "opposite_logprob_tokens": [ + -3.895582914352417 + ], + "logprob_reward": -0.020583007484674454, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583007484674454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1510559618473053, + 0.049411118030548096, + 0.058574482798576355, + 0.08505517989397049, + 0.08628391474485397, + 0.09784220159053802, + 0.046893443912267685, + 0.08361852169036865, + 0.3059249222278595, + 0.035340264439582825 + ], + "credit_map": [ + 0.01830376870930195, + 0.018038034439086914, + 0.027870330959558487, + 0.015629328787326813, + 0.008561896160244942, + 0.026429740712046623, + 0.01723456382751465, + 0.018988292664289474, + 0.0065858992747962475, + 0.0064304848201572895, + 0.0060819112695753574, + 0.0066751642152667046, + 0.005885581951588392, + 0.0062597207725048065, + 0.005853419192135334, + 0.0056389328092336655, + 0.007550499867647886, + 0.006350731011480093, + 0.0080567616969347, + 0.007537467405200005, + 0.00771230086684227, + 0.00712073128670454, + 0.008182967081665993, + 0.00606301287189126, + 0.00503908796235919, + 0.007393174339085817, + 0.012202384881675243, + 0.014347492717206478, + 0.007048639003187418, + 0.009100573137402534, + 0.014507665298879147, + 0.015416153706610203, + 0.007731038145720959, + 0.005556338932365179, + 0.013688823208212852, + 0.007477280683815479, + 0.009802019223570824, + 0.011227762326598167, + 0.014607386663556099, + 0.016193263232707977, + 0.011724662967026234, + 0.029664775356650352, + 0.007839690893888474, + 0.005663127172738314, + 0.00795032549649477, + 0.01100756786763668, + 0.005740896333009005, + 0.018251148983836174, + 0.0042618652805686, + 0.013516532257199287, + 0.008145720697939396, + 0.0047231935895979404, + 0.0042703161016106606, + 0.005989657249301672, + 0.0027341055683791637, + 0.003252046648412943, + 0.0111260199919343, + 0.009482992812991142, + 0.008989806286990643, + 0.012457400560379028, + 0.008537069894373417, + 0.011875388212502003, + 0.010777821764349937, + 0.010372006334364414, + 0.019028112292289734, + 0.02749484032392502, + 0.032696232199668884, + 0.02157079242169857, + 0.023806093260645866, + 0.0898360162973404, + 0.012141675688326359, + 0.07935113459825516, + 0.0029880062211304903, + 0.004435495473444462, + 0.007452191319316626, + 0.004874036181718111, + 0.003948764409869909, + 0.0030867026653140783, + 0.00385499419644475, + 0.004700071644037962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.020583007484674454 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, wide shot, the camera panning in a smooth, 360-degree orbit, captures an epic, surreal spectacle from every angle. Amidst a velvety, inky night sky dotted with distant stars, countless fiery meteors streaked like a torrential rain, trailing brilliant white and gold trails and tearing through the night from every angle. Their light briefly and dramatically illuminated a vast parametric holographic projection of a storm below, composed of countless shimmering blue and white light particles. Within this complex, translucent three-dimensional structure, simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors were clearly visible, the light particles undulating subtly as the meteors passed. At the very core of the holographic projection, an ominous crimson glow pulsed steadily and menacingly, its light piercing through layers of data and creating a stark, unsettling contrast against the calm blue and white tones of the surroundings. As the camera continues to move, each angle reveals new details about this meteorological spectacle where the universe and technology intertwine, creating an atmosphere that is both magnificent and awe-inspiring, yet also full of potential threats.", + "seed": 8752, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_3/index_7301_seed_8752_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": -0.321312502771616, + "vlm_logprob": -0.321312502771616, + "vlm_yes_no_margin": 2.175000011920929, + "vlm_token_credit": -0.321312502771616, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.3288183212280273 + ], + "opposite_logprob_tokens": [ + -0.3288183808326721 + ], + "logprob_reward": -1.3288183212280273, + "yes_no_margin_reward": -0.9999999403953552, + "token_credit_reward": -1.3288183212280273, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03680373355746269, + 0.5072900652885437, + 0.12076548486948013, + 0.019033940508961678, + 0.23662403225898743, + 0.011991421692073345, + 0.010232435539364815, + 0.020123662427067757, + 0.02887766622006893, + 0.008257529698312283 + ], + "credit_map": [ + 0.008197274059057236, + 0.0045826430432498455, + 0.0035244303289800882, + 0.0028943477664142847, + 0.004012543708086014, + 0.005914084147661924, + 0.003155319718644023, + 0.004523089621216059, + 0.05429510027170181, + 0.03436236456036568, + 0.12547433376312256, + 0.05593693628907204, + 0.04431625083088875, + 0.05445362254977226, + 0.0347888208925724, + 0.10366262495517731, + 0.008653313852846622, + 0.016838546842336655, + 0.013299706391990185, + 0.023312842473387718, + 0.008204477839171886, + 0.016167379915714264, + 0.014869711361825466, + 0.019419504329562187, + 0.002339625498279929, + 0.002782539464533329, + 0.0014711110852658749, + 0.0015865341993048787, + 0.004116448573768139, + 0.002937935758382082, + 0.0024232901632785797, + 0.0013764550676569343, + 0.028704337775707245, + 0.030852273106575012, + 0.02428458444774151, + 0.01683094911277294, + 0.04855140298604965, + 0.050558824092149734, + 0.013589214533567429, + 0.023252442479133606, + 0.0018622983479872346, + 0.001284307800233364, + 0.0006745363352820277, + 0.0012059154687449336, + 0.0019118305062875152, + 0.0024994155392050743, + 0.001545066712424159, + 0.0010080498177558184, + 0.0014208009233698249, + 0.0026464187540113926, + 0.0007210335461422801, + 0.0008749451371841133, + 0.0010124422842636704, + 0.0011520240223035216, + 0.0009076021378859878, + 0.0014971692580729723, + 0.002250988967716694, + 0.00250523560680449, + 0.0011947221355512738, + 0.003041860181838274, + 0.0033992247190326452, + 0.002414467977359891, + 0.003200436243787408, + 0.0021167269442230463, + 0.0029633648227900267, + 0.010310804471373558, + 0.0007575618801638484, + 0.0006743925623595715, + 0.0013285432942211628, + 0.011249346658587456, + 0.000873856944963336, + 0.0007197947707027197, + 0.0011772109428420663, + 0.0010952380253002048, + 0.0008681080071255565, + 0.0007177077932283282, + 0.0015288727590814233, + 0.0011113797081634402, + 0.0006839570123702288, + 0.0010750554502010345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -1.3288183212280273 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909006416797638 + ], + "opposite_logprob_tokens": [ + -2.46409010887146 + ], + "logprob_reward": -0.08909006416797638, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909006416797638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03403730317950249, + 0.48752790689468384, + 0.0232095904648304, + 0.059472616761922836, + 0.21984891593456268, + 0.024280991405248642, + 0.03934226185083389, + 0.02381991781294346, + 0.03694263473153114, + 0.0515177920460701 + ], + "credit_map": [ + 0.005295184440910816, + 0.003238074481487274, + 0.004777910187840462, + 0.002788018202409148, + 0.0037128038238734007, + 0.006142838392406702, + 0.0030035730451345444, + 0.0050789047963917255, + 0.05297662690281868, + 0.0346091203391552, + 0.11859189718961716, + 0.05091298371553421, + 0.04575871303677559, + 0.052471816539764404, + 0.033083584159612656, + 0.09912320226430893, + 0.0019412488909438252, + 0.003464748151600361, + 0.002561273518949747, + 0.0033123947214335203, + 0.0033922914881259203, + 0.0031184686813503504, + 0.0022060470655560493, + 0.003213119227439165, + 0.007326388265937567, + 0.009480494074523449, + 0.002780774375423789, + 0.004160652402788401, + 0.014873822219669819, + 0.010171689093112946, + 0.00657395925372839, + 0.004104845225811005, + 0.02697606384754181, + 0.028601370751857758, + 0.02280036173760891, + 0.01504061184823513, + 0.04562438279390335, + 0.04754973575472832, + 0.012094720266759396, + 0.02116168662905693, + 0.003167580347508192, + 0.002264553215354681, + 0.0018026507459580898, + 0.002905273577198386, + 0.003502943553030491, + 0.0044068023562431335, + 0.0037046687211841345, + 0.002526521682739258, + 0.005384466145187616, + 0.012183956801891327, + 0.001865743426606059, + 0.0031902340706437826, + 0.005817989818751812, + 0.005683259107172489, + 0.00248903501778841, + 0.0027275800239294767, + 0.0032547316513955593, + 0.0074311974458396435, + 0.0017548375762999058, + 0.001676637213677168, + 0.001887406688183546, + 0.003364172065630555, + 0.0019758171401917934, + 0.002475121058523655, + 0.004485928453505039, + 0.012098840437829494, + 0.0007217741804197431, + 0.0008076966041699052, + 0.0015980337047949433, + 0.014923105016350746, + 0.001003013108856976, + 0.0013042447390034795, + 0.00696520833298564, + 0.0075069693848490715, + 0.00463763065636158, + 0.0037986058741807938, + 0.012592834420502186, + 0.008009644225239754, + 0.003889933694154024, + 0.004116971045732498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.08909006416797638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069423407316208 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069423407316208, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069423407316208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1322908103466034, + 0.25393912196159363, + 0.1880308836698532, + 0.01636352203786373, + 0.13729213178157806, + 0.02436455897986889, + 0.030172839760780334, + 0.11753736436367035, + 0.07762830704450607, + 0.022380515933036804 + ], + "credit_map": [ + 0.025000296533107758, + 0.013434247113764286, + 0.018866216763854027, + 0.012390109710395336, + 0.0076466635800898075, + 0.021328900009393692, + 0.00864317175000906, + 0.024981200695037842, + 0.02704791910946369, + 0.017596008256077766, + 0.06275450438261032, + 0.027394693344831467, + 0.022459521889686584, + 0.027256185188889503, + 0.017449762672185898, + 0.051980502903461456, + 0.013045222498476505, + 0.026170598343014717, + 0.019498104229569435, + 0.03693640977144241, + 0.012167717330157757, + 0.025065356865525246, + 0.023881390690803528, + 0.03126607462763786, + 0.0015362348640337586, + 0.0029371962882578373, + 0.001121563371270895, + 0.0015237777261063457, + 0.003782590851187706, + 0.002441194374114275, + 0.0018175893928855658, + 0.001203375170007348, + 0.017150605097413063, + 0.017859583720564842, + 0.014560900628566742, + 0.010022269561886787, + 0.02604668028652668, + 0.029893681406974792, + 0.007858258672058582, + 0.01390014123171568, + 0.0030246106907725334, + 0.0021699077915400267, + 0.0029452245216816664, + 0.0038868195842951536, + 0.0026201107539236546, + 0.0031867804937064648, + 0.0028851665556430817, + 0.0036459355615079403, + 0.004300675820559263, + 0.009085327386856079, + 0.0013540846994146705, + 0.0024409808684140444, + 0.004171092063188553, + 0.00407665129750967, + 0.0015738221118226647, + 0.0031702029518783092, + 0.019052376970648766, + 0.04225144907832146, + 0.004461327567696571, + 0.0060122511349618435, + 0.00785822793841362, + 0.017422828823328018, + 0.0057384781539440155, + 0.014740411192178726, + 0.009121285751461983, + 0.026348654180765152, + 0.0013283430598676205, + 0.001667371136136353, + 0.0030944293830543756, + 0.03177202120423317, + 0.0020620578434318304, + 0.0022341380827128887, + 0.003183412365615368, + 0.00376882404088974, + 0.0016379603184759617, + 0.0014560670824721456, + 0.005523496773093939, + 0.0030740585643798113, + 0.001559299766086042, + 0.002177394460886717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.016069423407316208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.4591141939163208, + 0.06964147835969925, + 0.08668073266744614, + 0.017507554963231087, + 0.24475885927677155, + 0.00379738537594676, + 0.013614590279757977, + 0.0803612768650055, + 0.011888074688613415, + 0.012635856866836548 + ], + "credit_map": [ + 0.08604267984628677, + 0.04982709512114525, + 0.0605773888528347, + 0.040966447442770004, + 0.031060054898262024, + 0.07500764727592468, + 0.030079204589128494, + 0.08555366843938828, + 0.0074978903867304325, + 0.005050265695899725, + 0.017152830958366394, + 0.007198231760412455, + 0.006233263295143843, + 0.007393076550215483, + 0.005087474826723337, + 0.014028443954885006, + 0.006102466490119696, + 0.012562488205730915, + 0.008346601389348507, + 0.016941508278250694, + 0.007299554068595171, + 0.011016709730029106, + 0.01077242847532034, + 0.013638976961374283, + 0.0022355883847922087, + 0.003305742982774973, + 0.0013805071357637644, + 0.0012941864551976323, + 0.004076510202139616, + 0.002646825974807143, + 0.0013811610406264663, + 0.0011870324378833175, + 0.028871670365333557, + 0.03087475337088108, + 0.026242079213261604, + 0.016836853697896004, + 0.053965408354997635, + 0.05104902386665344, + 0.013450558297336102, + 0.023468511179089546, + 0.0005159831489436328, + 0.0004134529735893011, + 0.0003793701180256903, + 0.00046017265412956476, + 0.00042876292718574405, + 0.000672384281642735, + 0.0005307544488459826, + 0.00039650488179177046, + 0.0018162858905270696, + 0.003656518878415227, + 0.0010818728478625417, + 0.0016929346602410078, + 0.0015985475620254874, + 0.0015999076422303915, + 0.0009686034172773361, + 0.001199919730424881, + 0.013226889073848724, + 0.028726628050208092, + 0.0030481002759188414, + 0.003899323521181941, + 0.00481403386220336, + 0.011537973769009113, + 0.004045020788908005, + 0.01106330193579197, + 0.0016078848857432604, + 0.0017242870526388288, + 0.001383238355629146, + 0.0009443784947507083, + 0.0014941688859835267, + 0.0022842567414045334, + 0.0010987624991685152, + 0.0013510981807485223, + 0.0018743535038083792, + 0.0023679533042013645, + 0.0009872850496321917, + 0.0009111798135563731, + 0.0027056618127971888, + 0.0017062164843082428, + 0.0010036579333245754, + 0.001079548499546945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.029788486659526825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera executing a smooth 360-degree orbit around the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are countless fiery meteors streaking across a velvety, inky night sky dotted with distant stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a vast parametric holographic projection of a storm visible, composed of shimmering blue and white light particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is an ominous crimson glow visible at the core of the holographic projection?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are simulated swirling air currents, a precise grid of isobars, and pulsating lightning vectors visible within the holographic structure?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279621839523315 + ], + "opposite_logprob_tokens": [ + -2.017796277999878 + ], + "logprob_reward": -0.14279621839523315, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279621839523315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24182568490505219, + 0.18954864144325256, + 0.08232404291629791, + 0.035949304699897766, + 0.17010602355003357, + 0.030179493129253387, + 0.032247040420770645, + 0.10259350389242172, + 0.0367836132645607, + 0.07844264805316925 + ], + "credit_map": [ + 0.0544346421957016, + 0.025364775210618973, + 0.029712727293372154, + 0.019204560667276382, + 0.016517730429768562, + 0.04116859287023544, + 0.013050478883087635, + 0.0423722006380558, + 0.01957741566002369, + 0.01444541197270155, + 0.04607285186648369, + 0.020141838118433952, + 0.01949230581521988, + 0.019934019073843956, + 0.012700487859547138, + 0.03718430548906326, + 0.006242967210710049, + 0.010806490667164326, + 0.009469081647694111, + 0.015459229238331318, + 0.006437778938561678, + 0.010444278828799725, + 0.009740261361002922, + 0.013723956421017647, + 0.004125975538045168, + 0.007565586362034082, + 0.003312801942229271, + 0.003187937894836068, + 0.004730772227048874, + 0.006163196638226509, + 0.004415509756654501, + 0.00244752480648458, + 0.020163103938102722, + 0.023364556953310966, + 0.016789548099040985, + 0.012278194539248943, + 0.035656318068504333, + 0.03651345148682594, + 0.010598411783576012, + 0.014742457307875156, + 0.0039357407949864864, + 0.002715528942644596, + 0.002543342998251319, + 0.004279150161892176, + 0.0044722831808030605, + 0.00535348616540432, + 0.0034740816336125135, + 0.0034058799501508474, + 0.004567554220557213, + 0.00980154424905777, + 0.0015106040518730879, + 0.0025731881614774466, + 0.004466417711228132, + 0.004466380923986435, + 0.00173420540522784, + 0.0031271486077457666, + 0.016416149213910103, + 0.03656226396560669, + 0.0038525108247995377, + 0.0053793261758983135, + 0.007341964635998011, + 0.015160834416747093, + 0.005153505131602287, + 0.012726962566375732, + 0.006399204023182392, + 0.007730781566351652, + 0.0013347197091206908, + 0.0012478480348363519, + 0.0027606580406427383, + 0.01284381840378046, + 0.0016644243150949478, + 0.0028021596372127533, + 0.011078126728534698, + 0.012247370555996895, + 0.008339791558682919, + 0.005966972094029188, + 0.0156794935464859, + 0.008860009722411633, + 0.005989202298223972, + 0.010281681083142757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.14279621839523315 + } + ], + "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/0000055/group_4/index_tensor([2723, 2723])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_tensor([2723, 2723])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..649adac84f3276f6fc7377393d9152dc18eb48e4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_tensor([2723, 2723])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Ultra-slow-motion, abstract macrophotographic static close-ups capture a moment of controlled chaos under dramatic, high-contrast studio lighting. A heavy, polished steel spoon, its perfectly reflective bowl and elegant stem, captured in sharp, mirror-like highlights, succumbs to an unseen tremor. It deliberately tips over, its weight palpable as it begins its descent. Beneath it, a pool of viscous, opaque white paint spreads against a dark, perfectly matte background, its surface as smooth and still as glass. The spoon's thin leading edge makes contact first, plunging into the liquid with surgical precision. In extreme slow motion, the paint's surface dimples, resisting for a fraction of a second, then collapses as the spoon carves a clean, vertical path downward. This displacement creates a thick, rolling wave of white paint, cresting and curling outward, its texture similar to that of heavy cream, forming a perfect, silent, crown-like splash. High-contrast lighting shapes the scene, casting the background in deep shadow while illuminating the brilliant white of the paint and the cold metallic sheen of the partially submerged spoon, creating a spectacle that is tense, captivating, and purely physical.", + "seed": 8749, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_2723_seed_8749_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": -0.09395696918945759, + "vlm_logprob": -0.09395696918945759, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.09395696918945759, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06371927261352539 + ], + "opposite_logprob_tokens": [ + -2.8137192726135254 + ], + "logprob_reward": -0.06371927261352539, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06371927261352539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11477277427911758, + 0.034570060670375824, + 0.07511747628450394, + 0.14133860170841217, + 0.12998968362808228, + 0.049840524792671204, + 0.056400395929813385, + 0.1987791806459427, + 0.11366923898458481, + 0.08552201092243195 + ], + "credit_map": [ + 0.01590287685394287, + 0.01048904750496149, + 0.020681511610746384, + 0.016517043113708496, + 0.010793205350637436, + 0.013868783600628376, + 0.017328260466456413, + 0.009192047640681267, + 0.003892880864441395, + 0.005256184842437506, + 0.0053331670351326466, + 0.003063302719965577, + 0.005411540158092976, + 0.004006929229944944, + 0.004042116925120354, + 0.0035639412235468626, + 0.00631840992718935, + 0.008687571622431278, + 0.0073296185582876205, + 0.00527542270720005, + 0.011961490847170353, + 0.012342461384832859, + 0.012856727465987206, + 0.01034577563405037, + 0.014284863136708736, + 0.012989428825676441, + 0.02069883607327938, + 0.017005998641252518, + 0.024695705622434616, + 0.02466694824397564, + 0.01495286263525486, + 0.012043946422636509, + 0.010072458535432816, + 0.013602814637124538, + 0.02168162539601326, + 0.013623138889670372, + 0.021051790565252304, + 0.02095000445842743, + 0.015583670698106289, + 0.01342417486011982, + 0.0036763011012226343, + 0.006021607667207718, + 0.009563974104821682, + 0.008742013014853, + 0.00535613764077425, + 0.005958821624517441, + 0.00586592685431242, + 0.004655745346099138, + 0.004523091949522495, + 0.00830157846212387, + 0.008826547302305698, + 0.009802118875086308, + 0.008964327163994312, + 0.00525917112827301, + 0.006769951432943344, + 0.003953611943870783, + 0.00915239006280899, + 0.015016432851552963, + 0.0301443412899971, + 0.07374489307403564, + 0.009832385927438736, + 0.011503043584525585, + 0.03853826969861984, + 0.010847419500350952, + 0.010517408140003681, + 0.011619843542575836, + 0.013417249545454979, + 0.014002889394760132, + 0.01661751978099346, + 0.019795307889580727, + 0.016380589455366135, + 0.011318434961140156, + 0.007488214410841465, + 0.022141417488455772, + 0.007352353539317846, + 0.005854440852999687, + 0.016510004177689552, + 0.009639935567975044, + 0.007719564717262983, + 0.008816072717308998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.06371927261352539 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870641887187958 + ], + "opposite_logprob_tokens": [ + -1.1370642185211182 + ], + "logprob_reward": -0.3870641887187958, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870641887187958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46229562163352966, + 0.059773314744234085, + 0.05822979658842087, + 0.05307440832257271, + 0.07397342473268509, + 0.04029853269457817, + 0.06250595301389694, + 0.06832707673311234, + 0.0933697521686554, + 0.02815212309360504 + ], + "credit_map": [ + 0.07313724607229233, + 0.04893719032406807, + 0.09314806759357452, + 0.0453275665640831, + 0.07517925649881363, + 0.05652065575122833, + 0.03285389021039009, + 0.037191782146692276, + 0.0060854884795844555, + 0.006319835316389799, + 0.009635558351874352, + 0.00682630343362689, + 0.010095421224832535, + 0.006623789202421904, + 0.006901951972395182, + 0.007284966297447681, + 0.005323640536516905, + 0.004226212855428457, + 0.012292896397411823, + 0.0069437152706086636, + 0.005865380633622408, + 0.006685622502118349, + 0.010729077272117138, + 0.006163245998322964, + 0.007034733891487122, + 0.006449692416936159, + 0.012327451258897781, + 0.007984637282788754, + 0.005748530849814415, + 0.005638147704303265, + 0.004702271893620491, + 0.003188943723216653, + 0.005461629014462233, + 0.0055883475579321384, + 0.01529849972575903, + 0.015046974644064903, + 0.010561604052782059, + 0.008600770495831966, + 0.007050753571093082, + 0.006364842876791954, + 0.002736643422394991, + 0.0047597396187484264, + 0.00681900093331933, + 0.0052920617163181305, + 0.007265390362590551, + 0.005826142616569996, + 0.005015449598431587, + 0.002584105357527733, + 0.004924447275698185, + 0.009679069742560387, + 0.00951200071722269, + 0.008870230987668037, + 0.009306240826845169, + 0.006843095179647207, + 0.007854818366467953, + 0.005516047589480877, + 0.0030739877838641405, + 0.008392236195504665, + 0.008344217203557491, + 0.01895015500485897, + 0.008255247958004475, + 0.007450767792761326, + 0.010057756677269936, + 0.0038027099799364805, + 0.009369250386953354, + 0.01069430261850357, + 0.011899618431925774, + 0.010948535986244678, + 0.010919326916337013, + 0.017802828922867775, + 0.012049690820276737, + 0.009686199016869068, + 0.002937438664957881, + 0.0027672434225678444, + 0.004549880977720022, + 0.004724038764834404, + 0.002652153605595231, + 0.0023282309994101524, + 0.004521895665675402, + 0.0036712423898279667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.3870641887187958 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645840920507908 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008645840920507908, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645840920507908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09963994473218918, + 0.09214083850383759, + 0.061810899525880814, + 0.058642882853746414, + 0.07679454982280731, + 0.06764046102762222, + 0.11259741336107254, + 0.174153134226799, + 0.13256628811359406, + 0.12401361763477325 + ], + "credit_map": [ + 0.009720182046294212, + 0.010192780755460262, + 0.009723438881337643, + 0.016113704070448875, + 0.014564556069672108, + 0.00942434836179018, + 0.01820395328104496, + 0.011696981266140938, + 0.010180065408349037, + 0.009490108117461205, + 0.021686438471078873, + 0.009883414022624493, + 0.012090995907783508, + 0.009674828499555588, + 0.010163712315261364, + 0.008971271105110645, + 0.0068880063481628895, + 0.007083815056830645, + 0.009944958612322807, + 0.006429648492485285, + 0.007358700968325138, + 0.008230273611843586, + 0.008374705910682678, + 0.007500786799937487, + 0.006997732911258936, + 0.009824634529650211, + 0.008602931164205074, + 0.006968715228140354, + 0.00776003347709775, + 0.007269424386322498, + 0.006617074366658926, + 0.004602334927767515, + 0.0068192551843822, + 0.009843816049396992, + 0.012354327365756035, + 0.013781554065644741, + 0.009066282771527767, + 0.012834513559937477, + 0.007410549558699131, + 0.004684241954237223, + 0.006630785763263702, + 0.006842698901891708, + 0.010966653935611248, + 0.008761799894273281, + 0.008173095062375069, + 0.011892730370163918, + 0.006671222858130932, + 0.007701471447944641, + 0.01074922177940607, + 0.0124060632660985, + 0.02208571322262287, + 0.018636683002114296, + 0.011822852306067944, + 0.013964749872684479, + 0.014055547304451466, + 0.00887658167630434, + 0.007962223142385483, + 0.015411454252898693, + 0.022117288783192635, + 0.05871746316552162, + 0.012556412257254124, + 0.01793714240193367, + 0.03049449436366558, + 0.008956645615398884, + 0.013919804245233536, + 0.014602197334170341, + 0.015656644478440285, + 0.017754865810275078, + 0.014786558225750923, + 0.02649836800992489, + 0.015282656997442245, + 0.014065176248550415, + 0.011230587027966976, + 0.032275184988975525, + 0.00946715846657753, + 0.009427325800061226, + 0.02385891228914261, + 0.013301688246428967, + 0.01222148071974516, + 0.01223127543926239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.008645840920507908 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640758693218231 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640758693218231, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640758693218231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12685470283031464, + 0.0989084243774414, + 0.12870989739894867, + 0.05335346236824989, + 0.08954126387834549, + 0.08838581293821335, + 0.09014102071523666, + 0.19049887359142303, + 0.08237601071596146, + 0.05123056843876839 + ], + "credit_map": [ + 0.01594601571559906, + 0.014006190933287144, + 0.017230739817023277, + 0.014246255159378052, + 0.017666244879364967, + 0.01697581261396408, + 0.021688174456357956, + 0.009095271117985249, + 0.011714374646544456, + 0.009082271717488766, + 0.017994536086916924, + 0.010544689372181892, + 0.013419357128441334, + 0.011518474668264389, + 0.012915787287056446, + 0.0117189297452569, + 0.012680696323513985, + 0.011817350052297115, + 0.019272271543741226, + 0.012155629694461823, + 0.019622590392827988, + 0.02411489561200142, + 0.016760535538196564, + 0.012285922653973103, + 0.005398845300078392, + 0.005518133752048016, + 0.013553420081734657, + 0.008265587501227856, + 0.00538950739428401, + 0.0052411784417927265, + 0.005569778848439455, + 0.004417010582983494, + 0.01093385275453329, + 0.0172971710562706, + 0.012555887922644615, + 0.013072819449007511, + 0.008411469869315624, + 0.010326432064175606, + 0.011063244193792343, + 0.005880389828234911, + 0.00835161842405796, + 0.013443877920508385, + 0.010826962068676949, + 0.008323941379785538, + 0.014605136588215828, + 0.014048822224140167, + 0.010653139092028141, + 0.00813231896609068, + 0.006060934159904718, + 0.022291600704193115, + 0.011669538915157318, + 0.01053747907280922, + 0.01876833662390709, + 0.008101959712803364, + 0.00829189084470272, + 0.004419277422130108, + 0.010322773829102516, + 0.014899085275828838, + 0.028003497049212456, + 0.06428053975105286, + 0.0116767892614007, + 0.016173506155610085, + 0.03409108519554138, + 0.011051591485738754, + 0.008524352684617043, + 0.007535135839134455, + 0.012050948105752468, + 0.008815720677375793, + 0.009087998420000076, + 0.019799571484327316, + 0.007342447992414236, + 0.009219838306307793, + 0.004161786753684282, + 0.008411608636379242, + 0.005406325217336416, + 0.00544889597222209, + 0.006090944167226553, + 0.008648599497973919, + 0.006329826079308987, + 0.006732582114636898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.008640758693218231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017147850012406707 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017147850012406707, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017147850012406707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10903037339448929, + 0.03593830764293671, + 0.11864714324474335, + 0.044298332184553146, + 0.04294740408658981, + 0.03648381680250168, + 0.12825177609920502, + 0.42784205079078674, + 0.03632201626896858, + 0.020238781347870827 + ], + "credit_map": [ + 0.01805141754448414, + 0.011106948368251324, + 0.02394375391304493, + 0.01010938361287117, + 0.016387036070227623, + 0.01264192909002304, + 0.009341058321297169, + 0.007448842283338308, + 0.004051780793815851, + 0.004582306370139122, + 0.006625997833907604, + 0.0038930196315050125, + 0.003844891209155321, + 0.004859070759266615, + 0.0042533790692687035, + 0.0038278631400316954, + 0.009763947688043118, + 0.009122856892645359, + 0.022702133283019066, + 0.012043869122862816, + 0.01837666891515255, + 0.01751100644469261, + 0.01722494699060917, + 0.011901716701686382, + 0.0038516020867973566, + 0.0045342217199504375, + 0.010531122796237469, + 0.007626074831932783, + 0.005103680305182934, + 0.005279424600303173, + 0.0038787024095654488, + 0.0034935076255351305, + 0.003952761180698872, + 0.005301335826516151, + 0.008970008231699467, + 0.006451803259551525, + 0.005840027704834938, + 0.0052759721875190735, + 0.004257736727595329, + 0.002897758735343814, + 0.0036095345858484507, + 0.004113321658223867, + 0.007047930732369423, + 0.005318948999047279, + 0.005115616600960493, + 0.004922852851450443, + 0.003866031300276518, + 0.0024895782116800547, + 0.012334408238530159, + 0.01378650963306427, + 0.022363530471920967, + 0.020817862823605537, + 0.01562872715294361, + 0.019060853868722916, + 0.015532094053924084, + 0.008727798238396645, + 0.020740192383527756, + 0.03114306554198265, + 0.06362470239400864, + 0.16081267595291138, + 0.019189544022083282, + 0.025440162047743797, + 0.08443208038806915, + 0.022459642961621284, + 0.002351312665268779, + 0.008363311178982258, + 0.002889875089749694, + 0.005735236220061779, + 0.006665775552392006, + 0.004660665523260832, + 0.0035651354119181633, + 0.0020907041616737843, + 0.0021301244851201773, + 0.0037600903306156397, + 0.0021135960705578327, + 0.001953021390363574, + 0.0032250017393380404, + 0.003336583497002721, + 0.001973559847101569, + 0.0017468027072027326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.0017147850012406707 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-slow-motion, abstract macrophotographic static close-ups capture a moment of controlled chaos under dramatic, high-contrast studio lighting. A heavy, polished steel spoon, its perfectly reflective bowl and elegant stem, captured in sharp, mirror-like highlights, succumbs to an unseen tremor. It deliberately tips over, its weight palpable as it begins its descent. Beneath it, a pool of viscous, opaque white paint spreads against a dark, perfectly matte background, its surface as smooth and still as glass. The spoon's thin leading edge makes contact first, plunging into the liquid with surgical precision. In extreme slow motion, the paint's surface dimples, resisting for a fraction of a second, then collapses as the spoon carves a clean, vertical path downward. This displacement creates a thick, rolling wave of white paint, cresting and curling outward, its texture similar to that of heavy cream, forming a perfect, silent, crown-like splash. High-contrast lighting shapes the scene, casting the background in deep shadow while illuminating the brilliant white of the paint and the cold metallic sheen of the partially submerged spoon, creating a spectacle that is tense, captivating, and purely physical.", + "seed": 8750, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_2723_seed_8750_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": -0.09501558158081025, + "vlm_logprob": -0.09501558158081025, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.09501558158081025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1832817643880844 + ], + "opposite_logprob_tokens": [ + -1.8082817792892456 + ], + "logprob_reward": -0.1832817643880844, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1832817643880844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16685616970062256, + 0.10768494009971619, + 0.03440168499946594, + 0.25360384583473206, + 0.0804547369480133, + 0.1263614296913147, + 0.031223012134432793, + 0.1460268646478653, + 0.024789996445178986, + 0.02859724499285221 + ], + "credit_map": [ + 0.02011997625231743, + 0.01677214726805687, + 0.022448483854532242, + 0.02512417361140251, + 0.032262880355119705, + 0.01322876289486885, + 0.020525159314274788, + 0.016374580562114716, + 0.030151627957820892, + 0.017869090661406517, + 0.009563780389726162, + 0.008729418739676476, + 0.009134483523666859, + 0.012001083232462406, + 0.009163289330899715, + 0.01107217650860548, + 0.006014807149767876, + 0.003614705754444003, + 0.00863453559577465, + 0.004011816810816526, + 0.004167838022112846, + 0.002831535181030631, + 0.0035169923212379217, + 0.0016094569582492113, + 0.02781561017036438, + 0.025674695149064064, + 0.03186650574207306, + 0.03737187385559082, + 0.04872367903590202, + 0.021733039990067482, + 0.046057309955358505, + 0.014361148700118065, + 0.012623727321624756, + 0.005333655513823032, + 0.025358203798532486, + 0.012942930683493614, + 0.011315176263451576, + 0.005221949424594641, + 0.0050542899407446384, + 0.002604809822514653, + 0.03236609324812889, + 0.005664742551743984, + 0.028193388134241104, + 0.01567166857421398, + 0.009366577491164207, + 0.025600718334317207, + 0.00480496184900403, + 0.004693281836807728, + 0.0037913592532277107, + 0.002676951000466943, + 0.0035515634808689356, + 0.005333398003131151, + 0.00543670728802681, + 0.0032817835453897715, + 0.003342690411955118, + 0.0038085600826889277, + 0.010914408601820469, + 0.011860214173793793, + 0.013614005409181118, + 0.02580534853041172, + 0.016384324058890343, + 0.010992375202476978, + 0.03256112337112427, + 0.02389506995677948, + 0.0022933506406843662, + 0.0021457623224705458, + 0.002397579839453101, + 0.004507836885750294, + 0.0030775568448007107, + 0.0023218723945319653, + 0.004191123880445957, + 0.003854915965348482, + 0.0029000581707805395, + 0.002581645967438817, + 0.0032219933345913887, + 0.0046061137691140175, + 0.004635290242731571, + 0.0023445719853043556, + 0.004506902303546667, + 0.003800670150667429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.1832817643880844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28133654594421387 + ], + "opposite_logprob_tokens": [ + -1.4063365459442139 + ], + "logprob_reward": -0.28133654594421387, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28133654594421387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09010500460863113, + 0.10191521048545837, + 0.1114824041724205, + 0.058937039226293564, + 0.04274370148777962, + 0.4717934727668762, + 0.029250135645270348, + 0.05510222911834717, + 0.01648869924247265, + 0.022182052955031395 + ], + "credit_map": [ + 0.02322930470108986, + 0.005515203811228275, + 0.00724862702190876, + 0.006822928786277771, + 0.008489660918712616, + 0.024149004369974136, + 0.007634659297764301, + 0.007015634328126907, + 0.022119421511888504, + 0.032608333975076675, + 0.009616680443286896, + 0.008471901528537273, + 0.007190902717411518, + 0.01142818946391344, + 0.005507230293005705, + 0.004972565919160843, + 0.031238727271556854, + 0.037894681096076965, + 0.007742930669337511, + 0.007685662247240543, + 0.005954970605671406, + 0.01086985133588314, + 0.006886452902108431, + 0.003209139918908477, + 0.006805439479649067, + 0.004874619655311108, + 0.006147515960037708, + 0.010331605561077595, + 0.013274136930704117, + 0.006011662073433399, + 0.007286746054887772, + 0.004205323290079832, + 0.005225721746683121, + 0.004646498244255781, + 0.005589335225522518, + 0.007269364316016436, + 0.00885004736483097, + 0.00430777994915843, + 0.0044361744076013565, + 0.002418788382783532, + 0.12818486988544464, + 0.019827859476208687, + 0.10742972046136856, + 0.06181107461452484, + 0.02318037673830986, + 0.10540319979190826, + 0.012710083276033401, + 0.013246355578303337, + 0.0033718144986778498, + 0.0020899525843560696, + 0.003168606897816062, + 0.005591248162090778, + 0.004600995685905218, + 0.0023973898496478796, + 0.004267955664545298, + 0.003762176726013422, + 0.005589938256889582, + 0.005682140588760376, + 0.005102258175611496, + 0.008083479478955269, + 0.00885678455233574, + 0.005633680149912834, + 0.009066914208233356, + 0.0070870416238904, + 0.0011410940205678344, + 0.000987754319794476, + 0.0018910461803898215, + 0.0035191441420465708, + 0.0012823713477700949, + 0.0011635097907856107, + 0.0035514342598617077, + 0.0029523482080549, + 0.001668230746872723, + 0.001542615587823093, + 0.0028168810531497, + 0.004489339888095856, + 0.002367317210882902, + 0.0016469723777845502, + 0.004109448287636042, + 0.0035412500146776438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.28133654594421387 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16194160282611847, + 0.12764957547187805, + 0.11046618968248367, + 0.12185739725828171, + 0.12647691369056702, + 0.11676894128322601, + 0.04575272276997566, + 0.13280437886714935, + 0.023370984941720963, + 0.03291132673621178 + ], + "credit_map": [ + 0.013739255256950855, + 0.012662391178309917, + 0.045452434569597244, + 0.025315061211586, + 0.0177399143576622, + 0.021190229803323746, + 0.012111833319067955, + 0.013730489648878574, + 0.022178880870342255, + 0.03074575588107109, + 0.010338658466935158, + 0.00991667527705431, + 0.005956937558948994, + 0.023216096684336662, + 0.007926386781036854, + 0.017370184883475304, + 0.02911442518234253, + 0.033040814101696014, + 0.011196326464414597, + 0.008094879798591137, + 0.008028487674891949, + 0.010451789945363998, + 0.007500927429646254, + 0.0030385374557226896, + 0.013747074641287327, + 0.01216351892799139, + 0.013286424800753593, + 0.01664668880403042, + 0.02450552023947239, + 0.00940632726997137, + 0.02486918680369854, + 0.007232655305415392, + 0.01303199864923954, + 0.009880774654448032, + 0.026182489469647408, + 0.02636360004544258, + 0.02318466454744339, + 0.009729593992233276, + 0.012874734587967396, + 0.005229051224887371, + 0.035194505006074905, + 0.005821459461003542, + 0.011069200001657009, + 0.007640371099114418, + 0.009660432115197182, + 0.03690243139863014, + 0.004408485721796751, + 0.006072047166526318, + 0.004469158127903938, + 0.003745454363524914, + 0.004943627864122391, + 0.0077624935656785965, + 0.00860518403351307, + 0.0041001043282449245, + 0.005171470809727907, + 0.006955229677259922, + 0.013442990370094776, + 0.013174053281545639, + 0.014049571938812733, + 0.019793430343270302, + 0.025762001052498817, + 0.013497903943061829, + 0.017782453447580338, + 0.015301976352930069, + 0.001558186486363411, + 0.0016328224446624517, + 0.0025494121946394444, + 0.0042989919893443584, + 0.002603999339044094, + 0.0014328298857435584, + 0.005248076282441616, + 0.004046667367219925, + 0.0032723541371524334, + 0.0026232420932501554, + 0.0035042064264416695, + 0.005994515027850866, + 0.0052360291592776775, + 0.0022638968657702208, + 0.005573058035224676, + 0.00444402638822794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099062643945217 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004099062643945217, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099062643945217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2043820172548294, + 0.06583983451128006, + 0.059386007487773895, + 0.07920536398887634, + 0.18055693805217743, + 0.21178773045539856, + 0.07069563865661621, + 0.08249174803495407, + 0.02639223076403141, + 0.019262410700321198 + ], + "credit_map": [ + 0.05061585456132889, + 0.018069811165332794, + 0.019126607105135918, + 0.018506405875086784, + 0.022789383307099342, + 0.05143146589398384, + 0.014319007284939289, + 0.00952350627630949, + 0.012335987761616707, + 0.009293882176280022, + 0.007255297154188156, + 0.006235363427549601, + 0.0083279674872756, + 0.009410147555172443, + 0.005640492308884859, + 0.007340706884860992, + 0.009359053336083889, + 0.013428991660475731, + 0.008376951329410076, + 0.005703802686184645, + 0.006927116308361292, + 0.006363664288073778, + 0.005731235258281231, + 0.0034952068235725164, + 0.008844377472996712, + 0.007990448735654354, + 0.010645609349012375, + 0.011513900011777878, + 0.014145465567708015, + 0.007638842333108187, + 0.014082206413149834, + 0.0043445248156785965, + 0.02417783997952938, + 0.011239541694521904, + 0.06297404319047928, + 0.032642971724271774, + 0.022636881098151207, + 0.009833024814724922, + 0.00983090978115797, + 0.007221762556582689, + 0.045622773468494415, + 0.011681404896080494, + 0.03887299820780754, + 0.03368616849184036, + 0.02089986391365528, + 0.0412130281329155, + 0.01184623409062624, + 0.00796527974307537, + 0.008488498628139496, + 0.006561620160937309, + 0.006938781123608351, + 0.009612216614186764, + 0.018213601782917976, + 0.006820965558290482, + 0.006940738297998905, + 0.007119228597730398, + 0.006666042376309633, + 0.008241510950028896, + 0.008225240744650364, + 0.014913756400346756, + 0.011608987115323544, + 0.0071950205601751804, + 0.013022588565945625, + 0.012618609704077244, + 0.002156573813408613, + 0.0024747687857598066, + 0.003046258119866252, + 0.0049890377558767796, + 0.003105001989752054, + 0.002010882832109928, + 0.004628906492143869, + 0.003980806563049555, + 0.0019537003245204687, + 0.0019096983596682549, + 0.0016928913537412882, + 0.0034293963108211756, + 0.0037131262943148613, + 0.0021862396970391273, + 0.0021600157488137484, + 0.002217345405369997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004099062643945217 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017164511373266578 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.0017164511373266578, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017164511373266578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16416475176811218, + 0.18320836126804352, + 0.16887705028057098, + 0.09891225397586823, + 0.09261421859264374, + 0.1021219789981842, + 0.10508504509925842, + 0.03835121542215347, + 0.03181714192032814, + 0.014848020859062672 + ], + "credit_map": [ + 0.03541911765933037, + 0.016857292503118515, + 0.013402332551777363, + 0.014752291142940521, + 0.01949801854789257, + 0.033764999359846115, + 0.014421937987208366, + 0.01604875549674034, + 0.05105045065283775, + 0.04364959895610809, + 0.017965221777558327, + 0.014409180730581284, + 0.013430090621113777, + 0.012792548164725304, + 0.01416550762951374, + 0.01574576273560524, + 0.03475985676050186, + 0.06088414043188095, + 0.01809409074485302, + 0.011664869263768196, + 0.010289095342159271, + 0.01986776664853096, + 0.007680591195821762, + 0.005636643618345261, + 0.011551174335181713, + 0.011453275568783283, + 0.012650945223867893, + 0.012196624651551247, + 0.022765683010220528, + 0.0071927630342543125, + 0.016205551102757454, + 0.004896231926977634, + 0.010376346297562122, + 0.00797543115913868, + 0.012766345404088497, + 0.019022904336452484, + 0.01835300773382187, + 0.008548006415367126, + 0.010871732607483864, + 0.004700445104390383, + 0.01360436249524355, + 0.009223399683833122, + 0.018298998475074768, + 0.015518124215304852, + 0.018428092822432518, + 0.00931825302541256, + 0.010471127927303314, + 0.007259618490934372, + 0.009698189795017242, + 0.006369548384100199, + 0.018631644546985626, + 0.02691875398159027, + 0.00971180945634842, + 0.006974237039685249, + 0.01120948139578104, + 0.015571379102766514, + 0.0032475159969180822, + 0.0028069140389561653, + 0.005225729662925005, + 0.005910127889364958, + 0.0044008116237819195, + 0.002904093125835061, + 0.007043040357530117, + 0.006812980864197016, + 0.0017248281510546803, + 0.002062518149614334, + 0.0033958435524255037, + 0.005172946490347385, + 0.002526354044675827, + 0.002292068675160408, + 0.008541571907699108, + 0.006101008504629135, + 0.0011695531429722905, + 0.0012372961500659585, + 0.0015142452903091908, + 0.0028497132007032633, + 0.002300782361999154, + 0.0011249590897932649, + 0.0025624886620789766, + 0.002088982379063964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0017164511373266578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-slow-motion, abstract macrophotographic static close-ups capture a moment of controlled chaos under dramatic, high-contrast studio lighting. A heavy, polished steel spoon, its perfectly reflective bowl and elegant stem, captured in sharp, mirror-like highlights, succumbs to an unseen tremor. It deliberately tips over, its weight palpable as it begins its descent. Beneath it, a pool of viscous, opaque white paint spreads against a dark, perfectly matte background, its surface as smooth and still as glass. The spoon's thin leading edge makes contact first, plunging into the liquid with surgical precision. In extreme slow motion, the paint's surface dimples, resisting for a fraction of a second, then collapses as the spoon carves a clean, vertical path downward. This displacement creates a thick, rolling wave of white paint, cresting and curling outward, its texture similar to that of heavy cream, forming a perfect, silent, crown-like splash. High-contrast lighting shapes the scene, casting the background in deep shadow while illuminating the brilliant white of the paint and the cold metallic sheen of the partially submerged spoon, creating a spectacle that is tense, captivating, and purely physical.", + "seed": 8751, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_2723_seed_8751_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": -0.1433345858938992, + "vlm_logprob": -0.1433345858938992, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.1433345858938992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06561724841594696 + ], + "opposite_logprob_tokens": [ + -2.815617322921753 + ], + "logprob_reward": -0.06561724841594696, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06561724841594696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09637924283742905, + 0.09466857463121414, + 0.07663404941558838, + 0.054519783705472946, + 0.08361175656318665, + 0.2921423614025116, + 0.10814186185598373, + 0.061339423060417175, + 0.04780188947916031, + 0.08476108312606812 + ], + "credit_map": [ + 0.010162143036723137, + 0.008871844038367271, + 0.018416155129671097, + 0.010169648565351963, + 0.00784488394856453, + 0.012555770576000214, + 0.011024128645658493, + 0.01733466237783432, + 0.00788029469549656, + 0.009624028578400612, + 0.014681987464427948, + 0.011491063050925732, + 0.009680367074906826, + 0.014872155152261257, + 0.011616997420787811, + 0.014821666292846203, + 0.00871261302381754, + 0.0047238911502063274, + 0.008445575833320618, + 0.005691647529602051, + 0.011639765463769436, + 0.014319442212581635, + 0.010663076303899288, + 0.012438034638762474, + 0.009656219743192196, + 0.008984007872641087, + 0.0045686443336308, + 0.0030903720762580633, + 0.006377920974045992, + 0.00793592631816864, + 0.0068211629986763, + 0.00708552123978734, + 0.008603201247751713, + 0.010340048000216484, + 0.009131358936429024, + 0.012307283468544483, + 0.006054686848074198, + 0.014665202237665653, + 0.009371276013553143, + 0.013138690032064915, + 0.02752586081624031, + 0.013840578496456146, + 0.008525999262928963, + 0.013669261708855629, + 0.011383146978914738, + 0.11848810315132141, + 0.013214915059506893, + 0.08549445122480392, + 0.014962354674935341, + 0.018857093527913094, + 0.012750038877129555, + 0.010897742584347725, + 0.011509094387292862, + 0.012104887515306473, + 0.01314206700772047, + 0.013918579556047916, + 0.0061930096708238125, + 0.01049281470477581, + 0.0060203177854418755, + 0.006994576659053564, + 0.0065494682639837265, + 0.011259093880653381, + 0.006355130113661289, + 0.0074750082567334175, + 0.005694574676454067, + 0.006848884746432304, + 0.0039623090997338295, + 0.006340446881949902, + 0.004323310684412718, + 0.006974020041525364, + 0.00697364890947938, + 0.006684690248221159, + 0.01163556706160307, + 0.011803611181676388, + 0.007767425384372473, + 0.010934291407465935, + 0.010009881108999252, + 0.015300421044230461, + 0.007873536087572575, + 0.009436339139938354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.06561724841594696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328102946281433 + ], + "opposite_logprob_tokens": [ + -0.7578102946281433 + ], + "logprob_reward": -0.6328102946281433, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328102946281433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14129957556724548, + 0.06420164555311203, + 0.03432907536625862, + 0.033628061413764954, + 0.02856602519750595, + 0.0588078610599041, + 0.06569832563400269, + 0.11794590950012207, + 0.07121067494153976, + 0.3843128979206085 + ], + "credit_map": [ + 0.016993140801787376, + 0.010453073307871819, + 0.022504139691591263, + 0.012112762778997421, + 0.012902284041047096, + 0.0262324009090662, + 0.017858557403087616, + 0.02224322035908699, + 0.005404545459896326, + 0.005746930837631226, + 0.009315837174654007, + 0.008646249771118164, + 0.005922604817897081, + 0.010642198845744133, + 0.008286407217383385, + 0.010236872360110283, + 0.004058127291500568, + 0.0020797145552933216, + 0.003914009779691696, + 0.0031161049846559763, + 0.0038966683205217123, + 0.005799839738756418, + 0.006033329293131828, + 0.005431281868368387, + 0.004573166370391846, + 0.005667477380484343, + 0.0026076864451169968, + 0.0027023961301892996, + 0.0044944616965949535, + 0.005743922665715218, + 0.0034600715152919292, + 0.004378882702440023, + 0.0031487704254686832, + 0.0046555520966649055, + 0.0025901554618030787, + 0.003248426131904125, + 0.002234226558357477, + 0.004168117418885231, + 0.0033420755062252283, + 0.005178702995181084, + 0.006581549532711506, + 0.006541084963828325, + 0.004419267177581787, + 0.003842907724902034, + 0.004458507988601923, + 0.015260150656104088, + 0.004748163744807243, + 0.012956230901181698, + 0.005829473491758108, + 0.005246772896498442, + 0.004933537449687719, + 0.004965488333255053, + 0.005037460010498762, + 0.01935083605349064, + 0.0064054518006742, + 0.013929307460784912, + 0.013992046006023884, + 0.018998825922608376, + 0.011138633824884892, + 0.012490996159613132, + 0.010537594556808472, + 0.022283440455794334, + 0.012414166703820229, + 0.01609020307660103, + 0.006662186235189438, + 0.004907457157969475, + 0.0037397502455860376, + 0.005000000819563866, + 0.008197668939828873, + 0.02431032992899418, + 0.004488639999181032, + 0.01390464324504137, + 0.05111520364880562, + 0.03521129488945007, + 0.02933960221707821, + 0.04551833122968674, + 0.055288851261138916, + 0.05846595764160156, + 0.06236916035413742, + 0.04700449854135513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.6328102946281433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794510900974274 + ], + "opposite_logprob_tokens": [ + -4.63479471206665 + ], + "logprob_reward": -0.009794510900974274, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794510900974274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15369296073913574, + 0.048229534178972244, + 0.04848570376634598, + 0.08417104929685593, + 0.15975336730480194, + 0.07301792502403259, + 0.14648911356925964, + 0.09438598901033401, + 0.07060893625020981, + 0.12116540223360062 + ], + "credit_map": [ + 0.018152253702282906, + 0.008545567281544209, + 0.028010837733745575, + 0.0212392620742321, + 0.011214381083846092, + 0.025766612961888313, + 0.018479058519005775, + 0.022284990176558495, + 0.00469975546002388, + 0.0050810277462005615, + 0.0061906371265649796, + 0.004728452302515507, + 0.0050104702822864056, + 0.006728567648679018, + 0.007402472198009491, + 0.008388147689402103, + 0.007054361514747143, + 0.002887966576963663, + 0.005639508832246065, + 0.0031358362175524235, + 0.005657434929162264, + 0.008624042384326458, + 0.007623719517141581, + 0.007862832397222519, + 0.013940457254648209, + 0.015731194987893105, + 0.008427067659795284, + 0.005533355288207531, + 0.009832475334405899, + 0.010791584849357605, + 0.008632000535726547, + 0.011282912455499172, + 0.01707940176129341, + 0.023361220955848694, + 0.01751566119492054, + 0.0185974333435297, + 0.013744466006755829, + 0.026902969926595688, + 0.017475057393312454, + 0.02507714554667473, + 0.00914505124092102, + 0.007829255424439907, + 0.007218083366751671, + 0.008454249240458012, + 0.005969393998384476, + 0.014297512359917164, + 0.007819460704922676, + 0.012284917756915092, + 0.014925395138561726, + 0.01598271168768406, + 0.013515237718820572, + 0.013870169408619404, + 0.010862115770578384, + 0.032794397324323654, + 0.0221225805580616, + 0.022416498512029648, + 0.007250354625284672, + 0.007692267186939716, + 0.00394896324723959, + 0.005253860726952553, + 0.006679373327642679, + 0.0329819954931736, + 0.007850009948015213, + 0.022729167714715004, + 0.010889393277466297, + 0.00678005488589406, + 0.006569243036210537, + 0.007421120069921017, + 0.006068671587854624, + 0.011481246910989285, + 0.01100118923932314, + 0.010398018173873425, + 0.012962189503014088, + 0.019710713997483253, + 0.014214799739420414, + 0.01823291927576065, + 0.013371082954108715, + 0.01400922704488039, + 0.011582941748201847, + 0.01708153262734413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.009794510900974274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09268201142549515, + 0.049086544662714005, + 0.03289905935525894, + 0.033501118421554565, + 0.031024865806102753, + 0.2836320400238037, + 0.06723276525735855, + 0.08684008568525314, + 0.19991420209407806, + 0.12318722903728485 + ], + "credit_map": [ + 0.008791121654212475, + 0.00685877026990056, + 0.015411884523928165, + 0.013913881033658981, + 0.005241197068244219, + 0.011529066599905491, + 0.00966077670454979, + 0.021275322884321213, + 0.004687805660068989, + 0.004709366708993912, + 0.00639081047847867, + 0.006278628949075937, + 0.0033981865271925926, + 0.009154649451375008, + 0.006838236935436726, + 0.00762886181473732, + 0.004360951948910952, + 0.0024244871456176043, + 0.004086650907993317, + 0.0023472460452467203, + 0.002711922861635685, + 0.004845614545047283, + 0.005753911100327969, + 0.006368275266140699, + 0.00663498742505908, + 0.004688354674726725, + 0.0031585926190018654, + 0.0023270868696272373, + 0.0028040441684424877, + 0.005125999450683594, + 0.0032455353066325188, + 0.00551651930436492, + 0.004371758550405502, + 0.003848152467980981, + 0.0029389183036983013, + 0.0035143960267305374, + 0.00318136066198349, + 0.004876240156590939, + 0.00383017398416996, + 0.0044638714753091335, + 0.030022865161299706, + 0.006202492862939835, + 0.004357692785561085, + 0.009746226482093334, + 0.010305894538760185, + 0.1270550787448883, + 0.0071984585374593735, + 0.08874335139989853, + 0.0065729753114283085, + 0.007454751059412956, + 0.00459207221865654, + 0.003825429594144225, + 0.006087199784815311, + 0.020708566531538963, + 0.004803212825208902, + 0.013188561424612999, + 0.008777836337685585, + 0.013981695286929607, + 0.006142974365502596, + 0.010805356316268444, + 0.007006970699876547, + 0.019271915778517723, + 0.007992090657353401, + 0.012861262075603008, + 0.02332305535674095, + 0.02061075158417225, + 0.012759811244904995, + 0.01741923950612545, + 0.02045627310872078, + 0.05719507113099098, + 0.016216902062296867, + 0.031933121383190155, + 0.014889683574438095, + 0.016378216445446014, + 0.01401924341917038, + 0.0158929992467165, + 0.013721371069550514, + 0.02318650297820568, + 0.012096943333745003, + 0.013002281077206135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.005255098920315504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "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": [ + 0.05840263143181801, + 0.025745349004864693, + 0.0329921692609787, + 0.0234262365847826, + 0.046180956065654755, + 0.5814872980117798, + 0.054444920271635056, + 0.02315584011375904, + 0.0583999864757061, + 0.09576466679573059 + ], + "credit_map": [ + 0.0052835228852927685, + 0.0077192396856844425, + 0.008706892840564251, + 0.006656741723418236, + 0.0060988846234977245, + 0.010285953059792519, + 0.0052346643060445786, + 0.008416733704507351, + 0.002573959296569228, + 0.00279741152189672, + 0.003675558604300022, + 0.0027330610901117325, + 0.002703359117731452, + 0.0023827441036701202, + 0.0034587541595101357, + 0.005420498549938202, + 0.004519264213740826, + 0.0021295915357768536, + 0.004881796892732382, + 0.002455818699672818, + 0.0030835948418825865, + 0.004847092553973198, + 0.006041745189577341, + 0.005033262073993683, + 0.0029696065466850996, + 0.002210820559412241, + 0.002426199847832322, + 0.0016688333125784993, + 0.0022691923659294844, + 0.004715877119451761, + 0.0031719752587378025, + 0.003993728663772345, + 0.004599682986736298, + 0.005378600209951401, + 0.0043207504786551, + 0.005677038338035345, + 0.004147110506892204, + 0.009170063771307468, + 0.004786724224686623, + 0.008100980892777443, + 0.05985230207443237, + 0.029232798144221306, + 0.013837811537086964, + 0.024825412780046463, + 0.02284299023449421, + 0.23855704069137573, + 0.018904950469732285, + 0.17343398928642273, + 0.006961028091609478, + 0.011607863940298557, + 0.004899526014924049, + 0.005856518633663654, + 0.005491578485816717, + 0.0074322461150586605, + 0.005628295708447695, + 0.006567857693880796, + 0.0021707601845264435, + 0.003556558396667242, + 0.0017871004529297352, + 0.003083539428189397, + 0.0025369729846715927, + 0.0031926478259265423, + 0.004168565850704908, + 0.0026596910320222378, + 0.006431389134377241, + 0.00533835357055068, + 0.003008602187037468, + 0.00394545728340745, + 0.006433142349123955, + 0.019449522718787193, + 0.003095329273492098, + 0.010698176920413971, + 0.014389404095709324, + 0.011190459132194519, + 0.007641970179975033, + 0.011213819496333599, + 0.010243738070130348, + 0.016326136887073517, + 0.013100028038024902, + 0.011659100651741028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.003195776604115963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Ultra-slow-motion, abstract macrophotographic static close-ups capture a moment of controlled chaos under dramatic, high-contrast studio lighting. A heavy, polished steel spoon, its perfectly reflective bowl and elegant stem, captured in sharp, mirror-like highlights, succumbs to an unseen tremor. It deliberately tips over, its weight palpable as it begins its descent. Beneath it, a pool of viscous, opaque white paint spreads against a dark, perfectly matte background, its surface as smooth and still as glass. The spoon's thin leading edge makes contact first, plunging into the liquid with surgical precision. In extreme slow motion, the paint's surface dimples, resisting for a fraction of a second, then collapses as the spoon carves a clean, vertical path downward. This displacement creates a thick, rolling wave of white paint, cresting and curling outward, its texture similar to that of heavy cream, forming a perfect, silent, crown-like splash. High-contrast lighting shapes the scene, casting the background in deep shadow while illuminating the brilliant white of the paint and the cold metallic sheen of the partially submerged spoon, creating a spectacle that is tense, captivating, and purely physical.", + "seed": 8752, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_4/index_2723_seed_8752_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": -0.1190107244066894, + "vlm_logprob": -0.1190107244066894, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.1190107244066894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3933994472026825 + ], + "opposite_logprob_tokens": [ + -1.1433994770050049 + ], + "logprob_reward": -0.3933994472026825, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3933994472026825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10626315325498581, + 0.25065872073173523, + 0.08218617737293243, + 0.04817365109920502, + 0.041964951902627945, + 0.04679577797651291, + 0.22036084532737732, + 0.04180041328072548, + 0.08048787713050842, + 0.08130836486816406 + ], + "credit_map": [ + 0.01448084320873022, + 0.0157281756401062, + 0.00739262206479907, + 0.013290906324982643, + 0.020805170759558678, + 0.01673324592411518, + 0.011075076647102833, + 0.006757114548236132, + 0.011784285306930542, + 0.10801202803850174, + 0.0055467658676207066, + 0.003951675724238157, + 0.007159038446843624, + 0.10461241006851196, + 0.004952586721628904, + 0.004639957565814257, + 0.006074465345591307, + 0.022977391257882118, + 0.004804123193025589, + 0.006671397481113672, + 0.009652277454733849, + 0.022788364440202713, + 0.004791983403265476, + 0.004426170140504837, + 0.0065713440999388695, + 0.00624871114268899, + 0.007119838148355484, + 0.004279148764908314, + 0.008790313266217709, + 0.007049329113215208, + 0.0037886181380599737, + 0.004326347261667252, + 0.005692898295819759, + 0.0060754017904400826, + 0.006200940813869238, + 0.00359206460416317, + 0.007556280121207237, + 0.005461181979626417, + 0.003189916955307126, + 0.004196269903331995, + 0.005551071371883154, + 0.005986621603369713, + 0.004706164821982384, + 0.007035799324512482, + 0.009037029929459095, + 0.006307784002274275, + 0.005254270974546671, + 0.002917033853009343, + 0.03209622576832771, + 0.03894141688942909, + 0.025496099144220352, + 0.03016277588903904, + 0.030881768092513084, + 0.031511131674051285, + 0.01828761026263237, + 0.012983820401132107, + 0.005448043346405029, + 0.007100624497979879, + 0.004246388096362352, + 0.005541710183024406, + 0.0071213082410395145, + 0.006587259005755186, + 0.002946814289316535, + 0.002808262826874852, + 0.010524841025471687, + 0.010211392305791378, + 0.011746867559850216, + 0.008725544437766075, + 0.00970859918743372, + 0.010052965022623539, + 0.009427434764802456, + 0.010090231895446777, + 0.009229018352925777, + 0.010686101391911507, + 0.007967976853251457, + 0.008489063940942287, + 0.014065916649997234, + 0.009448900818824768, + 0.014464037492871284, + 0.0069573489017784595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.3933994472026825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992790043354034 + ], + "opposite_logprob_tokens": [ + -1.804927945137024 + ], + "logprob_reward": -0.17992790043354034, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992790043354034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07571298629045486, + 0.19852076470851898, + 0.12720966339111328, + 0.2716507315635681, + 0.028229927644133568, + 0.04193314164876938, + 0.10734318941831589, + 0.07432852685451508, + 0.05893240123987198, + 0.016138680279254913 + ], + "credit_map": [ + 0.00833758246153593, + 0.011585173197090626, + 0.009756267070770264, + 0.011058117263019085, + 0.01007012091577053, + 0.011043094098567963, + 0.007448262069374323, + 0.006414368283003569, + 0.011887154541909695, + 0.08270051330327988, + 0.005537624470889568, + 0.0049091218970716, + 0.005677565932273865, + 0.07578947395086288, + 0.0062597510404884815, + 0.005759565159678459, + 0.007070690393447876, + 0.0440431572496891, + 0.006030989810824394, + 0.008503074757754803, + 0.005549427587538958, + 0.04295290261507034, + 0.008156312629580498, + 0.004903104621917009, + 0.03982870653271675, + 0.035155534744262695, + 0.03203018754720688, + 0.02582065761089325, + 0.06101534515619278, + 0.0395253449678421, + 0.01880529709160328, + 0.019469670951366425, + 0.004024066962301731, + 0.003817723598331213, + 0.004004886373877525, + 0.0025137874763458967, + 0.006162922363728285, + 0.003764413995668292, + 0.001907064695842564, + 0.0020350636914372444, + 0.005453919991850853, + 0.0068527101539075375, + 0.005786481313407421, + 0.0051778098568320274, + 0.006643735338002443, + 0.005291211884468794, + 0.004036868456751108, + 0.0026904067490249872, + 0.013950088061392307, + 0.018756745383143425, + 0.011237677186727524, + 0.011900930665433407, + 0.01628747768700123, + 0.013992984779179096, + 0.012235824950039387, + 0.008981464430689812, + 0.009046574123203754, + 0.016391200944781303, + 0.007894434966146946, + 0.007971063256263733, + 0.01271101739257574, + 0.01176965981721878, + 0.00425313226878643, + 0.004291442688554525, + 0.0077482229098677635, + 0.010406248271465302, + 0.0062211365438997746, + 0.006085050292313099, + 0.009618847630918026, + 0.008405808359384537, + 0.006302439607679844, + 0.004144651349633932, + 0.002058005193248391, + 0.002107171108946204, + 0.0019776036497205496, + 0.001947662909515202, + 0.002585106296464801, + 0.002125245053321123, + 0.002182409167289734, + 0.0011554770171642303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.17992790043354034 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028276958037167788 + ], + "opposite_logprob_tokens": [ + -5.8778276443481445 + ], + "logprob_reward": -0.0028276958037167788, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028276958037167788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025070393458008766, + 0.8078376054763794, + 0.013950511813163757, + 0.01497724000364542, + 0.00693551916629076, + 0.025528935715556145, + 0.041417572647333145, + 0.014729195274412632, + 0.028517521917819977, + 0.021035537123680115 + ], + "credit_map": [ + 0.003088788827881217, + 0.003199400147423148, + 0.0018958814907819033, + 0.003535452764481306, + 0.005610807798802853, + 0.0036701136268675327, + 0.002514013322070241, + 0.0015559359453618526, + 0.03327503800392151, + 0.3742258548736572, + 0.00520024448633194, + 0.008424169383943081, + 0.015530061908066273, + 0.35693421959877014, + 0.005924221593886614, + 0.008323750458657742, + 0.0009925919584929943, + 0.0042209201492369175, + 0.0008344954694621265, + 0.0006737075163982809, + 0.0014869638253003359, + 0.00452389195561409, + 0.0007100655930116773, + 0.000507875403854996, + 0.0032046171836555004, + 0.0024632506538182497, + 0.0012272701133042574, + 0.0010511475848034024, + 0.001896718400530517, + 0.0025740996934473515, + 0.001043035532347858, + 0.0015171010745689273, + 0.0010585574200376868, + 0.0009475070401094854, + 0.000828560150694102, + 0.0005517835961654782, + 0.0017759528709575534, + 0.0008767141262069345, + 0.00048327696276828647, + 0.00041316726128570735, + 0.0034354717936366796, + 0.003139193169772625, + 0.0029655792750418186, + 0.0029595240484923124, + 0.004857259802520275, + 0.003723881207406521, + 0.0021938784047961235, + 0.002254149643704295, + 0.005653422325849533, + 0.007262796629220247, + 0.003330741310492158, + 0.005462354514747858, + 0.008850381709635258, + 0.006034914869815111, + 0.0027531934902071953, + 0.0020697657018899918, + 0.001535765826702118, + 0.002941111335530877, + 0.0013084921520203352, + 0.0023849329445511103, + 0.002125715371221304, + 0.0021770282182842493, + 0.0013425274519249797, + 0.0009136213338933885, + 0.003080492140725255, + 0.004095697309821844, + 0.004140119068324566, + 0.003147486597299576, + 0.004121692385524511, + 0.0036495451349765062, + 0.003911637235432863, + 0.0023708532098680735, + 0.002035196404904127, + 0.0034872661344707012, + 0.0023521112743765116, + 0.0020332233980298042, + 0.0037279832176864147, + 0.0028258718084543943, + 0.0028999762143939734, + 0.0016739090206101537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.0028276958037167788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075875610113144 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016075875610113144, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075875610113144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02115282602608204, + 0.7373149991035461, + 0.012190011329948902, + 0.05044923350214958, + 0.009169738739728928, + 0.02834247052669525, + 0.061660122126340866, + 0.02198415994644165, + 0.02352961339056492, + 0.034206852316856384 + ], + "credit_map": [ + 0.0025809830985963345, + 0.0023495585191994905, + 0.003429233795031905, + 0.003513009287416935, + 0.0036392284091562033, + 0.002554489066824317, + 0.0018413504585623741, + 0.0012449714122340083, + 0.031862303614616394, + 0.3411012291908264, + 0.004915271885693073, + 0.007934579625725746, + 0.014558013528585434, + 0.32390856742858887, + 0.0053281355649232864, + 0.007706813979893923, + 0.0013796198181807995, + 0.002489763079211116, + 0.0012698761420324445, + 0.0013262087013572454, + 0.0013605643762275577, + 0.002348938724026084, + 0.001046532765030861, + 0.0009685062104836106, + 0.008644325658679008, + 0.0075922696851193905, + 0.006046625319868326, + 0.003971555270254612, + 0.0060813226737082005, + 0.008425828069448471, + 0.0038088506553322077, + 0.00587845453992486, + 0.0014418089995160699, + 0.0012552404077723622, + 0.0014011638704687357, + 0.0010763827012851834, + 0.001661494723521173, + 0.0011671200627461076, + 0.0005522471037693322, + 0.0006142798811197281, + 0.003910484723746777, + 0.0038845455273985863, + 0.004059764556586742, + 0.0024477296974509954, + 0.0054710800759494305, + 0.0038040177896618843, + 0.0028381168376654387, + 0.0019267291063442826, + 0.008282029069960117, + 0.01064303144812584, + 0.005576477851718664, + 0.007541900500655174, + 0.008702755905687809, + 0.010001440532505512, + 0.00539503013715148, + 0.005517454817891121, + 0.0027317518834024668, + 0.004404851701110601, + 0.002388507593423128, + 0.0028113811276853085, + 0.0028740563429892063, + 0.0034701316617429256, + 0.0018107319483533502, + 0.001492746057920158, + 0.0026532933115959167, + 0.0034515224397182465, + 0.0027998127043247223, + 0.0026443975511938334, + 0.0036067438777536154, + 0.003294172929599881, + 0.0030214907601475716, + 0.002058176789432764, + 0.004018440376967192, + 0.005713144317269325, + 0.0034281769767403603, + 0.0021152300760149956, + 0.007354754954576492, + 0.004726886749267578, + 0.004011858254671097, + 0.002838356886059046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.016075875610113144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy, polished steel spoon with a reflective bowl and stem?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, perfectly matte surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a pool of viscous, opaque white paint spread beneath the spoon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the spoon tip over and plunge vertically into the paint, creating a crown-like splash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in ultra-slow motion with high-contrast studio lighting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822702983394265 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.002822702983394265, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822702983394265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03913388401269913, + 0.6212494969367981, + 0.05712183564901352, + 0.13710364699363708, + 0.024450408294796944, + 0.011936993338167667, + 0.046698469668626785, + 0.015036383643746376, + 0.020258184522390366, + 0.02701074257493019 + ], + "credit_map": [ + 0.004538354463875294, + 0.006082730367779732, + 0.004041107837110758, + 0.0036905459128320217, + 0.007544814608991146, + 0.006136707030236721, + 0.0038959451485425234, + 0.003203675150871277, + 0.026676585897803307, + 0.28676730394363403, + 0.004631675314158201, + 0.007348669692873955, + 0.012000766582787037, + 0.273098886013031, + 0.004532284568995237, + 0.006193273235112429, + 0.003330471459776163, + 0.021061284467577934, + 0.002266047289595008, + 0.0026760834734886885, + 0.0023072336334735155, + 0.020611263811588287, + 0.002767121186479926, + 0.0021023242734372616, + 0.022629741579294205, + 0.01755925826728344, + 0.017633525654673576, + 0.012464865110814571, + 0.028813816606998444, + 0.02001522108912468, + 0.008532109670341015, + 0.009455092251300812, + 0.00351197668351233, + 0.0025076204910874367, + 0.003735970240086317, + 0.002436336362734437, + 0.003989889286458492, + 0.0033451279159635305, + 0.0022199477534741163, + 0.0027035344392061234, + 0.0017998713301494718, + 0.0015914940740913153, + 0.0011663479963317513, + 0.001615859684534371, + 0.002119554439559579, + 0.0016949918353930116, + 0.0010233100038021803, + 0.0009255625773221254, + 0.0063850185833871365, + 0.006322030909359455, + 0.004429126624017954, + 0.008011898025870323, + 0.006670269183814526, + 0.006572495214641094, + 0.004423976875841618, + 0.003883647732436657, + 0.0017172576626762748, + 0.0024001740384846926, + 0.0015200391644611955, + 0.00286194309592247, + 0.0020168041810393333, + 0.0022665360011160374, + 0.0013075487222522497, + 0.0009460785076953471, + 0.0025371715892106295, + 0.0024880669079720974, + 0.0030958482529968023, + 0.002270778641104698, + 0.0023726311046630144, + 0.002455710666254163, + 0.0025815966073423624, + 0.0024563781917095184, + 0.0027180074248462915, + 0.0038179957773536444, + 0.0028714193031191826, + 0.0025880062021315098, + 0.00437420979142189, + 0.003225694177672267, + 0.0047194501385092735, + 0.0026959541719406843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.002822702983394265 + } + ], + "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/0000055/group_5/index_tensor([9189, 9189])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_tensor([9189, 9189])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..456b6fa8cf82b2100aea36da3e2d2da715ee8441 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_tensor([9189, 9189])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, upward-sweeping long shot, shot from the base of the mountain, steadily and slowly pulls back to reveal the full extent of the catastrophe. In the center of the frame, a majestic, precipitous volcano stands against a dark, crimson sky stained by volcanic ash. Its jagged crater rim becomes the source of destruction. A stream of incandescent, viscous, orange-red molten lava erupts with explosive force from the summit, forming a massive fountain that soars skyward, hurling countless flaming fragments of rock high into the sky. Meanwhile, a thick plume of ash and gas, like an awakening beast, twists and turns upward, its interior illuminated by the glow of the molten lava, punctuated by eerie flashes of red lightning. The scene is illuminated solely by this infernal eruption. The fiery orange-yellow contrasts sharply with the dark red of the sky and the inky black of the mountain, creating a suffocating, apocalyptic atmosphere imbued with primal destructive power.", + "seed": 11098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_9189_seed_11098_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": -0.16735328151844442, + "vlm_logprob": -0.16735328151844442, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.16735328151844442, + "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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7014861702919006 + ], + "opposite_logprob_tokens": [ + -0.7014861702919006 + ], + "logprob_reward": -0.7014861702919006, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7014861702919006, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05264010652899742, + 0.0495784766972065, + 0.07887890189886093, + 0.33378511667251587, + 0.029052626341581345, + 0.036617960780858994, + 0.08233383297920227, + 0.05687190964818001, + 0.14112679660320282, + 0.13911432027816772 + ], + "credit_map": [ + 0.004327204078435898, + 0.004405011422932148, + 0.010651892982423306, + 0.006328882183879614, + 0.006717880256474018, + 0.006703932303935289, + 0.007246604189276695, + 0.006258699577301741, + 0.004090209491550922, + 0.004527228884398937, + 0.011059774085879326, + 0.007215549238026142, + 0.006391453091055155, + 0.004839528817683458, + 0.0054425885900855064, + 0.006012142635881901, + 0.007191479671746492, + 0.018006252124905586, + 0.009523630142211914, + 0.006821934599429369, + 0.010674627497792244, + 0.017567284405231476, + 0.00476912222802639, + 0.0043245768174529076, + 0.023300284519791603, + 0.11401411145925522, + 0.013521106913685799, + 0.01056827511638403, + 0.059817999601364136, + 0.09382186830043793, + 0.007192492950707674, + 0.011548982001841068, + 0.0029510236345231533, + 0.00295179127715528, + 0.005857011303305626, + 0.003711051307618618, + 0.002785403747111559, + 0.00515030650421977, + 0.002323680091649294, + 0.0033223575446754694, + 0.0038932664319872856, + 0.001871710643172264, + 0.006848126649856567, + 0.007560084108263254, + 0.003050279337912798, + 0.005782749503850937, + 0.002930577378720045, + 0.004681165795773268, + 0.0045502278953790665, + 0.004492029547691345, + 0.0073125362396240234, + 0.012222044169902802, + 0.02029133401811123, + 0.019456367939710617, + 0.00576661666855216, + 0.00824267603456974, + 0.00518179964274168, + 0.004282230511307716, + 0.006774932146072388, + 0.009562811814248562, + 0.006179697811603546, + 0.007923584431409836, + 0.007371865678578615, + 0.009594983421266079, + 0.012427815236151218, + 0.016262488439679146, + 0.01612919382750988, + 0.03269428387284279, + 0.015566457062959671, + 0.015679799020290375, + 0.012022062204778194, + 0.020344696938991547, + 0.011018224991858006, + 0.012804677709937096, + 0.033209167420864105, + 0.023461908102035522, + 0.010345470160245895, + 0.018570007756352425, + 0.010631374083459377, + 0.019073477014899254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.7014861702919006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900354266166687 + ], + "opposite_logprob_tokens": [ + -2.5790035724639893 + ], + "logprob_reward": -0.07900354266166687, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900354266166687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04535882547497749, + 0.04129251465201378, + 0.05333122983574867, + 0.4060671925544739, + 0.02072242833673954, + 0.02702353522181511, + 0.023037239909172058, + 0.0369926393032074, + 0.14162404835224152, + 0.20455040037631989 + ], + "credit_map": [ + 0.004481357987970114, + 0.003141594585031271, + 0.01123867742717266, + 0.003823458682745695, + 0.003994733560830355, + 0.00865512527525425, + 0.004320806358009577, + 0.00570307020097971, + 0.007566777989268303, + 0.0043381256982684135, + 0.003769746981561184, + 0.002108838642016053, + 0.006460802163928747, + 0.012513742782175541, + 0.00230457354336977, + 0.0022299042902886868, + 0.005354791879653931, + 0.012313208542764187, + 0.006338371429592371, + 0.0035162202548235655, + 0.00711399968713522, + 0.012913858518004417, + 0.0025403255131095648, + 0.003240457270294428, + 0.02831188403069973, + 0.13705700635910034, + 0.017408041283488274, + 0.012788074091076851, + 0.07237234711647034, + 0.11394866555929184, + 0.009667651727795601, + 0.014513520523905754, + 0.002009750111028552, + 0.0028876366559416056, + 0.0025636625941842794, + 0.0028289775364100933, + 0.0022113320883363485, + 0.0031919421162456274, + 0.0017932706978172064, + 0.0032358565367758274, + 0.0027964559849351645, + 0.001592773711308837, + 0.005357823334634304, + 0.004963830579072237, + 0.002513024490326643, + 0.004084390122443438, + 0.0020254533737897873, + 0.0036897831596434116, + 0.0017329026013612747, + 0.0016450592083856463, + 0.0033296255860477686, + 0.0036108088679611683, + 0.004440605640411377, + 0.0050520747900009155, + 0.0015123803168535233, + 0.001713783130981028, + 0.0035157143138349056, + 0.0026398678310215473, + 0.00458128796890378, + 0.005484721623361111, + 0.004060991108417511, + 0.005259874276816845, + 0.0046258955262601376, + 0.0068242838606238365, + 0.013150453567504883, + 0.015862325206398964, + 0.016421878710389137, + 0.03350307047367096, + 0.014780485071241856, + 0.015442097559571266, + 0.011622354388237, + 0.020841378718614578, + 0.015339992940425873, + 0.01879320852458477, + 0.04944709315896034, + 0.037012502551078796, + 0.015808094292879105, + 0.02622479572892189, + 0.015059583820402622, + 0.026865126565098763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.07900354266166687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289471864700317 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289471864700317, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289471864700317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029684411361813545, + 0.08196622878313065, + 0.05346639081835747, + 0.5999870300292969, + 0.030269762501120567, + 0.06720420718193054, + 0.017143715173006058, + 0.02750188671052456, + 0.06429541110992432, + 0.028480958193540573 + ], + "credit_map": [ + 0.0023290382232517004, + 0.0037815223913639784, + 0.005536493845283985, + 0.00269987341016531, + 0.0041792187839746475, + 0.003893222426995635, + 0.003079079557210207, + 0.0041859629563987255, + 0.016001055017113686, + 0.004402463790029287, + 0.011669566854834557, + 0.0064727747812867165, + 0.008435633033514023, + 0.025252530351281166, + 0.003754391334950924, + 0.005977816414088011, + 0.005112101323902607, + 0.015970872715115547, + 0.0033497256226837635, + 0.002322874963283539, + 0.009333794005215168, + 0.013138345442712307, + 0.0021947133354842663, + 0.0020439624786376953, + 0.042193908244371414, + 0.2030322104692459, + 0.024955257773399353, + 0.018483297899365425, + 0.10556398332118988, + 0.17058493196964264, + 0.013799748383462429, + 0.0213736891746521, + 0.0032061117235571146, + 0.005376036278903484, + 0.004733460955321789, + 0.0027754155453294516, + 0.004280934575945139, + 0.0035386611707508564, + 0.002550992416217923, + 0.003808146808296442, + 0.005880606826394796, + 0.0041141207329928875, + 0.015099705196917057, + 0.016063151881098747, + 0.00635982071980834, + 0.00797702930867672, + 0.006896825507283211, + 0.00481294933706522, + 0.0014523238642141223, + 0.0013894180301576853, + 0.0022273960057646036, + 0.00448582461103797, + 0.002633280586451292, + 0.0022121614310890436, + 0.0014121835120022297, + 0.0013311272487044334, + 0.0022768666967749596, + 0.00292256404645741, + 0.0032975925132632256, + 0.0040076919831335545, + 0.004168477840721607, + 0.003103929804638028, + 0.00321503565646708, + 0.004509729333221912, + 0.006567481905221939, + 0.0058496566489338875, + 0.009039881639182568, + 0.01759890839457512, + 0.005288730375468731, + 0.007705855183303356, + 0.005464444402605295, + 0.006780452094972134, + 0.0023583145812153816, + 0.0025333391968160868, + 0.007057531736791134, + 0.004792542662471533, + 0.0022662857081741095, + 0.0034559881314635277, + 0.002698614727705717, + 0.00331834121607244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.023289471864700317 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797744005918503 + ], + "opposite_logprob_tokens": [ + -3.5297977924346924 + ], + "logprob_reward": -0.029797744005918503, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797744005918503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1132405698299408, + 0.09911470115184784, + 0.37698742747306824, + 0.06598452478647232, + 0.07361803203821182, + 0.04384841024875641, + 0.06350908428430557, + 0.07076049596071243, + 0.04862358048558235, + 0.04431314766407013 + ], + "credit_map": [ + 0.010164816863834858, + 0.011673646047711372, + 0.022196859121322632, + 0.011721687391400337, + 0.013819226995110512, + 0.01370198093354702, + 0.012641035951673985, + 0.01732131838798523, + 0.01597697287797928, + 0.007741569075733423, + 0.015732908621430397, + 0.008858841843903065, + 0.011810756288468838, + 0.023968635126948357, + 0.006512293126434088, + 0.008512728847563267, + 0.033025722950696945, + 0.0829329788684845, + 0.03143437206745148, + 0.025465577840805054, + 0.061600491404533386, + 0.07763899117708206, + 0.040940988808870316, + 0.023948291316628456, + 0.004570672754198313, + 0.01706252060830593, + 0.008657281287014484, + 0.004838571418076754, + 0.01047308649867773, + 0.011965110898017883, + 0.0037060114555060863, + 0.004711271263659, + 0.005464482121169567, + 0.01417337916791439, + 0.00959685817360878, + 0.00908464752137661, + 0.013897479511797428, + 0.00912149902433157, + 0.00534656411036849, + 0.006933121010661125, + 0.0038921113591641188, + 0.0031125580426305532, + 0.007384932599961758, + 0.007219369988888502, + 0.005640664603561163, + 0.006572242360562086, + 0.004169560503214598, + 0.005856968928128481, + 0.004662816412746906, + 0.003607938066124916, + 0.00863216444849968, + 0.016258368268609047, + 0.008381091058254242, + 0.010378861799836159, + 0.005651172250509262, + 0.005936670117080212, + 0.006512223742902279, + 0.005978711415082216, + 0.009942206554114819, + 0.01093856617808342, + 0.010083614848554134, + 0.009498056955635548, + 0.007888447493314743, + 0.009918664582073689, + 0.0035838286858052015, + 0.006932470947504044, + 0.005557060241699219, + 0.00851521734148264, + 0.007573770359158516, + 0.006059903651475906, + 0.005196486599743366, + 0.005204842425882816, + 0.0034392119850963354, + 0.003659393172711134, + 0.009921932592988014, + 0.007613022346049547, + 0.003878512652590871, + 0.0056455847807228565, + 0.003954252228140831, + 0.006201239302754402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.029797744005918503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04669693112373352, + 0.04563503712415695, + 0.03557367995381355, + 0.6588981747627258, + 0.03958507254719734, + 0.02480774186551571, + 0.052050866186618805, + 0.038605015724897385, + 0.030144698917865753, + 0.028002845123410225 + ], + "credit_map": [ + 0.004646584391593933, + 0.005226728040724993, + 0.00919802300632, + 0.004168510437011719, + 0.006382770836353302, + 0.007164471782743931, + 0.005934447515755892, + 0.003975395113229752, + 0.006830069702118635, + 0.004571574740111828, + 0.0071143764071166515, + 0.0036465793382376432, + 0.005969952791929245, + 0.009792294353246689, + 0.0036400577519088984, + 0.004070132039487362, + 0.0029527503065764904, + 0.009749600663781166, + 0.003380225272849202, + 0.002258084714412689, + 0.006317139603197575, + 0.0062282029539346695, + 0.002622937550768256, + 0.0020647388882935047, + 0.045763712376356125, + 0.22331839799880981, + 0.02794598415493965, + 0.02069207839667797, + 0.11701689660549164, + 0.18515489995479584, + 0.015530052594840527, + 0.023476121947169304, + 0.004536292050033808, + 0.006980074569582939, + 0.006514399312436581, + 0.0042068567126989365, + 0.00558863952755928, + 0.005186965689063072, + 0.003999633248895407, + 0.0025722100399434566, + 0.002764556324109435, + 0.001739084254950285, + 0.004443625453859568, + 0.005462245550006628, + 0.0027228770777583122, + 0.003182965097948909, + 0.0023477075155824423, + 0.0021446808241307735, + 0.0030778818763792515, + 0.0033484555315226316, + 0.00482757156714797, + 0.011503822170197964, + 0.012032835744321346, + 0.010968090035021305, + 0.003383982228115201, + 0.0029082249384373426, + 0.0032373853027820587, + 0.002773596439510584, + 0.005576164927333593, + 0.00712929293513298, + 0.0040418840944767, + 0.004916600417345762, + 0.004701881669461727, + 0.006228210870176554, + 0.0020770402625203133, + 0.003263498656451702, + 0.0034013050608336926, + 0.005181577056646347, + 0.004952833987772465, + 0.004046429414302111, + 0.004230435471981764, + 0.0029915787745267153, + 0.001956809312105179, + 0.003106242511421442, + 0.004421234130859375, + 0.0067575834691524506, + 0.002547272015362978, + 0.002988527761772275, + 0.0038049882277846336, + 0.0024201872292906046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.0031894787680357695 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, upward-sweeping long shot, shot from the base of the mountain, steadily and slowly pulls back to reveal the full extent of the catastrophe. In the center of the frame, a majestic, precipitous volcano stands against a dark, crimson sky stained by volcanic ash. Its jagged crater rim becomes the source of destruction. A stream of incandescent, viscous, orange-red molten lava erupts with explosive force from the summit, forming a massive fountain that soars skyward, hurling countless flaming fragments of rock high into the sky. Meanwhile, a thick plume of ash and gas, like an awakening beast, twists and turns upward, its interior illuminated by the glow of the molten lava, punctuated by eerie flashes of red lightning. The scene is illuminated solely by this infernal eruption. The fiery orange-yellow contrasts sharply with the dark red of the sky and the inky black of the mountain, creating a suffocating, apocalyptic atmosphere imbued with primal destructive power.", + "seed": 11099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_9189_seed_11099_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": -0.22474477863870562, + "vlm_logprob": -0.22474477863870562, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.22474477863870562, + "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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.767254650592804 + ], + "opposite_logprob_tokens": [ + -0.642254650592804 + ], + "logprob_reward": -0.767254650592804, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.767254650592804, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.177800714969635, + 0.17476828396320343, + 0.07375939190387726, + 0.0814317911863327, + 0.05572250857949257, + 0.18205957114696503, + 0.039854515343904495, + 0.06422427296638489, + 0.09706246852874756, + 0.05331640690565109 + ], + "credit_map": [ + 0.01806073822081089, + 0.021701138466596603, + 0.03581427037715912, + 0.02578536421060562, + 0.019087981432676315, + 0.020307419821619987, + 0.016467437148094177, + 0.020576393231749535, + 0.02724926732480526, + 0.018599417060613632, + 0.01701650768518448, + 0.012058228254318237, + 0.01825762540102005, + 0.0454699732363224, + 0.01118857879191637, + 0.024928687140345573, + 0.008710572496056557, + 0.012674449011683464, + 0.008424423635005951, + 0.009655952453613281, + 0.008052201010286808, + 0.009308157488703728, + 0.008082929067313671, + 0.008850709535181522, + 0.012436884455382824, + 0.007750080432742834, + 0.012464283965528011, + 0.014794490300118923, + 0.009439061395823956, + 0.010606512427330017, + 0.006708345375955105, + 0.0072321416810154915, + 0.008209295570850372, + 0.008042576722800732, + 0.006167631130665541, + 0.008329678326845169, + 0.005230967421084642, + 0.007622450590133667, + 0.004878605250269175, + 0.007241309154778719, + 0.018085725605487823, + 0.012732348404824734, + 0.027749078348279, + 0.021049635484814644, + 0.018589084967970848, + 0.025463629513978958, + 0.023474790155887604, + 0.034915290772914886, + 0.003382911207154393, + 0.0043897745199501514, + 0.006452188361436129, + 0.0030403989367187023, + 0.004344592802226543, + 0.006499819457530975, + 0.0055121490731835365, + 0.006232685875147581, + 0.006531034596264362, + 0.0069308350794017315, + 0.006619683001190424, + 0.006775786634534597, + 0.013362319208681583, + 0.009000807069242, + 0.009577139280736446, + 0.005426669493317604, + 0.010712925344705582, + 0.01019725389778614, + 0.010021019726991653, + 0.011416786350309849, + 0.015495328232645988, + 0.013282419182360172, + 0.013398993760347366, + 0.012537752278149128, + 0.008100664243102074, + 0.005834771785885096, + 0.00474317604675889, + 0.004119776654988527, + 0.009115410968661308, + 0.010297073051333427, + 0.005562819540500641, + 0.005542720668017864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.767254650592804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347861886024475 + ], + "opposite_logprob_tokens": [ + -1.3134785890579224 + ], + "logprob_reward": -0.31347861886024475, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31347861886024475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18651820719242096, + 0.11058488488197327, + 0.0892522856593132, + 0.08321875333786011, + 0.12059290707111359, + 0.10350476950407028, + 0.08972601592540741, + 0.0586291179060936, + 0.06836796551942825, + 0.08960519731044769 + ], + "credit_map": [ + 0.019994035363197327, + 0.017877792939543724, + 0.032042618840932846, + 0.01830022782087326, + 0.018277237191796303, + 0.04605324566364288, + 0.010397341102361679, + 0.02357567474246025, + 0.013817623257637024, + 0.011510241776704788, + 0.011785480193793774, + 0.01318383775651455, + 0.01757737062871456, + 0.01966983452439308, + 0.008268260397017002, + 0.014772220514714718, + 0.010115678422152996, + 0.018979359418153763, + 0.011234315112233162, + 0.01559302769601345, + 0.01076178066432476, + 0.009030950255692005, + 0.006293493323028088, + 0.00724366120994091, + 0.010939582251012325, + 0.008843193762004375, + 0.01171267218887806, + 0.01438808161765337, + 0.0077346633188426495, + 0.009549160487949848, + 0.007311737164855003, + 0.012739649042487144, + 0.014123335480690002, + 0.013479569926857948, + 0.014998509548604488, + 0.018543502315878868, + 0.017456788569688797, + 0.01660371944308281, + 0.011025719344615936, + 0.014361750334501266, + 0.011527067050337791, + 0.014857714995741844, + 0.015603773295879364, + 0.0113068837672472, + 0.012547224760055542, + 0.014033859595656395, + 0.011159791611135006, + 0.012468427419662476, + 0.008917749859392643, + 0.011199120432138443, + 0.016216319054365158, + 0.0064070685766637325, + 0.009905395098030567, + 0.013020932674407959, + 0.012890542857348919, + 0.011168869212269783, + 0.005962950177490711, + 0.006568579934537411, + 0.007178258150815964, + 0.005251735914498568, + 0.010421901941299438, + 0.0075516728684306145, + 0.0075209070928394794, + 0.008173106238245964, + 0.0070911250077188015, + 0.0066503193229436874, + 0.009373124688863754, + 0.0071664233691990376, + 0.009708636440336704, + 0.012072532437741756, + 0.009128948673605919, + 0.007176841609179974, + 0.007251919712871313, + 0.005536716431379318, + 0.008481624536216259, + 0.009048674255609512, + 0.017004378139972687, + 0.020700203254818916, + 0.00890093669295311, + 0.012680727988481522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.31347861886024475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300303146243095 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300303146243095, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300303146243095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22163140773773193, + 0.20245619118213654, + 0.16344600915908813, + 0.061233628541231155, + 0.0600014328956604, + 0.1552855372428894, + 0.049509625881910324, + 0.02669910341501236, + 0.030712036415934563, + 0.029024973511695862 + ], + "credit_map": [ + 0.02871202677488327, + 0.019193554297089577, + 0.024943895637989044, + 0.027603957802057266, + 0.02488698437809944, + 0.044528719037771225, + 0.017347313463687897, + 0.034414976835250854, + 0.04330141097307205, + 0.011327735148370266, + 0.018099207431077957, + 0.00894379336386919, + 0.01845737174153328, + 0.07429425418376923, + 0.008056513033807278, + 0.019975904375314713, + 0.018158871680498123, + 0.021655967459082603, + 0.031470008194446564, + 0.017694197595119476, + 0.013566658832132816, + 0.03132736682891846, + 0.013709875755012035, + 0.01586306095123291, + 0.008770082145929337, + 0.007186851464211941, + 0.008067552000284195, + 0.008505699224770069, + 0.0066339303739368916, + 0.008901054970920086, + 0.005722828675061464, + 0.007445633411407471, + 0.007548808585852385, + 0.007943086326122284, + 0.006770631298422813, + 0.010557916015386581, + 0.005561024881899357, + 0.007684797048568726, + 0.006620200350880623, + 0.007314973510801792, + 0.0143013596534729, + 0.019301939755678177, + 0.02001139521598816, + 0.02136683091521263, + 0.02034907601773739, + 0.02039441280066967, + 0.01815173216164112, + 0.021408796310424805, + 0.004852069541811943, + 0.006325558293610811, + 0.008407073095440865, + 0.0032001605723053217, + 0.005860473960638046, + 0.007302519865334034, + 0.006712683476507664, + 0.006849091500043869, + 0.0028883388731628656, + 0.002758361166343093, + 0.002215989399701357, + 0.0022996498737484217, + 0.006648065987974405, + 0.0037188862916082144, + 0.003621343756094575, + 0.0025484685320407152, + 0.003022189252078533, + 0.00309656816534698, + 0.003265133360400796, + 0.003102285088971257, + 0.005290511529892683, + 0.0044772145338356495, + 0.004543495364487171, + 0.00391464214771986, + 0.0029404498636722565, + 0.004206174518913031, + 0.002808287274092436, + 0.0027967486530542374, + 0.006048532668501139, + 0.004571602679789066, + 0.0026705628260970116, + 0.0029826152604073286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.023300303146243095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607564091682434 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.01607564091682434, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607564091682434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14717763662338257, + 0.32177555561065674, + 0.08569986373186111, + 0.1421523541212082, + 0.11263639479875565, + 0.06338892132043839, + 0.022857844829559326, + 0.030477374792099, + 0.03314794600009918, + 0.04068610072135925 + ], + "credit_map": [ + 0.01571042835712433, + 0.02145477384328842, + 0.026378922164440155, + 0.019009215757250786, + 0.02069241739809513, + 0.021225016564130783, + 0.00904857087880373, + 0.01365827675908804, + 0.055633533746004105, + 0.027951927855610847, + 0.030666403472423553, + 0.02633601427078247, + 0.041421569883823395, + 0.08912351727485657, + 0.02030336670577526, + 0.030339226126670837, + 0.009469948709011078, + 0.015753362327814102, + 0.01140082161873579, + 0.008494564332067966, + 0.010001873597502708, + 0.010833910666406155, + 0.00834805890917778, + 0.011397317051887512, + 0.01791108213365078, + 0.018100274726748466, + 0.019831109791994095, + 0.03242629021406174, + 0.014672980643808842, + 0.019875645637512207, + 0.01047641970217228, + 0.008858544752001762, + 0.011494063772261143, + 0.011848288588225842, + 0.010935472324490547, + 0.021628813818097115, + 0.017203165218234062, + 0.014138895086944103, + 0.011370575986802578, + 0.014017111621797085, + 0.006195705849677324, + 0.005817445460706949, + 0.012818576768040657, + 0.007844711653888226, + 0.008082208223640919, + 0.009230328723788261, + 0.006815081462264061, + 0.006584855727851391, + 0.0025321990251541138, + 0.0029902574606239796, + 0.003450945019721985, + 0.0018363363342359662, + 0.0028261893894523382, + 0.0035842375364154577, + 0.0027572307735681534, + 0.002880445681512356, + 0.0034346713218837976, + 0.0038378664758056402, + 0.002939498284831643, + 0.003188986098393798, + 0.006466597318649292, + 0.004350164905190468, + 0.0036687972024083138, + 0.002590789692476392, + 0.004579740110784769, + 0.003811108646914363, + 0.003612758358940482, + 0.00304840668104589, + 0.005650136154145002, + 0.0039274971932172775, + 0.003953225445002317, + 0.0045650736428797245, + 0.003896459238603711, + 0.004959185607731342, + 0.003962673246860504, + 0.00332921021617949, + 0.007412887644022703, + 0.010636433959007263, + 0.0032101268880069256, + 0.0032791169360280037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.01607564091682434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614679677411914 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003614679677411914, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614679677411914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16470867395401, + 0.16481059789657593, + 0.09377428889274597, + 0.14329753816127777, + 0.13790801167488098, + 0.1356656700372696, + 0.04121984541416168, + 0.03823961690068245, + 0.046703677624464035, + 0.0336720310151577 + ], + "credit_map": [ + 0.02286510169506073, + 0.02066393755376339, + 0.025012021884322166, + 0.02426593005657196, + 0.02426893077790737, + 0.018257856369018555, + 0.0123304957523942, + 0.017044411972165108, + 0.024418136104941368, + 0.013916690833866596, + 0.015828389674425125, + 0.013574471697211266, + 0.015799332410097122, + 0.04461716115474701, + 0.011257058009505272, + 0.025399353355169296, + 0.011124055832624435, + 0.00833254586905241, + 0.02081707864999771, + 0.007552198134362698, + 0.009050439111888409, + 0.018713142722845078, + 0.007946992293000221, + 0.010237842798233032, + 0.01872866228222847, + 0.01938348077237606, + 0.02244824543595314, + 0.020213423296809196, + 0.016678636893630028, + 0.018231190741062164, + 0.01245265081524849, + 0.015161242336034775, + 0.014052458107471466, + 0.016076045110821724, + 0.016574790701270103, + 0.019350042566657066, + 0.019606364890933037, + 0.01818782649934292, + 0.013195933774113655, + 0.02086455747485161, + 0.016528338193893433, + 0.015177278779447079, + 0.019636090844869614, + 0.015591918490827084, + 0.014932206831872463, + 0.018929217010736465, + 0.015741392970085144, + 0.019129224121570587, + 0.004168551880866289, + 0.005536471027880907, + 0.006288902834057808, + 0.0031754174269735813, + 0.007117222994565964, + 0.005351802799850702, + 0.005627990700304508, + 0.003953485284000635, + 0.004104342777282, + 0.0033663404174149036, + 0.0035087750293314457, + 0.003670933423563838, + 0.008246895857155323, + 0.0058400193229317665, + 0.005737938452512026, + 0.0037643732503056526, + 0.004350449424237013, + 0.005097055342048407, + 0.004839887842535973, + 0.004794015083462, + 0.00864442065358162, + 0.008192835375666618, + 0.005418166518211365, + 0.005366849713027477, + 0.00481026154011488, + 0.00406638765707612, + 0.0033244339283555746, + 0.0036443399731069803, + 0.005125182680785656, + 0.004441215191036463, + 0.004054792691022158, + 0.004205417353659868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.003614679677411914 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, upward-sweeping long shot, shot from the base of the mountain, steadily and slowly pulls back to reveal the full extent of the catastrophe. In the center of the frame, a majestic, precipitous volcano stands against a dark, crimson sky stained by volcanic ash. Its jagged crater rim becomes the source of destruction. A stream of incandescent, viscous, orange-red molten lava erupts with explosive force from the summit, forming a massive fountain that soars skyward, hurling countless flaming fragments of rock high into the sky. Meanwhile, a thick plume of ash and gas, like an awakening beast, twists and turns upward, its interior illuminated by the glow of the molten lava, punctuated by eerie flashes of red lightning. The scene is illuminated solely by this infernal eruption. The fiery orange-yellow contrasts sharply with the dark red of the sky and the inky black of the mountain, creating a suffocating, apocalyptic atmosphere imbued with primal destructive power.", + "seed": 11100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_9189_seed_11100_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": -0.22445724662393332, + "vlm_logprob": -0.22445724662393332, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.22445724662393332, + "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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8312411308288574 + ], + "opposite_logprob_tokens": [ + -0.5812411308288574 + ], + "logprob_reward": -0.8312411308288574, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8312411308288574, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07783722132444382, + 0.02988336980342865, + 0.03101508691906929, + 0.03220868483185768, + 0.05946590006351471, + 0.16874577105045319, + 0.280210018157959, + 0.21566776931285858, + 0.02267436496913433, + 0.0822918638586998 + ], + "credit_map": [ + 0.007742486894130707, + 0.006978187710046768, + 0.010051804594695568, + 0.012609832920134068, + 0.00688177440315485, + 0.009370514191687107, + 0.007221485022455454, + 0.016981130465865135, + 0.004239100497215986, + 0.003236607415601611, + 0.00328911398537457, + 0.0049888137727975845, + 0.0018924708710983396, + 0.004446102771908045, + 0.0023283965419977903, + 0.005462764296680689, + 0.003595632966607809, + 0.0022415274288505316, + 0.004525437019765377, + 0.0049202898517251015, + 0.002411018591374159, + 0.004432427231222391, + 0.0021819889079779387, + 0.00670676538720727, + 0.0035896438639611006, + 0.0038693402893841267, + 0.004164750222116709, + 0.004197115544229746, + 0.004477580077946186, + 0.0025686800945550203, + 0.0035922741517424583, + 0.005749301984906197, + 0.007256849203258753, + 0.00711028603836894, + 0.008559027686715126, + 0.00816162209957838, + 0.006617492064833641, + 0.008882585912942886, + 0.005061022471636534, + 0.007817012257874012, + 0.024292699992656708, + 0.017462722957134247, + 0.022231588140130043, + 0.015757907181978226, + 0.02083527110517025, + 0.035504408180713654, + 0.013193823397159576, + 0.01946733333170414, + 0.03198562189936638, + 0.008825724013149738, + 0.08413510024547577, + 0.015593254007399082, + 0.011882208287715912, + 0.10800375044345856, + 0.007597888354212046, + 0.01218645554035902, + 0.03126950562000275, + 0.008765938691794872, + 0.05120270699262619, + 0.013717668130993843, + 0.01164898369461298, + 0.07940629869699478, + 0.007621332071721554, + 0.01203532237559557, + 0.002661109669134021, + 0.002808190416544676, + 0.003972009755671024, + 0.0033472583163529634, + 0.0023900314699858427, + 0.0031643924303352833, + 0.0015655012102797627, + 0.002765872748568654, + 0.007482975721359253, + 0.01014160830527544, + 0.013099217787384987, + 0.012704707682132721, + 0.01258146483451128, + 0.009200683794915676, + 0.008035832084715366, + 0.009045378305017948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.8312411308288574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016000598669052 + ], + "opposite_logprob_tokens": [ + -1.7016000747680664 + ], + "logprob_reward": -0.2016000598669052, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016000598669052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17687548696994781, + 0.06515081971883774, + 0.04605015367269516, + 0.039412956684827805, + 0.0708429217338562, + 0.04506112262606621, + 0.3504841923713684, + 0.06754259765148163, + 0.04109368100762367, + 0.09748604148626328 + ], + "credit_map": [ + 0.021985379979014397, + 0.018008753657341003, + 0.017077859491109848, + 0.012339522130787373, + 0.025721997022628784, + 0.053310927003622055, + 0.008011693134903908, + 0.020419351756572723, + 0.00597607996314764, + 0.004479276016354561, + 0.011580497026443481, + 0.008290096186101437, + 0.005740826018154621, + 0.020476534962654114, + 0.0032423820812255144, + 0.0053651295602321625, + 0.005294044502079487, + 0.002381990198045969, + 0.007254988886415958, + 0.006926325615495443, + 0.00347351492382586, + 0.007285995874553919, + 0.0028328795451670885, + 0.010600416921079159, + 0.0034418783616274595, + 0.005899226292967796, + 0.005086381919682026, + 0.007801887113600969, + 0.004550807643681765, + 0.003785406006500125, + 0.0037841838784515858, + 0.005063187330961227, + 0.006839987821877003, + 0.009394879452884197, + 0.01110449805855751, + 0.006743209436535835, + 0.009879588149487972, + 0.015070714987814426, + 0.0059712836518883705, + 0.005838763900101185, + 0.007764345034956932, + 0.004204931203275919, + 0.005542553961277008, + 0.004094600677490234, + 0.005539296194911003, + 0.008662931621074677, + 0.0035799983888864517, + 0.005672463681548834, + 0.04683935269713402, + 0.024412475526332855, + 0.07946920394897461, + 0.027800356969237328, + 0.02098855935037136, + 0.12044496834278107, + 0.012271429412066936, + 0.01825784519314766, + 0.011351996101439, + 0.00797167420387268, + 0.008733524940907955, + 0.004456012509763241, + 0.005666963290423155, + 0.017179910093545914, + 0.0037072421982884407, + 0.008475272916257381, + 0.005418754648417234, + 0.004685915075242519, + 0.00703734066337347, + 0.0054923295974731445, + 0.004275510553270578, + 0.004999849945306778, + 0.0033096785191446543, + 0.0058742999099195, + 0.010735026560723782, + 0.010560089722275734, + 0.01548717450350523, + 0.015923159196972847, + 0.011826608330011368, + 0.012359555810689926, + 0.009150446392595768, + 0.01144398469477892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.2016000598669052 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062056127935647964 + ], + "opposite_logprob_tokens": [ + -2.812056064605713 + ], + "logprob_reward": -0.062056127935647964, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062056127935647964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23756591975688934, + 0.04103719815611839, + 0.05540890991687775, + 0.04737687110900879, + 0.03423710912466049, + 0.09258370101451874, + 0.22047601640224457, + 0.15022696554660797, + 0.037801846861839294, + 0.08328549563884735 + ], + "credit_map": [ + 0.026949869468808174, + 0.021712400019168854, + 0.02049594558775425, + 0.024569392204284668, + 0.03062031790614128, + 0.06549228727817535, + 0.014509964734315872, + 0.033215731382369995, + 0.005190382245928049, + 0.00476663326844573, + 0.00482171168550849, + 0.005081190727651119, + 0.0037326666060835123, + 0.008234958164393902, + 0.0023357535246759653, + 0.00687389587983489, + 0.007910922169685364, + 0.004053068812936544, + 0.01010869350284338, + 0.006274052895605564, + 0.005301421973854303, + 0.008212387561798096, + 0.0030382636468857527, + 0.0105100953951478, + 0.005178343970328569, + 0.005614347755908966, + 0.006817822810262442, + 0.006563948001712561, + 0.007032568566501141, + 0.005174282006919384, + 0.005090821534395218, + 0.005904729478061199, + 0.0036419895477592945, + 0.004167445469647646, + 0.005225810222327709, + 0.004358942154794931, + 0.004641635809093714, + 0.005028189159929752, + 0.0030153205152601004, + 0.004157770425081253, + 0.011587142944335938, + 0.01305991131812334, + 0.010950564406812191, + 0.009246254339814186, + 0.012670960277318954, + 0.015295916236937046, + 0.009081048890948296, + 0.010691897012293339, + 0.023784123361110687, + 0.01696421205997467, + 0.055937137454748154, + 0.017739616334438324, + 0.01147136464715004, + 0.07251615822315216, + 0.009228657931089401, + 0.012834731489419937, + 0.02157769165933132, + 0.008389562368392944, + 0.031764037907123566, + 0.01298784464597702, + 0.008867661468684673, + 0.049918144941329956, + 0.006770513951778412, + 0.009951482526957989, + 0.004311800934374332, + 0.005376523360610008, + 0.0063322847709059715, + 0.005005495622754097, + 0.0037079141475260258, + 0.004201587755233049, + 0.0037686650175601244, + 0.005097571760416031, + 0.008783518336713314, + 0.011775832623243332, + 0.013663851656019688, + 0.009761194698512554, + 0.009416834451258183, + 0.011709899641573429, + 0.007270416710525751, + 0.010903940536081791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.062056127935647964 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32651033997535706, + 0.06370148807764053, + 0.057966746389865875, + 0.03628232330083847, + 0.015077638439834118, + 0.03374724090099335, + 0.08632388710975647, + 0.10996688902378082, + 0.031929340213537216, + 0.23849406838417053 + ], + "credit_map": [ + 0.03811505436897278, + 0.031245816498994827, + 0.024709677323698997, + 0.03289835900068283, + 0.042670149356126785, + 0.09777681529521942, + 0.01950988918542862, + 0.03958461433649063, + 0.008596817962825298, + 0.004161762539297342, + 0.008715040981769562, + 0.007738586515188217, + 0.007148866541683674, + 0.01774377003312111, + 0.0027867103926837444, + 0.006809932179749012, + 0.006302225403487682, + 0.0032541840337216854, + 0.015885207802057266, + 0.00558512844145298, + 0.003992649260908365, + 0.013698806054890156, + 0.003126674797385931, + 0.006121878046542406, + 0.0046470798552036285, + 0.004201407544314861, + 0.005860939156264067, + 0.007063387893140316, + 0.002961712423712015, + 0.003838244127109647, + 0.0029867547564208508, + 0.004722800571471453, + 0.001779592945240438, + 0.0017525161383673549, + 0.001934441039338708, + 0.0019833571277558804, + 0.0018174154683947563, + 0.00211502006277442, + 0.0019436248112469912, + 0.0017516734078526497, + 0.0039025922305881977, + 0.004555020481348038, + 0.0037222288083285093, + 0.004582391120493412, + 0.004892164375633001, + 0.00467256223782897, + 0.0037636379711329937, + 0.0036566446069628, + 0.011850117705762386, + 0.0027729333378374577, + 0.02218036912381649, + 0.006148018408566713, + 0.004061587154865265, + 0.03236960619688034, + 0.002947015454992652, + 0.003994250204414129, + 0.01331492979079485, + 0.00657869316637516, + 0.02598770335316658, + 0.009655172936618328, + 0.00684240972623229, + 0.03739113360643387, + 0.004278420004993677, + 0.005918442737311125, + 0.003272479632869363, + 0.004264829680323601, + 0.003808331675827503, + 0.004500074312090874, + 0.0039801993407309055, + 0.004453575238585472, + 0.002520807320252061, + 0.005129044409841299, + 0.01922077313065529, + 0.043816909193992615, + 0.03264426812529564, + 0.02970229648053646, + 0.042217403650283813, + 0.020236553624272346, + 0.027917033061385155, + 0.022738855332136154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409455131739378 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.00409455131739378, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409455131739378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2827012836933136, + 0.06472691148519516, + 0.12611296772956848, + 0.07962893694639206, + 0.04127241671085358, + 0.06694181263446808, + 0.15895549952983856, + 0.06933849304914474, + 0.036605849862098694, + 0.07371582090854645 + ], + "credit_map": [ + 0.03421316295862198, + 0.024147797375917435, + 0.026235530152916908, + 0.0294769499450922, + 0.032246630638837814, + 0.08320675790309906, + 0.017546307295560837, + 0.035628121346235275, + 0.006925627589225769, + 0.00799825880676508, + 0.007697415538132191, + 0.011280420236289501, + 0.005963142029941082, + 0.007752158213406801, + 0.005550324451178312, + 0.011559562757611275, + 0.013250620104372501, + 0.007629856001585722, + 0.03103928454220295, + 0.01368582621216774, + 0.010023612529039383, + 0.024857908487319946, + 0.00931950006633997, + 0.016306370496749878, + 0.007441255263984203, + 0.010797057300806046, + 0.011308462359011173, + 0.009221130050718784, + 0.01152665913105011, + 0.008229123428463936, + 0.007658433634787798, + 0.013446814380586147, + 0.0050496552139520645, + 0.006663547363132238, + 0.005211732350289822, + 0.006050467491149902, + 0.004142685327678919, + 0.0046761659905314445, + 0.004849943332374096, + 0.004628223832696676, + 0.008075965568423271, + 0.009409490041434765, + 0.0073759788647294044, + 0.009203165769577026, + 0.008221489377319813, + 0.009571513161063194, + 0.007532726973295212, + 0.007551484741270542, + 0.020033299922943115, + 0.005954906810075045, + 0.04310966655611992, + 0.00909523107111454, + 0.008410069160163403, + 0.05974378436803818, + 0.005203195847570896, + 0.007405330892652273, + 0.006671923212707043, + 0.00820741057395935, + 0.013567536137998104, + 0.006753376219421625, + 0.007337877526879311, + 0.017641901969909668, + 0.003602386452257633, + 0.005556079093366861, + 0.0037532628048211336, + 0.006200659554451704, + 0.005020389333367348, + 0.003536676988005638, + 0.004635092802345753, + 0.006140449550002813, + 0.0023771394044160843, + 0.004942179657518864, + 0.006804401520639658, + 0.011784643866121769, + 0.011563528329133987, + 0.010355920530855656, + 0.007610832341015339, + 0.009810542687773705, + 0.0064489408396184444, + 0.00933701079338789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.00409455131739378 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, upward-sweeping long shot, shot from the base of the mountain, steadily and slowly pulls back to reveal the full extent of the catastrophe. In the center of the frame, a majestic, precipitous volcano stands against a dark, crimson sky stained by volcanic ash. Its jagged crater rim becomes the source of destruction. A stream of incandescent, viscous, orange-red molten lava erupts with explosive force from the summit, forming a massive fountain that soars skyward, hurling countless flaming fragments of rock high into the sky. Meanwhile, a thick plume of ash and gas, like an awakening beast, twists and turns upward, its interior illuminated by the glow of the molten lava, punctuated by eerie flashes of red lightning. The scene is illuminated solely by this infernal eruption. The fiery orange-yellow contrasts sharply with the dark red of the sky and the inky black of the mountain, creating a suffocating, apocalyptic atmosphere imbued with primal destructive power.", + "seed": 11101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_5/index_9189_seed_11101_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": -0.18610200693365186, + "vlm_logprob": -0.18610200693365186, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.18610200693365186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5305426120758057 + ], + "opposite_logprob_tokens": [ + -0.9055426120758057 + ], + "logprob_reward": -0.5305426120758057, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5305426120758057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10415849834680557, + 0.10664740949869156, + 0.03184227645397186, + 0.162832111120224, + 0.09666308015584946, + 0.16327957808971405, + 0.10289046913385391, + 0.05733312666416168, + 0.06396625936031342, + 0.11038718372583389 + ], + "credit_map": [ + 0.012295834720134735, + 0.011917957104742527, + 0.021125730127096176, + 0.012823582626879215, + 0.007832449860870838, + 0.01391869131475687, + 0.007947428151965141, + 0.01629682630300522, + 0.014985183253884315, + 0.007048026658594608, + 0.02193489484488964, + 0.016178755089640617, + 0.006765257101505995, + 0.016254033893346786, + 0.008782913908362389, + 0.014698338694870472, + 0.0032918660435825586, + 0.0037559587508440018, + 0.003981612157076597, + 0.0036115285474807024, + 0.0038745803758502007, + 0.004652068018913269, + 0.0049358876422047615, + 0.003738776780664921, + 0.014388179406523705, + 0.017237236723303795, + 0.019902924075722694, + 0.022008469328284264, + 0.01890714466571808, + 0.023905325680971146, + 0.026369646191596985, + 0.020113179460167885, + 0.012128712609410286, + 0.010744821280241013, + 0.00963177066296339, + 0.009767035022377968, + 0.012425048276782036, + 0.017048370093107224, + 0.013565083965659142, + 0.01135224662721157, + 0.029304977506399155, + 0.013620666228234768, + 0.02424064837396145, + 0.017086245119571686, + 0.014607510529458523, + 0.03523089364171028, + 0.013443493284285069, + 0.01574513129889965, + 0.021723506972193718, + 0.022346479818224907, + 0.012299527414143085, + 0.00967369507998228, + 0.011637331917881966, + 0.009442249312996864, + 0.007611280772835016, + 0.00815640576183796, + 0.006117077544331551, + 0.004596878308802843, + 0.008958782069385052, + 0.008144395425915718, + 0.007324310019612312, + 0.009736413136124611, + 0.00594665203243494, + 0.006508620921522379, + 0.006568879820406437, + 0.005529784597456455, + 0.008142948150634766, + 0.012330728583037853, + 0.007177179679274559, + 0.007969502359628677, + 0.006420658901333809, + 0.00982657726854086, + 0.01200089044868946, + 0.019005296751856804, + 0.011281908489763737, + 0.01535979937762022, + 0.015513403341174126, + 0.015995820984244347, + 0.008023916743695736, + 0.01320615317672491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.5305426120758057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870685398578644 + ], + "opposite_logprob_tokens": [ + -1.137068510055542 + ], + "logprob_reward": -0.3870685398578644, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870685398578644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14850759506225586, + 0.04872330650687218, + 0.032844435423612595, + 0.1799740195274353, + 0.14277005195617676, + 0.17580603063106537, + 0.0748123899102211, + 0.10220231115818024, + 0.046371087431907654, + 0.04798874258995056 + ], + "credit_map": [ + 0.020792989060282707, + 0.02950139530003071, + 0.01369822584092617, + 0.014468569308519363, + 0.019948570057749748, + 0.02004767581820488, + 0.01367376558482647, + 0.01637638919055462, + 0.005849296692758799, + 0.004017123952507973, + 0.007903259247541428, + 0.007820096798241138, + 0.00435095839202404, + 0.0062060728669166565, + 0.005770653951913118, + 0.006805845070630312, + 0.0042251888662576675, + 0.0034329283516854048, + 0.005658908747136593, + 0.003859755350276828, + 0.003504763590171933, + 0.003534774761646986, + 0.004981012083590031, + 0.003647102043032646, + 0.024209802970290184, + 0.023171160370111465, + 0.021116578951478004, + 0.017559444531798363, + 0.02571732923388481, + 0.027894319966435432, + 0.013793425634503365, + 0.026511970907449722, + 0.01766684092581272, + 0.015111721120774746, + 0.025409594178199768, + 0.024065222591161728, + 0.016295501962304115, + 0.013542696833610535, + 0.010393710806965828, + 0.02028476633131504, + 0.045966025441884995, + 0.008519488386809826, + 0.01953488402068615, + 0.015357707627117634, + 0.009481912478804588, + 0.046504825353622437, + 0.010384179651737213, + 0.02005700394511223, + 0.015317840501666069, + 0.01640341989696026, + 0.010230314917862415, + 0.006033712066709995, + 0.008662117645144463, + 0.006417854223400354, + 0.006836761254817247, + 0.0049103666096925735, + 0.012729641981422901, + 0.012775038368999958, + 0.014555064029991627, + 0.014399291947484016, + 0.01263703778386116, + 0.010595645755529404, + 0.011724095791578293, + 0.012786493636667728, + 0.0047773034311831, + 0.004348532762378454, + 0.004790525417774916, + 0.0073314537294209, + 0.006608361843973398, + 0.006580864544957876, + 0.005281142424792051, + 0.006652905605733395, + 0.0058795493096113205, + 0.007205537986010313, + 0.006115325726568699, + 0.006140953861176968, + 0.006409359164535999, + 0.004964297637343407, + 0.004176471382379532, + 0.007097246125340462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.3870685398578644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745188031345606 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745188031345606, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745188031345606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10928045213222504, + 0.24327239394187927, + 0.09750515967607498, + 0.20970690250396729, + 0.0507938414812088, + 0.14030048251152039, + 0.045040566474199295, + 0.030918346717953682, + 0.020295381546020508, + 0.05288650095462799 + ], + "credit_map": [ + 0.011755534447729588, + 0.02096569538116455, + 0.016014669090509415, + 0.007914229296147823, + 0.014746325090527534, + 0.013116966933012009, + 0.008452851325273514, + 0.016314173117280006, + 0.0402340367436409, + 0.019973022863268852, + 0.03634406253695488, + 0.029300572350621223, + 0.02357196994125843, + 0.037457264959812164, + 0.02309446781873703, + 0.033296987414360046, + 0.012415675446391106, + 0.006213850807398558, + 0.021563461050391197, + 0.010716213844716549, + 0.006417348049581051, + 0.014322662726044655, + 0.013990653678774834, + 0.011865288950502872, + 0.030870256945490837, + 0.015187273733317852, + 0.026280177757143974, + 0.026153426617383957, + 0.02066515013575554, + 0.041160859167575836, + 0.018674815073609352, + 0.03071492537856102, + 0.0056154620833694935, + 0.005231909453868866, + 0.007544756401330233, + 0.006315034348517656, + 0.005860733799636364, + 0.007110470440238714, + 0.007650545332580805, + 0.005464921239763498, + 0.020726054906845093, + 0.014144782908260822, + 0.020379289984703064, + 0.022082863375544548, + 0.01349591463804245, + 0.013626223430037498, + 0.012711389921605587, + 0.02313394844532013, + 0.007815389893949032, + 0.011127911508083344, + 0.005638458300381899, + 0.003945975564420223, + 0.0052684759721159935, + 0.003844280494377017, + 0.0037635844200849533, + 0.0036364872939884663, + 0.0029358675237745047, + 0.0027448912151157856, + 0.0038958322256803513, + 0.005043714307248592, + 0.004622869659215212, + 0.004430670291185379, + 0.0043030837550759315, + 0.002941415412351489, + 0.0023202779702842236, + 0.0017779103945940733, + 0.0025442028418183327, + 0.00415684562176466, + 0.001973262522369623, + 0.0022699732799082994, + 0.002719614654779434, + 0.0025332928635179996, + 0.0070695579051971436, + 0.005336225964128971, + 0.008313379250466824, + 0.007924385368824005, + 0.008368275128304958, + 0.0060347700491547585, + 0.004222034011036158, + 0.005617866292595863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.006745188031345606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29290711879730225, + 0.08086800575256348, + 0.17559561133384705, + 0.12774892151355743, + 0.01902940310537815, + 0.15367691218852997, + 0.06276634335517883, + 0.026238106191158295, + 0.026753390207886696, + 0.034416161477565765 + ], + "credit_map": [ + 0.041220523416996, + 0.04204172268509865, + 0.03724358603358269, + 0.022943273186683655, + 0.040161263197660446, + 0.04492270201444626, + 0.020575258880853653, + 0.043798819184303284, + 0.011061025783419609, + 0.007924147881567478, + 0.010975544340908527, + 0.010109157301485538, + 0.01181623712182045, + 0.010313566774129868, + 0.009663970209658146, + 0.009004364721477032, + 0.019926417618989944, + 0.022337963804602623, + 0.026095541194081306, + 0.021807778626680374, + 0.01810704916715622, + 0.025361442938447, + 0.016102828085422516, + 0.02585659734904766, + 0.016320815309882164, + 0.01726251095533371, + 0.014997588470578194, + 0.01268234383314848, + 0.017395902425050735, + 0.015909958630800247, + 0.011279228143393993, + 0.021900588646531105, + 0.002526278141885996, + 0.0018459116108715534, + 0.00241687404923141, + 0.001982603222131729, + 0.0016996752237901092, + 0.002842455869540572, + 0.0021293258760124445, + 0.003586280159652233, + 0.041716594249010086, + 0.008015032857656479, + 0.016239145770668983, + 0.008667131885886192, + 0.009185543283820152, + 0.04610917717218399, + 0.008614137768745422, + 0.015130153857171535, + 0.010185988619923592, + 0.015156402252614498, + 0.00808282196521759, + 0.005735715851187706, + 0.007302692160010338, + 0.0056450809352099895, + 0.004370861686766148, + 0.006286784540861845, + 0.004264777060598135, + 0.0032883980311453342, + 0.0023994215298444033, + 0.0033595862332731485, + 0.002956791315227747, + 0.003301725722849369, + 0.003398759989067912, + 0.0032686463091522455, + 0.003689858829602599, + 0.0021682651713490486, + 0.002496791072189808, + 0.004425423219799995, + 0.003993081860244274, + 0.0026280530728399754, + 0.0032426866237074137, + 0.004109232220798731, + 0.0036294024903327227, + 0.0073798843659460545, + 0.0026302780024707317, + 0.0044877962209284306, + 0.005024653859436512, + 0.0033629408571869135, + 0.0034165987744927406, + 0.004484610632061958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.004640405531972647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 upward-sweeping long shot starting from the base of the mountain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera steadily and slowly pull back to reveal the full extent of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a precipitous volcano with a jagged crater rim visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky dark and crimson, stained by volcanic ash, with a thick plume of ash and gas twisting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a stream of incandescent, orange-red molten lava erupting from the summit as a massive fountain hurling flaming rock fragments?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015132891712710261 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015132891712710261, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015132891712710261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14204774796962738, + 0.04019325599074364, + 0.14351552724838257, + 0.22750622034072876, + 0.06589142233133316, + 0.14479206502437592, + 0.06626183539628983, + 0.11657049506902695, + 0.021066410467028618, + 0.03215501457452774 + ], + "credit_map": [ + 0.018220441415905952, + 0.02339419163763523, + 0.0174484271556139, + 0.013605120591819286, + 0.01733909733593464, + 0.016338640823960304, + 0.01326720044016838, + 0.022434620186686516, + 0.005650867708027363, + 0.0035667980555444956, + 0.005958730820566416, + 0.0038249939680099487, + 0.0052918726578354836, + 0.0067440601997077465, + 0.002974357455968857, + 0.006181578617542982, + 0.018565716221928596, + 0.013473396189510822, + 0.024207210168242455, + 0.021517477929592133, + 0.015055115334689617, + 0.016085708513855934, + 0.015190832316875458, + 0.019420074298977852, + 0.03186364844441414, + 0.027359498664736748, + 0.027838822454214096, + 0.026301156729459763, + 0.03351449966430664, + 0.03550240024924278, + 0.013406636193394661, + 0.031719569116830826, + 0.009546741843223572, + 0.0074020856991410255, + 0.0104294428601861, + 0.01044498197734356, + 0.0052415053360164165, + 0.010936080478131771, + 0.00573649350553751, + 0.006154093891382217, + 0.017177818343043327, + 0.017570238560438156, + 0.021731184795498848, + 0.02173609286546707, + 0.018853701651096344, + 0.018134210258722305, + 0.014976071193814278, + 0.01461276225745678, + 0.010537499561905861, + 0.00552890170365572, + 0.010357706807553768, + 0.008633804507553577, + 0.006937889847904444, + 0.006685428321361542, + 0.006144910119473934, + 0.011435696855187416, + 0.011604820378124714, + 0.012950633652508259, + 0.01164295244961977, + 0.019439157098531723, + 0.017655149102211, + 0.01861594244837761, + 0.013222754932940006, + 0.011439084075391293, + 0.0019411706598475575, + 0.002077288692817092, + 0.002090176334604621, + 0.003724900772795081, + 0.0031378285493701696, + 0.003032251028344035, + 0.002945280633866787, + 0.002117513446137309, + 0.003285251324996352, + 0.00508995121344924, + 0.0027661165222525597, + 0.005900043528527021, + 0.0035797010641545057, + 0.0035371084231883287, + 0.003701388370245695, + 0.00429545296356082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.0015132891712710261 + } + ], + "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/0000055/group_6/index_tensor([9062, 9062])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_tensor([9062, 9062])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..664d2fab617da5087a2cd5cb0555d7e08c7ad6fe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_tensor([9062, 9062])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, taken from a slightly elevated angle, focuses on an elderly craftsman with graying hair, dressed in a faded dark blue cotton work shirt. His head bowed, his forehead crisscrossed with deep wrinkles, his brows slightly furrowed as he gazes intently at his work. His shirt sleeves are neatly rolled up to his elbows, revealing sturdy forearms etched with veins and age spots. A warm overhead light slants down from the upper right, perfectly illuminating his hands and the gradually forming wood carving, while casting soft shadows on the slightly cluttered woodworking workshop in the background, which is filled with various tools. One of his calloused hands firmly holds a square block of unpolished, light-colored wood. With the other, he deftly and precisely manipulates a razor-edged carving knife. With steady, deliberate force, the tip of the knife slowly pushes across the wood, shaving off tiny, curled chips as thin as a cicada's wing. These tiny chips flutter briefly in the beam of light before silently landing on his dark workbench. The entire scene exudes a sense of profound tranquility and concentration. The warm tones of light and the air's almost permeating wood fragrance combine to create a tranquil atmosphere of frozen time and imbued with the spirit of craftsmanship.", + "seed": 14564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14564_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": -0.13879335671663284, + "vlm_logprob": -0.13879335671663284, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.13879335671663284, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.11955952644348145 + ], + "opposite_logprob_tokens": [ + -2.2445595264434814 + ], + "logprob_reward": -0.11955952644348145, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11955952644348145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27978071570396423, + 0.08827721327543259, + 0.0437798872590065, + 0.03313324227929115, + 0.11383076012134552, + 0.0640430897474289, + 0.10174906998872757, + 0.07346836477518082, + 0.06776196509599686, + 0.13417567312717438 + ], + "credit_map": [ + 0.04978031665086746, + 0.04001064971089363, + 0.05246420204639435, + 0.027850061655044556, + 0.03969131037592888, + 0.025843817740678787, + 0.01905238814651966, + 0.02508796751499176, + 0.013210045173764229, + 0.012451868504285812, + 0.010885986499488354, + 0.008804379031062126, + 0.013023395091295242, + 0.012407612055540085, + 0.009362106211483479, + 0.008131819777190685, + 0.005127657204866409, + 0.00670640217140317, + 0.007078728172928095, + 0.005453372839838266, + 0.006624753586947918, + 0.0039256964810192585, + 0.0051513006910681725, + 0.0037119779735803604, + 0.004093044903129339, + 0.004659584723412991, + 0.003590560983866453, + 0.00289262761361897, + 0.00715789757668972, + 0.0029470915906131268, + 0.0040017664432525635, + 0.0037906672805547714, + 0.01849079690873623, + 0.016345622017979622, + 0.01619844138622284, + 0.009235713630914688, + 0.020952368155121803, + 0.007823275402188301, + 0.01095810066908598, + 0.013826444745063782, + 0.011228940449655056, + 0.008204548619687557, + 0.011449834331870079, + 0.007301881443709135, + 0.005551330745220184, + 0.004464464727789164, + 0.010460563004016876, + 0.005381530150771141, + 0.0159621462225914, + 0.01362090464681387, + 0.016589591279625893, + 0.01271823886781931, + 0.017925363034009933, + 0.008120821788907051, + 0.009357940405607224, + 0.00745405675843358, + 0.01091802679002285, + 0.009686844423413277, + 0.012769735418260098, + 0.01058725081384182, + 0.011637324467301369, + 0.005299318581819534, + 0.007393247447907925, + 0.0051766145043075085, + 0.008441336452960968, + 0.009759240783751011, + 0.008963868953287601, + 0.01278621330857277, + 0.008990147151052952, + 0.006044052075594664, + 0.005948768928647041, + 0.006828333716839552, + 0.022006480023264885, + 0.02569843828678131, + 0.033456116914749146, + 0.024344077333807945, + 0.008807419799268246, + 0.00605239300057292, + 0.00833155121654272, + 0.005479191429913044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.11955952644348145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.018224645406007767 + ], + "opposite_logprob_tokens": [ + -4.018224716186523 + ], + "logprob_reward": -0.018224645406007767, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018224645406007767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14286483824253082, + 0.13266974687576294, + 0.06575940549373627, + 0.06860711425542831, + 0.0766490027308464, + 0.1298324465751648, + 0.07636968791484833, + 0.14552360773086548, + 0.06699512153863907, + 0.09472903609275818 + ], + "credit_map": [ + 0.019761981442570686, + 0.01816421002149582, + 0.01595606654882431, + 0.011244854889810085, + 0.026293650269508362, + 0.014119571074843407, + 0.012224217876791954, + 0.025100281462073326, + 0.02458454854786396, + 0.016684487462043762, + 0.01869640313088894, + 0.014378217048943043, + 0.01865745708346367, + 0.010141849517822266, + 0.016900477930903435, + 0.012626294046640396, + 0.010228976607322693, + 0.008617423474788666, + 0.00915884505957365, + 0.007637815084308386, + 0.010627067647874355, + 0.005802527070045471, + 0.00733884796500206, + 0.006347897928208113, + 0.010224459692835808, + 0.01030123420059681, + 0.012718062847852707, + 0.004456778988242149, + 0.010634713806211948, + 0.00565315131098032, + 0.0062082745134830475, + 0.008410435169935226, + 0.009022500365972519, + 0.008617507293820381, + 0.013625011779367924, + 0.004957793280482292, + 0.009685234166681767, + 0.005566610489040613, + 0.010734024457633495, + 0.014440315775573254, + 0.02722736820578575, + 0.018463347107172012, + 0.02535843849182129, + 0.01343453023582697, + 0.011689931154251099, + 0.009200718253850937, + 0.015599205158650875, + 0.008858902379870415, + 0.010225147008895874, + 0.010494605638086796, + 0.009440887719392776, + 0.005819276906549931, + 0.020538071170449257, + 0.006842512637376785, + 0.006888267118483782, + 0.006120915524661541, + 0.021882856264710426, + 0.018822450190782547, + 0.02154192142188549, + 0.023482918739318848, + 0.016864394769072533, + 0.013258037157356739, + 0.018451958894729614, + 0.011219063773751259, + 0.010012040846049786, + 0.00936136208474636, + 0.008033977821469307, + 0.007847420871257782, + 0.012274897657334805, + 0.007456459570676088, + 0.005956463515758514, + 0.0060524968430399895, + 0.017027612775564194, + 0.018950263038277626, + 0.012454485520720482, + 0.009172213263809681, + 0.01240912638604641, + 0.0086423484608531, + 0.007023675367236137, + 0.009049301967024803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.018224645406007767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.014216337352991104 + ], + "opposite_logprob_tokens": [ + -4.264216423034668 + ], + "logprob_reward": -0.014216337352991104, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014216337352991104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23304201662540436, + 0.08222024142742157, + 0.04946038872003555, + 0.029975490644574165, + 0.04477659612894058, + 0.018751095980405807, + 0.040878262370824814, + 0.22515958547592163, + 0.12082646042108536, + 0.15490984916687012 + ], + "credit_map": [ + 0.02656649239361286, + 0.026609402149915695, + 0.028589261695742607, + 0.015094400383532047, + 0.05173507705330849, + 0.022334082052111626, + 0.021845011040568352, + 0.040268298238515854, + 0.011152845807373524, + 0.009718276560306549, + 0.00753421476110816, + 0.006358742713928223, + 0.012649578973650932, + 0.011520491912961006, + 0.00808981154114008, + 0.015196284279227257, + 0.006095675751566887, + 0.005786670837551355, + 0.005199488718062639, + 0.004811553750187159, + 0.009679088369011879, + 0.0051818666979670525, + 0.005372573155909777, + 0.007333471905440092, + 0.004060985054820776, + 0.004291749559342861, + 0.0037123360671103, + 0.0025643303524702787, + 0.006821630522608757, + 0.0025266739539802074, + 0.003455014666542411, + 0.002542771166190505, + 0.005438418127596378, + 0.006770933046936989, + 0.007377695757895708, + 0.003592836670577526, + 0.008852592669427395, + 0.004497603978961706, + 0.004480884876102209, + 0.0037656291387975216, + 0.0036755234468728304, + 0.002528778277337551, + 0.002632895251736045, + 0.0020954736974090338, + 0.002274298807606101, + 0.0014969265321269631, + 0.0019327132031321526, + 0.002114487113431096, + 0.004473603796213865, + 0.004721999168395996, + 0.00503241503611207, + 0.00743250735104084, + 0.005786082241684198, + 0.0032195623498409986, + 0.0047078682109713554, + 0.0055042230524122715, + 0.026610149070620537, + 0.0294918492436409, + 0.03752006217837334, + 0.051473721861839294, + 0.020107349380850792, + 0.015094507485628128, + 0.023165175691246986, + 0.021696787327528, + 0.015300385653972626, + 0.015635324642062187, + 0.020987970754504204, + 0.028171343728899956, + 0.011984236538410187, + 0.007889246568083763, + 0.012164042331278324, + 0.00869391206651926, + 0.027708247303962708, + 0.031151870265603065, + 0.025681156665086746, + 0.022829046472907066, + 0.01583104208111763, + 0.009869803674519062, + 0.010559833608567715, + 0.01127883791923523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.014216337352991104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.11299644410610199 + ], + "opposite_logprob_tokens": [ + -2.2379963397979736 + ], + "logprob_reward": -0.11299644410610199, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11299644410610199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30619922280311584, + 0.07150239497423172, + 0.055749669671058655, + 0.05404859781265259, + 0.04864013195037842, + 0.032718122005462646, + 0.023927373811602592, + 0.07594642043113708, + 0.05047086253762245, + 0.28079715371131897 + ], + "credit_map": [ + 0.032453104853630066, + 0.03526153042912483, + 0.03578602522611618, + 0.022463016211986542, + 0.05310387164354324, + 0.031052641570568085, + 0.030033595860004425, + 0.06604544073343277, + 0.010572901926934719, + 0.009147945791482925, + 0.010743770748376846, + 0.006232572253793478, + 0.008661234751343727, + 0.007012527901679277, + 0.007203045301139355, + 0.011928396299481392, + 0.00864278245717287, + 0.0066789728589355946, + 0.009228337556123734, + 0.0059159365482628345, + 0.006022165529429913, + 0.004495672415941954, + 0.00741195073351264, + 0.007353857159614563, + 0.005678032524883747, + 0.0068199425004422665, + 0.006263061426579952, + 0.004463194403797388, + 0.01377430371940136, + 0.004866228438913822, + 0.006470661144703627, + 0.005713173653930426, + 0.006276149768382311, + 0.006647830829024315, + 0.006340117659419775, + 0.004688289947807789, + 0.00812519621104002, + 0.003746778005734086, + 0.007573168259114027, + 0.005242600105702877, + 0.004394145216792822, + 0.004178905393928289, + 0.0043731932528316975, + 0.0038686594925820827, + 0.006199480500072241, + 0.0021381061524152756, + 0.0050877658650279045, + 0.002477868227288127, + 0.0032089739106595516, + 0.00325793563388288, + 0.003329368308186531, + 0.004339258186519146, + 0.002850556280463934, + 0.0016650233883410692, + 0.0033072177320718765, + 0.001969039672985673, + 0.008915594778954983, + 0.010341615416109562, + 0.016112972050905228, + 0.014170144684612751, + 0.007834790274500847, + 0.005247147288173437, + 0.00866373348981142, + 0.004660425707697868, + 0.006805090699344873, + 0.005745431408286095, + 0.007654514629393816, + 0.009865830652415752, + 0.00656156986951828, + 0.004791935905814171, + 0.005068203900009394, + 0.003978288732469082, + 0.04987603798508644, + 0.05033539608120918, + 0.06188312545418739, + 0.05132715776562691, + 0.017783286049962044, + 0.016531163826584816, + 0.018848663195967674, + 0.014212355017662048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.11299644410610199 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.4289698302745819 + ], + "opposite_logprob_tokens": [ + -1.0539698600769043 + ], + "logprob_reward": -0.4289698302745819, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289698302745819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19597142934799194, + 0.07183671742677689, + 0.05726790055632591, + 0.05322878435254097, + 0.02949092537164688, + 0.02356577105820179, + 0.034407418221235275, + 0.07231276482343674, + 0.11379655450582504, + 0.3481217324733734 + ], + "credit_map": [ + 0.024668997153639793, + 0.02178974263370037, + 0.023486841470003128, + 0.01261257752776146, + 0.026479646563529968, + 0.01590726710855961, + 0.024793917313218117, + 0.0462324433028698, + 0.008922383189201355, + 0.009983964264392853, + 0.007838767021894455, + 0.0048352875746786594, + 0.008545896038413048, + 0.008850234560668468, + 0.008384636603295803, + 0.014475546777248383, + 0.006958717945963144, + 0.008195179514586926, + 0.00547423679381609, + 0.003980507142841816, + 0.00816202349960804, + 0.005560120102018118, + 0.006558889988809824, + 0.012378225103020668, + 0.006159109063446522, + 0.007871512323617935, + 0.004817126784473658, + 0.004062287975102663, + 0.012371992692351341, + 0.005274963099509478, + 0.004934382159262896, + 0.007737410720437765, + 0.003441921202465892, + 0.0037029199302196503, + 0.0038508514408022165, + 0.002972399117425084, + 0.0037093032151460648, + 0.002900840248912573, + 0.004069278948009014, + 0.004843410570174456, + 0.0029463607352226973, + 0.002712713088840246, + 0.0032174517400562763, + 0.002204340184107423, + 0.0033060284331440926, + 0.002084595151245594, + 0.003844508668407798, + 0.003249774221330881, + 0.004450284875929356, + 0.00409218855202198, + 0.00473459716886282, + 0.005493880715221167, + 0.006309154909104109, + 0.002349184826016426, + 0.0045316219329833984, + 0.0024465059395879507, + 0.009594106115400791, + 0.009699168615043163, + 0.010598643682897091, + 0.013261187821626663, + 0.006293110083788633, + 0.005347306374460459, + 0.009111888706684113, + 0.008407356217503548, + 0.011510300450026989, + 0.013660429045557976, + 0.0172010138630867, + 0.028815511614084244, + 0.01462714932858944, + 0.008164539933204651, + 0.009048514999449253, + 0.010769096203148365, + 0.05581308528780937, + 0.06496431678533554, + 0.06137020140886307, + 0.05042574927210808, + 0.032090749591588974, + 0.025609787553548813, + 0.025410182774066925, + 0.03243767097592354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.4289698302745819 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slightly elevated angle, focuses on an elderly craftsman with graying hair, dressed in a faded dark blue cotton work shirt. His head bowed, his forehead crisscrossed with deep wrinkles, his brows slightly furrowed as he gazes intently at his work. His shirt sleeves are neatly rolled up to his elbows, revealing sturdy forearms etched with veins and age spots. A warm overhead light slants down from the upper right, perfectly illuminating his hands and the gradually forming wood carving, while casting soft shadows on the slightly cluttered woodworking workshop in the background, which is filled with various tools. One of his calloused hands firmly holds a square block of unpolished, light-colored wood. With the other, he deftly and precisely manipulates a razor-edged carving knife. With steady, deliberate force, the tip of the knife slowly pushes across the wood, shaving off tiny, curled chips as thin as a cicada's wing. These tiny chips flutter briefly in the beam of light before silently landing on his dark workbench. The entire scene exudes a sense of profound tranquility and concentration. The warm tones of light and the air's almost permeating wood fragrance combine to create a tranquil atmosphere of frozen time and imbued with the spirit of craftsmanship.", + "seed": 14565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14565_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": -0.04971076287329197, + "vlm_logprob": -0.04971076287329197, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04971076287329197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.045553747564554214 + ], + "opposite_logprob_tokens": [ + -3.170553684234619 + ], + "logprob_reward": -0.045553747564554214, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045553747564554214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20334173738956451, + 0.20626288652420044, + 0.07525046169757843, + 0.24107275903224945, + 0.10954216867685318, + 0.03503543511033058, + 0.03687606379389763, + 0.035073161125183105, + 0.0313374325633049, + 0.026207948103547096 + ], + "credit_map": [ + 0.029964275658130646, + 0.02850058302283287, + 0.03499972075223923, + 0.019956419244408607, + 0.01955745555460453, + 0.018629437312483788, + 0.020566273480653763, + 0.031167563050985336, + 0.02653048001229763, + 0.06496916711330414, + 0.020860867574810982, + 0.042580362409353256, + 0.012108922004699707, + 0.012759477831423283, + 0.010908088646829128, + 0.015545505098998547, + 0.006860142573714256, + 0.009258218109607697, + 0.005836799740791321, + 0.027207517996430397, + 0.007623376790434122, + 0.0047524115070700645, + 0.00888288114219904, + 0.004829106852412224, + 0.03155047819018364, + 0.032217975705862045, + 0.08103013038635254, + 0.047434523701667786, + 0.016073288396000862, + 0.010818053036928177, + 0.012491986155509949, + 0.00945630669593811, + 0.027904780581593513, + 0.01607542671263218, + 0.03764016926288605, + 0.008543988689780235, + 0.007420435547828674, + 0.0041993046179413795, + 0.003161378437653184, + 0.004596673417836428, + 0.004688771907240152, + 0.0035991331096738577, + 0.005401581525802612, + 0.0062906863167881966, + 0.003534188959747553, + 0.0027832663618028164, + 0.006139725912362337, + 0.002598075894638896, + 0.0060914685018360615, + 0.004720199387520552, + 0.006094092037528753, + 0.004645016044378281, + 0.005386611446738243, + 0.003292683744803071, + 0.004131281282752752, + 0.002514705527573824, + 0.005196124315261841, + 0.0034254419151693583, + 0.0034378687851130962, + 0.004737482406198978, + 0.007552545052021742, + 0.003050740109756589, + 0.0032821055501699448, + 0.004390845075249672, + 0.005010913126170635, + 0.002452881308272481, + 0.0027498216368258, + 0.009795645251870155, + 0.002870741533115506, + 0.00323912943713367, + 0.002889548195526004, + 0.002328745787963271, + 0.002602682216092944, + 0.0035415226593613625, + 0.004276389721781015, + 0.003569971537217498, + 0.0035114057827740908, + 0.003491166280582547, + 0.002418141346424818, + 0.0027966671623289585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.045553747564554214 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.020597605034708977 + ], + "opposite_logprob_tokens": [ + -3.8955976963043213 + ], + "logprob_reward": -0.020597605034708977, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597605034708977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09776879847049713, + 0.1862279623746872, + 0.06248823553323746, + 0.03435734286904335, + 0.4208616614341736, + 0.07847417145967484, + 0.0373300164937973, + 0.028182514011859894, + 0.018709734082221985, + 0.03559952229261398 + ], + "credit_map": [ + 0.013821820728480816, + 0.01656717248260975, + 0.017667677253484726, + 0.008475967682898045, + 0.012112872675061226, + 0.009131522849202156, + 0.00869635958224535, + 0.01129541452974081, + 0.04532160237431526, + 0.07022687792778015, + 0.029063355177640915, + 0.009569285437464714, + 0.01141386479139328, + 0.00799711886793375, + 0.00601740600541234, + 0.00661845225840807, + 0.0058259000070393085, + 0.0069512310437858105, + 0.005045488942414522, + 0.02466951310634613, + 0.00525872316211462, + 0.0027206172235310078, + 0.009175625629723072, + 0.0028411350212991238, + 0.002452586079016328, + 0.004154682159423828, + 0.008439665660262108, + 0.008325467817485332, + 0.0033592742402106524, + 0.0026482122484594584, + 0.0031272636260837317, + 0.0018501932499930263, + 0.126441091299057, + 0.05365189537405968, + 0.1589706391096115, + 0.03414035588502884, + 0.015418007038533688, + 0.010444958694279194, + 0.011638634838163853, + 0.01015608198940754, + 0.010146431624889374, + 0.012173512950539589, + 0.017516138032078743, + 0.007839391939342022, + 0.010904526337981224, + 0.007294113282114267, + 0.00779060460627079, + 0.004809453152120113, + 0.005487274844199419, + 0.005275699310004711, + 0.005365975666791201, + 0.003349588718265295, + 0.004911196883767843, + 0.0059777237474918365, + 0.003084952011704445, + 0.0038776046130806208, + 0.0037123537622392178, + 0.0034854677505791187, + 0.0042357430793344975, + 0.0029355769511312246, + 0.0037678151857107878, + 0.004015981685370207, + 0.0027022704016417265, + 0.003327305894345045, + 0.0031311640050262213, + 0.0016189838061109185, + 0.0024867469910532236, + 0.0036309293936938047, + 0.002862930065020919, + 0.001915391650982201, + 0.0017772992141544819, + 0.001286288141272962, + 0.004347995854914188, + 0.004907798022031784, + 0.006509858183562756, + 0.005224138032644987, + 0.004493289161473513, + 0.0024591502733528614, + 0.003847452811896801, + 0.0038098411168903112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.020597605034708977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.009789552539587021 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789552539587021, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789552539587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16436411440372467, + 0.11545094102621078, + 0.09597902745008469, + 0.08512461930513382, + 0.2547086775302887, + 0.04357079043984413, + 0.09678208827972412, + 0.04489113390445709, + 0.04010018706321716, + 0.05902845039963722 + ], + "credit_map": [ + 0.022412927821278572, + 0.02409963123500347, + 0.031137820333242416, + 0.015062780119478703, + 0.016584092751145363, + 0.019938480108976364, + 0.014771303161978722, + 0.020357079803943634, + 0.009030908346176147, + 0.019365185871720314, + 0.009170557372272015, + 0.04181161895394325, + 0.00922994315624237, + 0.00864444300532341, + 0.011139752343297005, + 0.007058524992316961, + 0.00821366161108017, + 0.012028025463223457, + 0.006661905441433191, + 0.032853517681360245, + 0.009350725449621677, + 0.007560554891824722, + 0.01139635406434536, + 0.00791428703814745, + 0.012269262224435806, + 0.011024890467524529, + 0.027055662125349045, + 0.009753919206559658, + 0.005766300484538078, + 0.0058404612354934216, + 0.005858963821083307, + 0.007555161602795124, + 0.08081234246492386, + 0.030053984373807907, + 0.09575200080871582, + 0.015486651100218296, + 0.010220104828476906, + 0.007847056724131107, + 0.0074351150542497635, + 0.007101424969732761, + 0.006400854792445898, + 0.0034723482094705105, + 0.006600181572139263, + 0.004412530455738306, + 0.005443231668323278, + 0.007015781477093697, + 0.005077442619949579, + 0.005148420576006174, + 0.01577923074364662, + 0.009525597095489502, + 0.013077371753752232, + 0.007645019330084324, + 0.01306708063930273, + 0.01738460175693035, + 0.008515752851963043, + 0.011787437833845615, + 0.006906984839588404, + 0.004114773590117693, + 0.0053483955562114716, + 0.006912844255566597, + 0.006021228153258562, + 0.005609431304037571, + 0.005257135257124901, + 0.0047203414142131805, + 0.0067540546879172325, + 0.00344679388217628, + 0.004724304191768169, + 0.009223057888448238, + 0.004941433668136597, + 0.004017469007521868, + 0.0038814775180071592, + 0.003111595753580332, + 0.006325185764580965, + 0.006734300870448351, + 0.009407352656126022, + 0.007867119275033474, + 0.009289499372243881, + 0.006875974591821432, + 0.005168383475393057, + 0.007360632065683603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.009789552539587021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.029807809740304947 + ], + "opposite_logprob_tokens": [ + -3.5298078060150146 + ], + "logprob_reward": -0.029807809740304947, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807809740304947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16801470518112183, + 0.18348954617977142, + 0.033040959388017654, + 0.11714272201061249, + 0.14962169528007507, + 0.04118746146559715, + 0.08132095634937286, + 0.12954337894916534, + 0.05722407251596451, + 0.039414480328559875 + ], + "credit_map": [ + 0.020203575491905212, + 0.017613569274544716, + 0.021345356479287148, + 0.019472382962703705, + 0.02202005870640278, + 0.01967787928879261, + 0.025393137708306313, + 0.02228875458240509, + 0.02486748993396759, + 0.060758769512176514, + 0.02008538879454136, + 0.03556891903281212, + 0.010616766288876534, + 0.009666706435382366, + 0.007820147089660168, + 0.014105369336903095, + 0.004382426850497723, + 0.003749268827959895, + 0.004459402523934841, + 0.0040609524585306644, + 0.004486893303692341, + 0.003224046668037772, + 0.004741248209029436, + 0.0039367214776575565, + 0.008379141800105572, + 0.01222673337906599, + 0.025232229381799698, + 0.03045472502708435, + 0.00969670806080103, + 0.011605228297412395, + 0.010541136376559734, + 0.009006827138364315, + 0.03348413482308388, + 0.018778815865516663, + 0.04254783317446709, + 0.018074534833431244, + 0.011316931806504726, + 0.00895099900662899, + 0.009246336296200752, + 0.007222101092338562, + 0.005243273917585611, + 0.004417775198817253, + 0.004535331856459379, + 0.005229907110333443, + 0.00559975765645504, + 0.0054929908365011215, + 0.005687444470822811, + 0.004980979487299919, + 0.012579729780554771, + 0.00934587512165308, + 0.009158042259514332, + 0.006757684051990509, + 0.011232786811888218, + 0.013462402857840061, + 0.007389092352241278, + 0.011395350098609924, + 0.01889353059232235, + 0.01585260219871998, + 0.015827372670173645, + 0.011639955453574657, + 0.025198299437761307, + 0.016393670812249184, + 0.011506163515150547, + 0.014231793582439423, + 0.008944670669734478, + 0.005215936340391636, + 0.0075892056338489056, + 0.00699516711756587, + 0.011713596060872078, + 0.007545101456344128, + 0.005545756313949823, + 0.0036746372934430838, + 0.004518942441791296, + 0.003943521063774824, + 0.003890946740284562, + 0.006681251339614391, + 0.006165075581520796, + 0.004508808720856905, + 0.004121887031942606, + 0.0055840457789599895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.029807809740304947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.1428050994873047 + ], + "opposite_logprob_tokens": [ + -2.0178050994873047 + ], + "logprob_reward": -0.1428050994873047, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428050994873047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07672339677810669, + 0.08679806441068649, + 0.04545459523797035, + 0.07481785863637924, + 0.30913403630256653, + 0.07822684198617935, + 0.1592121720314026, + 0.0538502037525177, + 0.018697617575526237, + 0.09708528220653534 + ], + "credit_map": [ + 0.010051664896309376, + 0.012154800817370415, + 0.010728084482252598, + 0.006807873956859112, + 0.008278819732367992, + 0.01140464935451746, + 0.005884163081645966, + 0.011413333006203175, + 0.00893725361675024, + 0.01682351343333721, + 0.006508388090878725, + 0.025723949074745178, + 0.005847310181707144, + 0.007642634678632021, + 0.0064606075175106525, + 0.008854400366544724, + 0.0038132723420858383, + 0.00522804819047451, + 0.0028935454320162535, + 0.01669410988688469, + 0.004362418781965971, + 0.0027634422294795513, + 0.006093793082982302, + 0.0036059613339602947, + 0.008970565162599087, + 0.007306080777198076, + 0.008093081414699554, + 0.015435008332133293, + 0.008063012734055519, + 0.009437446482479572, + 0.011116459034383297, + 0.006396202836185694, + 0.08977585285902023, + 0.03906800597906113, + 0.11240894347429276, + 0.021853351965546608, + 0.013765229843556881, + 0.013737346045672894, + 0.008320709690451622, + 0.010204567573964596, + 0.012469182722270489, + 0.005841896403580904, + 0.008883853442966938, + 0.006461139302700758, + 0.009321392513811588, + 0.01376835722476244, + 0.011759980581700802, + 0.009721033275127411, + 0.022116197273135185, + 0.01809488609433174, + 0.018268421292304993, + 0.0168407391756773, + 0.018907975405454636, + 0.02651277557015419, + 0.018516914919018745, + 0.01995423622429371, + 0.007694277446717024, + 0.005473893601447344, + 0.00545599265024066, + 0.005657044239342213, + 0.008615761063992977, + 0.010191580280661583, + 0.004593330901116133, + 0.006168316584080458, + 0.0030010209884494543, + 0.001648863428272307, + 0.002108160173520446, + 0.004891854710876942, + 0.0023689279332756996, + 0.0017069884343072772, + 0.0017731654224917293, + 0.0011986326426267624, + 0.01061997190117836, + 0.012494794093072414, + 0.012823183089494705, + 0.013131489977240562, + 0.015555301681160927, + 0.013250431045889854, + 0.007168966345489025, + 0.012041134759783745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.1428050994873047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slightly elevated angle, focuses on an elderly craftsman with graying hair, dressed in a faded dark blue cotton work shirt. His head bowed, his forehead crisscrossed with deep wrinkles, his brows slightly furrowed as he gazes intently at his work. His shirt sleeves are neatly rolled up to his elbows, revealing sturdy forearms etched with veins and age spots. A warm overhead light slants down from the upper right, perfectly illuminating his hands and the gradually forming wood carving, while casting soft shadows on the slightly cluttered woodworking workshop in the background, which is filled with various tools. One of his calloused hands firmly holds a square block of unpolished, light-colored wood. With the other, he deftly and precisely manipulates a razor-edged carving knife. With steady, deliberate force, the tip of the knife slowly pushes across the wood, shaving off tiny, curled chips as thin as a cicada's wing. These tiny chips flutter briefly in the beam of light before silently landing on his dark workbench. The entire scene exudes a sense of profound tranquility and concentration. The warm tones of light and the air's almost permeating wood fragrance combine to create a tranquil atmosphere of frozen time and imbued with the spirit of craftsmanship.", + "seed": 14566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14566_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": -0.041673199646174906, + "vlm_logprob": -0.041673199646174906, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.041673199646174906, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.058165498077869415 + ], + "opposite_logprob_tokens": [ + -2.9331655502319336 + ], + "logprob_reward": -0.058165498077869415, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.058165498077869415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4350883960723877, + 0.10428441315889359, + 0.0701105073094368, + 0.04427960887551308, + 0.08834430575370789, + 0.06761457771062851, + 0.07161068916320801, + 0.044597260653972626, + 0.04995442554354668, + 0.0241159126162529 + ], + "credit_map": [ + 0.026249408721923828, + 0.05662360414862633, + 0.046441320329904556, + 0.03482820838689804, + 0.10782992094755173, + 0.0593712255358696, + 0.06902039796113968, + 0.034724246710538864, + 0.008401750586926937, + 0.016081340610980988, + 0.013210163451731205, + 0.009988554753363132, + 0.02225792407989502, + 0.013634989969432354, + 0.010923071764409542, + 0.009786597453057766, + 0.007787203881889582, + 0.011075073853135109, + 0.008898586966097355, + 0.006462607067078352, + 0.012505307793617249, + 0.009014680050313473, + 0.006690802052617073, + 0.007676233071833849, + 0.0028903724160045385, + 0.006498653441667557, + 0.004654900636523962, + 0.003960015717893839, + 0.008525676093995571, + 0.005704744253307581, + 0.005064026452600956, + 0.006981213577091694, + 0.013524799607694149, + 0.012024527415633202, + 0.010574216023087502, + 0.006347564049065113, + 0.019978616386651993, + 0.011187711730599403, + 0.006603570654988289, + 0.008103285916149616, + 0.008071882650256157, + 0.008081954903900623, + 0.006156011018902063, + 0.003870986169204116, + 0.013456092216074467, + 0.00900461059063673, + 0.005299975164234638, + 0.013673055917024612, + 0.00612475024536252, + 0.01037110947072506, + 0.005282961763441563, + 0.005174967460334301, + 0.01853913627564907, + 0.008802211843430996, + 0.006833639927208424, + 0.010481903329491615, + 0.006205812096595764, + 0.005920868832617998, + 0.00802310649305582, + 0.004216786939650774, + 0.007283328101038933, + 0.006084228400141001, + 0.003563421778380871, + 0.0032997007947415113, + 0.0038575055077672005, + 0.006083344575017691, + 0.003977956250309944, + 0.003973338752985001, + 0.014478183351457119, + 0.006893291138112545, + 0.003374351654201746, + 0.007316442206501961, + 0.002565642585977912, + 0.004794824402779341, + 0.002843811409547925, + 0.0022500576451420784, + 0.003333709668368101, + 0.0031577988993376493, + 0.002029625466093421, + 0.003140437649562955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.058165498077869415 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.020611269399523735 + ], + "opposite_logprob_tokens": [ + -3.89561128616333 + ], + "logprob_reward": -0.020611269399523735, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611269399523735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30884701013565063, + 0.18700553476810455, + 0.08655381202697754, + 0.02856738306581974, + 0.054123371839523315, + 0.037117645144462585, + 0.058378539979457855, + 0.16231325268745422, + 0.024744151160120964, + 0.052349191159009933 + ], + "credit_map": [ + 0.02099345438182354, + 0.036087702959775925, + 0.030497293919324875, + 0.026545118540525436, + 0.062119726091623306, + 0.04017701372504234, + 0.05695662274956703, + 0.0354701392352581, + 0.015554165467619896, + 0.025420770049095154, + 0.02227703481912613, + 0.022412266582250595, + 0.027686387300491333, + 0.020692134276032448, + 0.03189384564757347, + 0.021068952977657318, + 0.00943164061754942, + 0.01541083212941885, + 0.009577198885381222, + 0.009220338426530361, + 0.013155006803572178, + 0.011450088582932949, + 0.008780174888670444, + 0.009528551250696182, + 0.002329274546355009, + 0.004839364904910326, + 0.0030133258551359177, + 0.003883102908730507, + 0.004520627669990063, + 0.004083678126335144, + 0.003380340989679098, + 0.002517673186957836, + 0.006380678620189428, + 0.007308982778340578, + 0.005194393452256918, + 0.004432331770658493, + 0.009494666010141373, + 0.006314761471003294, + 0.006366400513797998, + 0.008631165139377117, + 0.004778340924531221, + 0.005498024169355631, + 0.0035255909897387028, + 0.0025863274931907654, + 0.0054543293081223965, + 0.005048219580203295, + 0.003264357103034854, + 0.006962464191019535, + 0.00646197097375989, + 0.010505556128919125, + 0.006103754974901676, + 0.007297388277947903, + 0.009690423496067524, + 0.006757630966603756, + 0.005646931007504463, + 0.0059148939326405525, + 0.022445539012551308, + 0.02116398699581623, + 0.025613270699977875, + 0.013187726959586143, + 0.03156343847513199, + 0.020343847572803497, + 0.010845252312719822, + 0.01715022325515747, + 0.0028338953852653503, + 0.002870821161195636, + 0.0030532379169017076, + 0.002707595704123378, + 0.0026254693511873484, + 0.0031456188298761845, + 0.0032111648470163345, + 0.004296350292861462, + 0.004215410444885492, + 0.008197682909667492, + 0.005203655455261469, + 0.004590505734086037, + 0.013648062944412231, + 0.007676974404603243, + 0.003984183073043823, + 0.00483272410929203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020611269399523735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.012553364969789982 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553364969789982, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553364969789982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20913244783878326, + 0.15792620182037354, + 0.10001832991838455, + 0.07196126878261566, + 0.14080819487571716, + 0.07053177803754807, + 0.0841970443725586, + 0.053586095571517944, + 0.04959438368678093, + 0.0622442364692688 + ], + "credit_map": [ + 0.014138370752334595, + 0.021179649978876114, + 0.03052169643342495, + 0.016652196645736694, + 0.032282836735248566, + 0.024360984563827515, + 0.039333175867795944, + 0.03066352568566799, + 0.01943398080766201, + 0.016224203631281853, + 0.019657472148537636, + 0.011968771927058697, + 0.034195128828287125, + 0.01849738135933876, + 0.01961543783545494, + 0.01833382435142994, + 0.010404834523797035, + 0.012127385474741459, + 0.010251078754663467, + 0.010071544907987118, + 0.0156612079590559, + 0.011602353304624557, + 0.015216565690934658, + 0.014683364890515804, + 0.006363329943269491, + 0.010413417592644691, + 0.005909331142902374, + 0.005591524299234152, + 0.012293471023440361, + 0.009069337509572506, + 0.009071063250303268, + 0.013249794021248817, + 0.016791382804512978, + 0.017020514234900475, + 0.013255040161311626, + 0.009871014393866062, + 0.032581884413957596, + 0.017260994762182236, + 0.012579431757330894, + 0.02144794724881649, + 0.008249226957559586, + 0.007798958569765091, + 0.006173992063850164, + 0.005958531051874161, + 0.01259531732648611, + 0.010292408987879753, + 0.005546012427657843, + 0.013917329721152782, + 0.010812895372509956, + 0.009519098326563835, + 0.010632829740643501, + 0.007752955891191959, + 0.014345575124025345, + 0.010538286529481411, + 0.008689322508871555, + 0.011906086467206478, + 0.00766025111079216, + 0.009071055799722672, + 0.007099835202097893, + 0.004777524620294571, + 0.008209853433072567, + 0.006460926495492458, + 0.004573460668325424, + 0.0057331835851073265, + 0.006251940969377756, + 0.006689922884106636, + 0.00690611032769084, + 0.0040766275487840176, + 0.006270534358918667, + 0.00619844626635313, + 0.004658535122871399, + 0.008542264811694622, + 0.004212754312902689, + 0.008173213340342045, + 0.0076853190548717976, + 0.009577461518347263, + 0.008552064187824726, + 0.0077806697227060795, + 0.0073088002391159534, + 0.008953951299190521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.012553364969789982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.05498713254928589 + ], + "opposite_logprob_tokens": [ + -2.9299871921539307 + ], + "logprob_reward": -0.05498713254928589, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498713254928589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2089027613401413, + 0.18670673668384552, + 0.1969778835773468, + 0.045884568244218826, + 0.12192206084728241, + 0.04082730785012245, + 0.048523321747779846, + 0.06892137974500656, + 0.036480095237493515, + 0.044853921979665756 + ], + "credit_map": [ + 0.011521585285663605, + 0.03140529990196228, + 0.02243117243051529, + 0.01827181875705719, + 0.03621940687298775, + 0.02249332331120968, + 0.0326089970767498, + 0.033951178193092346, + 0.013584106229245663, + 0.020467540249228477, + 0.01967431604862213, + 0.013565797358751297, + 0.036861371248960495, + 0.020779909566044807, + 0.021648557856678963, + 0.04012514650821686, + 0.01647748425602913, + 0.019738316535949707, + 0.01998664066195488, + 0.017661340534687042, + 0.03600355610251427, + 0.025463569909334183, + 0.028652485460042953, + 0.032994478940963745, + 0.0036626067012548447, + 0.00789868924766779, + 0.00424763560295105, + 0.004461176227778196, + 0.00727473758161068, + 0.005792065057903528, + 0.004898122977465391, + 0.007649533916264772, + 0.00967852957546711, + 0.015035933814942837, + 0.011259152553975582, + 0.008557469584047794, + 0.02442900463938713, + 0.014986750669777393, + 0.014525233767926693, + 0.023449981585144997, + 0.004130496643483639, + 0.007152173202484846, + 0.003273853100836277, + 0.003112182253971696, + 0.007165706716477871, + 0.005409058183431625, + 0.004155337810516357, + 0.006428502034395933, + 0.004555823281407356, + 0.00701631885021925, + 0.0043993499130010605, + 0.0028097759932279587, + 0.009133621118962765, + 0.006028192583471537, + 0.005499416030943394, + 0.009080825373530388, + 0.007760099601000547, + 0.008663566783070564, + 0.010092312470078468, + 0.005948312114924192, + 0.011008384637534618, + 0.00877260323613882, + 0.008425909094512463, + 0.008250197395682335, + 0.003139140084385872, + 0.005121543072164059, + 0.0046063922345638275, + 0.003497257363051176, + 0.0064355190843343735, + 0.0047561246901750565, + 0.0052335080690681934, + 0.00369060761295259, + 0.0034489077515900135, + 0.007735243067145348, + 0.0038076930213719606, + 0.0036463697906583548, + 0.008918637409806252, + 0.006095839664340019, + 0.00485654640942812, + 0.006344683934003115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05498713254928589 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.06204873323440552 + ], + "opposite_logprob_tokens": [ + -2.8120486736297607 + ], + "logprob_reward": -0.06204873323440552, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204873323440552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15480996668338776, + 0.2630394399166107, + 0.27975037693977356, + 0.09765297174453735, + 0.05861775949597359, + 0.04954433813691139, + 0.029937928542494774, + 0.03187697008252144, + 0.019834449514746666, + 0.014935752376914024 + ], + "credit_map": [ + 0.010531744919717312, + 0.018337475135922432, + 0.01660943776369095, + 0.0140284588560462, + 0.02616024948656559, + 0.018591444939374924, + 0.026066087186336517, + 0.024485070258378983, + 0.016913220286369324, + 0.03575519844889641, + 0.0219716876745224, + 0.019334975630044937, + 0.06325990706682205, + 0.03251897543668747, + 0.03166935592889786, + 0.04161614179611206, + 0.023164592683315277, + 0.03261527419090271, + 0.02638145722448826, + 0.021434098482131958, + 0.05322752892971039, + 0.036262672394514084, + 0.03225884214043617, + 0.05440593138337135, + 0.0075064473785459995, + 0.015066326595842838, + 0.009777350351214409, + 0.010373663157224655, + 0.025392306968569756, + 0.012885235249996185, + 0.008080502040684223, + 0.008571140468120575, + 0.0045615253038704395, + 0.005847746040672064, + 0.004483475349843502, + 0.00446972344070673, + 0.009930863976478577, + 0.008477515541017056, + 0.006245430558919907, + 0.014601477421820164, + 0.0038224782329052687, + 0.004879956133663654, + 0.0036333599127829075, + 0.0030530602671205997, + 0.010755358263850212, + 0.00726536987349391, + 0.004310877528041601, + 0.011823875829577446, + 0.002932672156020999, + 0.004178040660917759, + 0.002854833146557212, + 0.002601597225293517, + 0.00664517842233181, + 0.0036340332590043545, + 0.0025217225775122643, + 0.004569850862026215, + 0.0034734592773020267, + 0.003908894956111908, + 0.0037621904630213976, + 0.003221749560907483, + 0.006282403599470854, + 0.00421115942299366, + 0.003133978694677353, + 0.0038831373676657677, + 0.002141448901966214, + 0.0027414681389927864, + 0.0024372790940105915, + 0.0021520464215427637, + 0.0027348087169229984, + 0.0024009470362216234, + 0.0028306622989475727, + 0.0023957882076501846, + 0.0012835055822506547, + 0.002261082176119089, + 0.0016754830721765757, + 0.0011099484981968999, + 0.0019259945256635547, + 0.0020516677759587765, + 0.0015082326717674732, + 0.0031198386568576097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.06204873323440552 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, taken from a slightly elevated angle, focuses on an elderly craftsman with graying hair, dressed in a faded dark blue cotton work shirt. His head bowed, his forehead crisscrossed with deep wrinkles, his brows slightly furrowed as he gazes intently at his work. His shirt sleeves are neatly rolled up to his elbows, revealing sturdy forearms etched with veins and age spots. A warm overhead light slants down from the upper right, perfectly illuminating his hands and the gradually forming wood carving, while casting soft shadows on the slightly cluttered woodworking workshop in the background, which is filled with various tools. One of his calloused hands firmly holds a square block of unpolished, light-colored wood. With the other, he deftly and precisely manipulates a razor-edged carving knife. With steady, deliberate force, the tip of the knife slowly pushes across the wood, shaving off tiny, curled chips as thin as a cicada's wing. These tiny chips flutter briefly in the beam of light before silently landing on his dark workbench. The entire scene exudes a sense of profound tranquility and concentration. The warm tones of light and the air's almost permeating wood fragrance combine to create a tranquil atmosphere of frozen time and imbued with the spirit of craftsmanship.", + "seed": 14567, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_6/index_9062_seed_14567_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": -0.053322970494627954, + "vlm_logprob": -0.053322970494627954, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.053322970494627954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.041721973568201065 + ], + "opposite_logprob_tokens": [ + -3.291722059249878 + ], + "logprob_reward": -0.041721973568201065, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.041721973568201065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101491391658783, + 0.19370941817760468, + 0.18357476592063904, + 0.059321701526641846, + 0.08689489960670471, + 0.03754260763525963, + 0.021674012765288353, + 0.020129917189478874, + 0.04430752247571945, + 0.14269600808620453 + ], + "credit_map": [ + 0.02291552536189556, + 0.035399798303842545, + 0.02901040017604828, + 0.019389837980270386, + 0.04362419620156288, + 0.03171071410179138, + 0.015373344533145428, + 0.012725318782031536, + 0.021921100094914436, + 0.03731122612953186, + 0.023974521085619926, + 0.017921172082424164, + 0.037527747452259064, + 0.030472662299871445, + 0.013890701346099377, + 0.010690282098948956, + 0.017121748998761177, + 0.027440551668405533, + 0.03544831648468971, + 0.015038050711154938, + 0.02239716425538063, + 0.04207242652773857, + 0.015602483414113522, + 0.008454021997749805, + 0.007862822152674198, + 0.008930589072406292, + 0.007559644989669323, + 0.007582115475088358, + 0.009542903862893581, + 0.009154695086181164, + 0.004134726244956255, + 0.004554204177111387, + 0.01276906207203865, + 0.006689761765301228, + 0.011840497143566608, + 0.006647239439189434, + 0.021093549206852913, + 0.014621403068304062, + 0.004431045614182949, + 0.008802345022559166, + 0.0046977698802948, + 0.004802662879228592, + 0.005182804074138403, + 0.00244937464594841, + 0.00845140591263771, + 0.005743850488215685, + 0.0026898242067545652, + 0.003524915548041463, + 0.0036770268343389034, + 0.003584173507988453, + 0.0025606139097362757, + 0.001904246979393065, + 0.003963636700063944, + 0.0028193958569318056, + 0.001161343534477055, + 0.0020035752095282078, + 0.002000458538532257, + 0.0029307131189852953, + 0.002986739156767726, + 0.002730049891397357, + 0.00395140191540122, + 0.002579127438366413, + 0.0013502847868949175, + 0.0016011418774724007, + 0.006004604510962963, + 0.005096709821373224, + 0.007013130467385054, + 0.005961134098470211, + 0.007051341701298952, + 0.0050149522721767426, + 0.003435665275901556, + 0.00472998758777976, + 0.018101144582033157, + 0.02313731238245964, + 0.016943739727139473, + 0.02490491420030594, + 0.021548647433519363, + 0.01683083176612854, + 0.009104791097342968, + 0.012124619446694851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.041721973568201065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.020605897530913353 + ], + "opposite_logprob_tokens": [ + -3.8956058025360107 + ], + "logprob_reward": -0.020605897530913353, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020605897530913353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12231692671775818, + 0.10290735960006714, + 0.07930036634206772, + 0.07062920182943344, + 0.0769493356347084, + 0.07518436759710312, + 0.03510555624961853, + 0.06583211570978165, + 0.10100992769002914, + 0.27076485753059387 + ], + "credit_map": [ + 0.01977424882352352, + 0.013146218843758106, + 0.01831553317606449, + 0.014028600417077541, + 0.013562323525547981, + 0.017083002254366875, + 0.01589759811758995, + 0.010509401559829712, + 0.01714085228741169, + 0.011569795198738575, + 0.02612544223666191, + 0.008850552141666412, + 0.008906351402401924, + 0.01656508259475231, + 0.007815508171916008, + 0.005933766718953848, + 0.009244637563824654, + 0.014790165238082409, + 0.009053719229996204, + 0.009050470776855946, + 0.009436830878257751, + 0.013929304666817188, + 0.009129276499152184, + 0.004665955901145935, + 0.008491674438118935, + 0.01062753051519394, + 0.013556386344134808, + 0.009414893575012684, + 0.005851299036294222, + 0.010736629366874695, + 0.005310737527906895, + 0.0066400510258972645, + 0.011802684515714645, + 0.0065110777504742146, + 0.011122692376375198, + 0.007404280360788107, + 0.014685509726405144, + 0.011268600821495056, + 0.005665892735123634, + 0.008488593623042107, + 0.01249274518340826, + 0.010987295769155025, + 0.009811989031732082, + 0.007104135118424892, + 0.013231632299721241, + 0.01055152527987957, + 0.004697203636169434, + 0.00630784360691905, + 0.0033961727749556303, + 0.0062283738516271114, + 0.003646803554147482, + 0.003219981910660863, + 0.007896039634943008, + 0.004691277630627155, + 0.00392600242048502, + 0.0021009044721722603, + 0.00749423960223794, + 0.012614598497748375, + 0.007542911916971207, + 0.005537998862564564, + 0.010719308629631996, + 0.008640595711767673, + 0.005157432984560728, + 0.00812503695487976, + 0.01180283073335886, + 0.01198258250951767, + 0.012866687960922718, + 0.014878991059958935, + 0.016939381137490273, + 0.0110152093693614, + 0.007895738817751408, + 0.013628506101667881, + 0.03605697676539421, + 0.049176108092069626, + 0.03184068202972412, + 0.039621345698833466, + 0.045379091054201126, + 0.030478889122605324, + 0.016920514404773712, + 0.021291231736540794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.020605897530913353 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.0205991230905056 + ], + "opposite_logprob_tokens": [ + -3.895599126815796 + ], + "logprob_reward": -0.0205991230905056, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.0205991230905056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12718360126018524, + 0.09941904991865158, + 0.08820077031850815, + 0.08629706501960754, + 0.06419194489717484, + 0.09594512730836868, + 0.03218260034918785, + 0.07290655374526978, + 0.1786123812198639, + 0.15506096184253693 + ], + "credit_map": [ + 0.0218201931566, + 0.012983250431716442, + 0.019878610968589783, + 0.012945294380187988, + 0.014816999435424805, + 0.018852686509490013, + 0.01465565711259842, + 0.011230896227061749, + 0.011758655309677124, + 0.012886318378150463, + 0.020459137856960297, + 0.012218517251312733, + 0.009136334992945194, + 0.015429669991135597, + 0.010545494966208935, + 0.006984916515648365, + 0.012640228495001793, + 0.010607976466417313, + 0.0178680382668972, + 0.00949313584715128, + 0.008402463980019093, + 0.017474595457315445, + 0.006481911987066269, + 0.005232424009591341, + 0.011440026573836803, + 0.011900126934051514, + 0.01009289175271988, + 0.008277899585664272, + 0.014694459736347198, + 0.013495534658432007, + 0.011003440245985985, + 0.005392682738602161, + 0.007118452340364456, + 0.006234309170395136, + 0.007662155199795961, + 0.008374882861971855, + 0.012249910272657871, + 0.008859834633767605, + 0.007698932662606239, + 0.0059934682212769985, + 0.010180099867284298, + 0.008887609466910362, + 0.014638916589319706, + 0.01207161694765091, + 0.02002287097275257, + 0.015508540906012058, + 0.008267877623438835, + 0.0063675944693386555, + 0.003101249458268285, + 0.005545288790017366, + 0.002848678734153509, + 0.004547849763184786, + 0.005440970417112112, + 0.0038879758212715387, + 0.002999568823724985, + 0.0038110176101326942, + 0.0068708257749676704, + 0.011027704924345016, + 0.009068217128515244, + 0.011980227194726467, + 0.00967380777001381, + 0.008535455912351608, + 0.010196633636951447, + 0.005553686060011387, + 0.0181984081864357, + 0.02344805933535099, + 0.02009637840092182, + 0.03318406268954277, + 0.027266351506114006, + 0.018746644258499146, + 0.01761174574494362, + 0.020060742273926735, + 0.02036430686712265, + 0.028196178376674652, + 0.01839436963200569, + 0.02402198687195778, + 0.0243263840675354, + 0.01773405633866787, + 0.009880300611257553, + 0.012143385596573353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0205991230905056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.023305542767047882 + ], + "opposite_logprob_tokens": [ + -3.773305654525757 + ], + "logprob_reward": -0.023305542767047882, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305542767047882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19522625207901, + 0.23474089801311493, + 0.19304552674293518, + 0.05193192511796951, + 0.030189428478479385, + 0.02649320848286152, + 0.05199089273810387, + 0.025745026767253876, + 0.03404591605067253, + 0.15659090876579285 + ], + "credit_map": [ + 0.030623627826571465, + 0.024861695244908333, + 0.028318533673882484, + 0.015347917564213276, + 0.027283594012260437, + 0.031429070979356766, + 0.01873837411403656, + 0.018623463809490204, + 0.02742588520050049, + 0.05499459058046341, + 0.022477196529507637, + 0.026515129953622818, + 0.030792193487286568, + 0.03493581339716911, + 0.01721431128680706, + 0.020385783165693283, + 0.025104664266109467, + 0.03269138187170029, + 0.028107693418860435, + 0.025994306430220604, + 0.023150211200118065, + 0.02967851236462593, + 0.015863539651036263, + 0.012455224990844727, + 0.005081922747194767, + 0.004325158428400755, + 0.005695679225027561, + 0.005840596742928028, + 0.006844445131719112, + 0.008085651323199272, + 0.009831818751990795, + 0.006226660218089819, + 0.003854756709188223, + 0.002827669959515333, + 0.0044117034412920475, + 0.003628608537837863, + 0.0038332364056259394, + 0.004649049136787653, + 0.002701343270018697, + 0.004283062648028135, + 0.003892940003424883, + 0.0026578085962682962, + 0.0036690831184387207, + 0.0028178596403449774, + 0.0032230254728347063, + 0.003235785523429513, + 0.0028853039257228374, + 0.004111406859010458, + 0.0049188449047505856, + 0.010863536968827248, + 0.0048102280125021935, + 0.005199894774705172, + 0.011697186157107353, + 0.0073994919657707214, + 0.0038997726514935493, + 0.0032019433565437794, + 0.0033889126498252153, + 0.004026737529784441, + 0.003447451861575246, + 0.002106998348608613, + 0.004620825406163931, + 0.0032959128729999065, + 0.002124341670423746, + 0.0027338487561792135, + 0.004110227804630995, + 0.006858520209789276, + 0.004467623308300972, + 0.0031115098390728235, + 0.006719451397657394, + 0.004159088712185621, + 0.0018768274458125234, + 0.0027426686137914658, + 0.019826781004667282, + 0.025961436331272125, + 0.020964419469237328, + 0.023963598534464836, + 0.023920871317386627, + 0.018371371552348137, + 0.009600981138646603, + 0.01398145779967308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023305542767047882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 craftsman with graying hair and deep wrinkles on his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the craftsman wearing a faded dark blue cotton work shirt with sleeves rolled up to the elbows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the craftsman using a razor-edged carving knife to shave tiny curled chips off a block of light-colored wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a warm overhead light coming from the upper right, casting soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a cluttered woodworking workshop with various tools visible 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": [ + -0.16038231551647186 + ], + "opposite_logprob_tokens": [ + -1.9103822708129883 + ], + "logprob_reward": -0.16038231551647186, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038231551647186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09960281103849411, + 0.3617219626903534, + 0.17354069650173187, + 0.03884921595454216, + 0.041580092161893845, + 0.10874681919813156, + 0.025511987507343292, + 0.0472286231815815, + 0.03831131383776665, + 0.06490658223628998 + ], + "credit_map": [ + 0.013399592600762844, + 0.014986280351877213, + 0.012567685917019844, + 0.009198357351124287, + 0.015094701200723648, + 0.014687499031424522, + 0.008612271398305893, + 0.011056416667997837, + 0.04334401339292526, + 0.05692087486386299, + 0.0720343291759491, + 0.026464467868208885, + 0.05806507542729378, + 0.05823105573654175, + 0.02815745398402214, + 0.01850464753806591, + 0.020673109218478203, + 0.026921089738607407, + 0.03361581638455391, + 0.013303051702678204, + 0.021462909877300262, + 0.037172313779592514, + 0.011042676866054535, + 0.009349705651402473, + 0.0042263902723789215, + 0.005568124819546938, + 0.0043211448937654495, + 0.004915399011224508, + 0.005537990480661392, + 0.005441032350063324, + 0.005474030505865812, + 0.003365102456882596, + 0.006144228391349316, + 0.0034568640403449535, + 0.005613506305962801, + 0.003326441626995802, + 0.007315733935683966, + 0.006805749610066414, + 0.0032339487224817276, + 0.00568361533805728, + 0.017530659213662148, + 0.014284037984907627, + 0.01814313232898712, + 0.010881712660193443, + 0.01494547724723816, + 0.01507807057350874, + 0.008194446563720703, + 0.009689278900623322, + 0.0018378611421212554, + 0.0038395579904317856, + 0.00196973979473114, + 0.0038589611649513245, + 0.005294403526932001, + 0.003123380709439516, + 0.003319842042401433, + 0.002268239390105009, + 0.00629185326397419, + 0.006775387097150087, + 0.005144323222339153, + 0.00718563050031662, + 0.008466750383377075, + 0.006329639814794064, + 0.00267515005543828, + 0.0043598804622888565, + 0.0037849145010113716, + 0.0074177877977490425, + 0.00504394480958581, + 0.0036567740608006716, + 0.008165525272488594, + 0.0045428164303302765, + 0.0025786624755710363, + 0.003120885929092765, + 0.0077104526571929455, + 0.010555725544691086, + 0.008775471709668636, + 0.00931969191879034, + 0.009892246685922146, + 0.007873082533478737, + 0.004145000129938126, + 0.006634906865656376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.16038231551647186 + } + ], + "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/0000055/group_7/index_tensor([6874, 6874])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_tensor([6874, 6874])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..700bcc26393ebc19c856d6e47f1a5cbca9370e03 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_tensor([6874, 6874])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, low-angle tracking shot follows a chariot, its wheels tumbling along a bumpy ancient road, creating a subtle yet authentic jolt of movement. The magnificent chariot, crafted from dark lacquered wood, its body intricately carved with cloud patterns and relief carvings of mythical beasts, gleams with a deep patina etched by time, its bronze components reflecting a cold, hard gleam against the gloomy sky. Within the carriage, a stern-faced, middle-aged emperor sits poised, clad in a dragon robe of deep black silk. A five-clawed golden dragon, delicately embroidered with gold thread, appears faintly visible in the dim light, its scales seemingly undulating. The emperor's brow furrows, his lips pressed into a tight line. His deep gaze gazes into the void ahead, revealing a complex blend of worry and determination. His hands clasp the jade belt at his knees, his knuckles white with the strain. A chariot travels along an ancient loess road deeply carved by ruts. Its massive wooden wheels roll over the dry ground, stirring up clouds of fine dust. The sky is covered by thick, leaden clouds, and the diffuse, oppressive light casts a cold, desaturated hue throughout the painting. Only the dark gold threads of the dragon robes provide the sole warm accent, creating a tense and austere atmosphere of impending storm.", + "seed": 29258, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29258_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": -0.034998444654047486, + "vlm_logprob": -0.034998444654047486, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.034998444654047486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04036088287830353 + ], + "opposite_logprob_tokens": [ + -3.415360927581787 + ], + "logprob_reward": -0.04036088287830353, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04036088287830353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14414562284946442, + 0.08583172410726547, + 0.14994701743125916, + 0.046070560812950134, + 0.03240486606955528, + 0.07076069712638855, + 0.03898432478308678, + 0.24298420548439026, + 0.101959228515625, + 0.08691174536943436 + ], + "credit_map": [ + 0.02159225195646286, + 0.013327889144420624, + 0.03200143575668335, + 0.0159091018140316, + 0.013991925865411758, + 0.018566859886050224, + 0.010210195556282997, + 0.018545979633927345, + 0.009409814141690731, + 0.007735267747193575, + 0.017844069749116898, + 0.012419520877301693, + 0.010954344645142555, + 0.008698183111846447, + 0.00843826774507761, + 0.010332264006137848, + 0.02499905787408352, + 0.014872722327709198, + 0.027234502136707306, + 0.01324552670121193, + 0.016562048345804214, + 0.0195115115493536, + 0.01018431130796671, + 0.0233373511582613, + 0.004763155709952116, + 0.007759831380099058, + 0.0037058622110635042, + 0.004299732390791178, + 0.007126398850232363, + 0.00580583093687892, + 0.006054040044546127, + 0.006555712781846523, + 0.0026078135706484318, + 0.0030125603079795837, + 0.004803034942597151, + 0.00621076300740242, + 0.0032017030753195286, + 0.003054468659684062, + 0.006197442766278982, + 0.003317087423056364, + 0.008257823996245861, + 0.014433331787586212, + 0.008611193858087063, + 0.007771998643875122, + 0.0077339401468634605, + 0.008097324520349503, + 0.00937238335609436, + 0.006482711527496576, + 0.0038328003138303757, + 0.009874521754682064, + 0.0029290146194398403, + 0.0032723168842494488, + 0.004452108405530453, + 0.005714572500437498, + 0.005153718404471874, + 0.0037552756257355213, + 0.0640830397605896, + 0.014209010638296604, + 0.010750028304755688, + 0.03338359296321869, + 0.009864448569715023, + 0.014918862842023373, + 0.017085708677768707, + 0.07868953049182892, + 0.010452983900904655, + 0.010036586783826351, + 0.01318406406790018, + 0.019412405788898468, + 0.008815893903374672, + 0.008269819431006908, + 0.021754300221800804, + 0.01003318466246128, + 0.015888627618551254, + 0.030705010518431664, + 0.004135260358452797, + 0.0070121497847139835, + 0.007597226649522781, + 0.004782229661941528, + 0.005622340366244316, + 0.011168910190463066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.04036088287830353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812243044376373 + ], + "opposite_logprob_tokens": [ + -3.2881224155426025 + ], + "logprob_reward": -0.03812243044376373, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812243044376373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10206729918718338, + 0.02972356602549553, + 0.10149829089641571, + 0.030230175703763962, + 0.026304930448532104, + 0.029030660167336464, + 0.04279370605945587, + 0.25031211972236633, + 0.2583859860897064, + 0.12965326011180878 + ], + "credit_map": [ + 0.014293729327619076, + 0.007193647790700197, + 0.027492115274071693, + 0.008734063245356083, + 0.007646644022315741, + 0.013789892196655273, + 0.007200040854513645, + 0.01571715995669365, + 0.0030708599369972944, + 0.002817657543346286, + 0.007270469795912504, + 0.0032292066607624292, + 0.0020241031888872385, + 0.00353857665322721, + 0.002134398790076375, + 0.005638292990624905, + 0.016488956287503242, + 0.011248504742980003, + 0.017426177859306335, + 0.009306900203227997, + 0.010725644417107105, + 0.012893445789813995, + 0.008394448086619377, + 0.015014221891760826, + 0.0026448394637554884, + 0.005680704489350319, + 0.003162994049489498, + 0.0036612770054489374, + 0.0036272192373871803, + 0.003357126610353589, + 0.005078911781311035, + 0.003017102601006627, + 0.002010904485359788, + 0.004267322365194559, + 0.002297688275575638, + 0.003984103910624981, + 0.0032043964602053165, + 0.0024943228345364332, + 0.0036231507547199726, + 0.004423042293637991, + 0.002384297316893935, + 0.004497867077589035, + 0.002446903847157955, + 0.0038943977560847998, + 0.005538491997867823, + 0.003932660911232233, + 0.003864436177536845, + 0.002471605548635125, + 0.0057987915351986885, + 0.004872757010161877, + 0.005096116568893194, + 0.004121690057218075, + 0.00548037514090538, + 0.005446608178317547, + 0.007263504434376955, + 0.004713861737400293, + 0.05205896496772766, + 0.02115733176469803, + 0.015747860074043274, + 0.028624150902032852, + 0.024550113826990128, + 0.029862169176340103, + 0.017446281388401985, + 0.060865238308906555, + 0.023347826674580574, + 0.022577958181500435, + 0.02859180048108101, + 0.03970770537853241, + 0.02535293996334076, + 0.021851619705557823, + 0.03346913307905197, + 0.06348702311515808, + 0.01786559633910656, + 0.01869744248688221, + 0.007860071025788784, + 0.012085412628948689, + 0.026385294273495674, + 0.006810297258198261, + 0.010968203656375408, + 0.028980934992432594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.03812243044376373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646263778209686 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646263778209686, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646263778209686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12826862931251526, + 0.06814676523208618, + 0.1013977974653244, + 0.06165146455168724, + 0.11871541291475296, + 0.07571873813867569, + 0.021429045125842094, + 0.1261882483959198, + 0.11507237702608109, + 0.1834115833044052 + ], + "credit_map": [ + 0.009190930984914303, + 0.014259436167776585, + 0.02254413440823555, + 0.013593261130154133, + 0.015288560651242733, + 0.01613842509686947, + 0.01285743061453104, + 0.024396447464823723, + 0.006959892343729734, + 0.005935697816312313, + 0.013685851357877254, + 0.008850207552313805, + 0.008126417174935341, + 0.0073853787034749985, + 0.007459304295480251, + 0.009744014590978622, + 0.013485505245625973, + 0.017570456489920616, + 0.01282233651727438, + 0.016814565286040306, + 0.010618321597576141, + 0.01169996615499258, + 0.008786511607468128, + 0.009600131772458553, + 0.004676579497754574, + 0.013974527828395367, + 0.006125477142632008, + 0.00567229837179184, + 0.007365458644926548, + 0.006761983968317509, + 0.011595596559345722, + 0.0054795462638139725, + 0.013106988742947578, + 0.021385839208960533, + 0.011159474030137062, + 0.012516124173998833, + 0.020271282643079758, + 0.01383493933826685, + 0.0133181968703866, + 0.013122564181685448, + 0.007068642880767584, + 0.017651280388236046, + 0.008838718757033348, + 0.007817486301064491, + 0.009006463922560215, + 0.009454423561692238, + 0.008487770333886147, + 0.007393950130790472, + 0.002951717469841242, + 0.003967844415456057, + 0.0019049415132030845, + 0.0017874941695481539, + 0.0032073899637907743, + 0.002660742960870266, + 0.002611849457025528, + 0.0023370638955384493, + 0.0224072877317667, + 0.011475356295704842, + 0.007876642979681492, + 0.016532180830836296, + 0.00994651298969984, + 0.014239446260035038, + 0.00887385755777359, + 0.03483695164322853, + 0.013313437812030315, + 0.00795195996761322, + 0.009336552582681179, + 0.019689034670591354, + 0.012456012889742851, + 0.006884606555104256, + 0.02993970736861229, + 0.01550106331706047, + 0.04073135182261467, + 0.07000339776277542, + 0.0031929095275700092, + 0.011247429996728897, + 0.017620282247662544, + 0.007962221279740334, + 0.01081570703536272, + 0.021838298067450523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.048646263778209686 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777394123375416 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777394123375416, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777394123375416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0819382444024086, + 0.2138376384973526, + 0.19521795213222504, + 0.033871643245220184, + 0.05292164534330368, + 0.11255593597888947, + 0.05745169147849083, + 0.1003788411617279, + 0.07130195200443268, + 0.08052439987659454 + ], + "credit_map": [ + 0.015187056735157967, + 0.006701043341308832, + 0.016487881541252136, + 0.007986405864357948, + 0.006229369435459375, + 0.010270749218761921, + 0.008245743811130524, + 0.010830005630850792, + 0.03199676796793938, + 0.026962492614984512, + 0.032992519438266754, + 0.03477625921368599, + 0.014116930775344372, + 0.02417307160794735, + 0.016918689012527466, + 0.031900931149721146, + 0.03356405720114708, + 0.018415793776512146, + 0.03474799171090126, + 0.016589755192399025, + 0.022544246166944504, + 0.025919854640960693, + 0.013611935079097748, + 0.02982434257864952, + 0.0026501361280679703, + 0.0042086318135261536, + 0.004553717095404863, + 0.004424453712999821, + 0.002906263340264559, + 0.0034025509376078844, + 0.007386817596852779, + 0.0043390775099396706, + 0.009933268651366234, + 0.00698811374604702, + 0.006078042089939117, + 0.006392078474164009, + 0.00566437654197216, + 0.006843158043920994, + 0.0057952506467700005, + 0.005227359477430582, + 0.019879832863807678, + 0.02076215110719204, + 0.010599757544696331, + 0.008672809228301048, + 0.015508076176047325, + 0.016890598461031914, + 0.01092485897243023, + 0.009317855350673199, + 0.009652621112763882, + 0.009820198640227318, + 0.004797960631549358, + 0.00466320663690567, + 0.008052147924900055, + 0.0061251879669725895, + 0.0063921669498085976, + 0.007948214188218117, + 0.023837827146053314, + 0.006148149725049734, + 0.005195342469960451, + 0.014271366409957409, + 0.005716685671359301, + 0.0065954336896538734, + 0.007693167310208082, + 0.030920878052711487, + 0.006289568264037371, + 0.006944486405700445, + 0.004974762909114361, + 0.008040184155106544, + 0.01943657360970974, + 0.005987926386296749, + 0.00769351189956069, + 0.011934945359826088, + 0.014198757708072662, + 0.01102708000689745, + 0.003594833891838789, + 0.00643688952550292, + 0.02209516614675522, + 0.005113129038363695, + 0.006908675655722618, + 0.011149878613650799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.009777394123375416 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808525204658508 + ], + "opposite_logprob_tokens": [ + -3.2880852222442627 + ], + "logprob_reward": -0.03808525204658508, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808525204658508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09037873893976212, + 0.021534467115998268, + 0.06709080934524536, + 0.10342710465192795, + 0.048764605075120926, + 0.028989730402827263, + 0.02737707644701004, + 0.5119916200637817, + 0.055211909115314484, + 0.045233920216560364 + ], + "credit_map": [ + 0.008462249301373959, + 0.0062057822942733765, + 0.019236046820878983, + 0.010117704048752785, + 0.009126820601522923, + 0.010664020664989948, + 0.007577889133244753, + 0.018988223746418953, + 0.002073065610602498, + 0.002574716228991747, + 0.003438018960878253, + 0.003389573423191905, + 0.003004223806783557, + 0.0022225333377718925, + 0.0019844321068376303, + 0.00284790457226336, + 0.010443181730806828, + 0.007496749982237816, + 0.011624178849160671, + 0.0066282181069254875, + 0.006388024426996708, + 0.008429798297584057, + 0.005556835792958736, + 0.010523820295929909, + 0.009513000957667828, + 0.015556592494249344, + 0.012333856895565987, + 0.013375922106206417, + 0.010220890864729881, + 0.01107422448694706, + 0.01846589893102646, + 0.012886718846857548, + 0.0035813332069665194, + 0.005020117852836847, + 0.005166849587112665, + 0.007079176604747772, + 0.010172302834689617, + 0.00677771819755435, + 0.006265934556722641, + 0.004701173398643732, + 0.0023539692629128695, + 0.004696539603173733, + 0.003647375153377652, + 0.002445430727675557, + 0.004129824694246054, + 0.0033913301303982735, + 0.00439988961443305, + 0.0039253695867955685, + 0.0032046083360910416, + 0.004310096148401499, + 0.002059318358078599, + 0.0019909769762307405, + 0.006639573257416487, + 0.004079338163137436, + 0.0025987981352955103, + 0.0024943677708506584, + 0.1314629316329956, + 0.030265571549534798, + 0.025506360456347466, + 0.07451657950878143, + 0.0195060633122921, + 0.02335461787879467, + 0.03797300159931183, + 0.1694064736366272, + 0.005483731161803007, + 0.0047887591645121574, + 0.006462377030402422, + 0.008775263093411922, + 0.005316044669598341, + 0.007431738544255495, + 0.009521248750388622, + 0.007432745303958654, + 0.009463703259825706, + 0.0053619458340108395, + 0.0022782613523304462, + 0.003765889909118414, + 0.0037982785142958164, + 0.002645724220201373, + 0.005863100755959749, + 0.012057015672326088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.03808525204658508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle tracking shot follows a chariot, its wheels tumbling along a bumpy ancient road, creating a subtle yet authentic jolt of movement. The magnificent chariot, crafted from dark lacquered wood, its body intricately carved with cloud patterns and relief carvings of mythical beasts, gleams with a deep patina etched by time, its bronze components reflecting a cold, hard gleam against the gloomy sky. Within the carriage, a stern-faced, middle-aged emperor sits poised, clad in a dragon robe of deep black silk. A five-clawed golden dragon, delicately embroidered with gold thread, appears faintly visible in the dim light, its scales seemingly undulating. The emperor's brow furrows, his lips pressed into a tight line. His deep gaze gazes into the void ahead, revealing a complex blend of worry and determination. His hands clasp the jade belt at his knees, his knuckles white with the strain. A chariot travels along an ancient loess road deeply carved by ruts. Its massive wooden wheels roll over the dry ground, stirring up clouds of fine dust. The sky is covered by thick, leaden clouds, and the diffuse, oppressive light casts a cold, desaturated hue throughout the painting. Only the dark gold threads of the dragon robes provide the sole warm accent, creating a tense and austere atmosphere of impending storm.", + "seed": 29259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29259_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": -0.16805061139166355, + "vlm_logprob": -0.16805061139166355, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.16805061139166355, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6964678168296814 + ], + "opposite_logprob_tokens": [ + -0.6964678168296814 + ], + "logprob_reward": -0.6964678168296814, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6964678168296814, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.30202293395996094, + 0.2161361128091812, + 0.09885554015636444, + 0.0704597681760788, + 0.04727170616388321, + 0.07613467425107956, + 0.03363395854830742, + 0.06939616054296494, + 0.03100258857011795, + 0.05508657544851303 + ], + "credit_map": [ + 0.044119659811258316, + 0.0349886529147625, + 0.033943019807338715, + 0.049356233328580856, + 0.02836376056075096, + 0.02969692088663578, + 0.03859230503439903, + 0.04296236112713814, + 0.026241116225719452, + 0.025842200964689255, + 0.02578064054250717, + 0.04648863151669502, + 0.021568231284618378, + 0.024724451825022697, + 0.025193145498633385, + 0.020297693088650703, + 0.015884796157479286, + 0.010783223435282707, + 0.015832945704460144, + 0.02160385064780712, + 0.00740937190130353, + 0.006546179763972759, + 0.010531246662139893, + 0.010263919830322266, + 0.012254372239112854, + 0.008304770104587078, + 0.009553260169923306, + 0.008993609808385372, + 0.007450913079082966, + 0.00454275356605649, + 0.005826003383845091, + 0.013534088619053364, + 0.004331246484071016, + 0.004365609958767891, + 0.004152171779423952, + 0.01108283456414938, + 0.004911339376121759, + 0.004192500840872526, + 0.004025363828986883, + 0.010210639797151089, + 0.007675316650420427, + 0.007625325582921505, + 0.011085491627454758, + 0.010726863518357277, + 0.015013902448117733, + 0.005478404927998781, + 0.01049733255058527, + 0.008032038807868958, + 0.003933211788535118, + 0.002620631828904152, + 0.005283844191581011, + 0.005250742193311453, + 0.004150742199271917, + 0.003180483356118202, + 0.0042341710068285465, + 0.004980132915079594, + 0.007827774621546268, + 0.005102986469864845, + 0.00835023820400238, + 0.017067372798919678, + 0.009089424274861813, + 0.008473830297589302, + 0.006679713260382414, + 0.0068048229441046715, + 0.002201734809204936, + 0.0024641347117722034, + 0.0028230745811015368, + 0.0030566288623958826, + 0.0052152774296700954, + 0.0067777070216834545, + 0.0034155691973865032, + 0.005048463121056557, + 0.0038162083365023136, + 0.0046324292197823524, + 0.004187043756246567, + 0.008564779534935951, + 0.01073980238288641, + 0.01127526443451643, + 0.006122561637312174, + 0.005748485215008259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.6964678168296814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373015716671944 + ], + "opposite_logprob_tokens": [ + -3.6513729095458984 + ], + "logprob_reward": -0.026373015716671944, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373015716671944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11696702986955643, + 0.18389944732189178, + 0.08550567924976349, + 0.07312316447496414, + 0.09154143929481506, + 0.06554882228374481, + 0.062439002096652985, + 0.07684962451457977, + 0.04883792996406555, + 0.19528786838054657 + ], + "credit_map": [ + 0.015585449524223804, + 0.014806232415139675, + 0.014972559176385403, + 0.021622946485877037, + 0.011582856997847557, + 0.0113960150629282, + 0.015940556302666664, + 0.011060409247875214, + 0.019548747688531876, + 0.01980539783835411, + 0.023072712123394012, + 0.04213276505470276, + 0.01977487839758396, + 0.019403772428631783, + 0.022978056222200394, + 0.017183110117912292, + 0.01793881319463253, + 0.011047868989408016, + 0.009594482369720936, + 0.016195228323340416, + 0.009216533042490482, + 0.005385161843150854, + 0.008648689836263657, + 0.007478902116417885, + 0.013567773625254631, + 0.00504120159894228, + 0.008605276234447956, + 0.00935856718569994, + 0.0073287105187773705, + 0.003496340708807111, + 0.006323814392089844, + 0.01940147951245308, + 0.009512464515864849, + 0.005144990980625153, + 0.01209181733429432, + 0.02102779969573021, + 0.005876278970390558, + 0.004120121710002422, + 0.006522456184029579, + 0.027245506644248962, + 0.007473058998584747, + 0.003946271724998951, + 0.00524565763771534, + 0.009635066613554955, + 0.014689357951283455, + 0.004931323695927858, + 0.009003764018416405, + 0.010624317452311516, + 0.00583706796169281, + 0.006235096137970686, + 0.010947133414447308, + 0.010331494733691216, + 0.00881392601877451, + 0.005575411021709442, + 0.00837317481637001, + 0.006325698457658291, + 0.006424295250326395, + 0.007678699214011431, + 0.009354481473565102, + 0.014678039588034153, + 0.013799666427075863, + 0.009379620663821697, + 0.009139583446085453, + 0.006395238917320967, + 0.004165316466242075, + 0.0036638067103922367, + 0.004549061879515648, + 0.005587932653725147, + 0.008576218038797379, + 0.009156120009720325, + 0.005183268338441849, + 0.007956207729876041, + 0.014101197011768818, + 0.018837766721844673, + 0.011319315060973167, + 0.07633967697620392, + 0.01366460882127285, + 0.031394388526678085, + 0.010890643112361431, + 0.018740274012088776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.026373015716671944 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864751175045967 + ], + "opposite_logprob_tokens": [ + -3.048647403717041 + ], + "logprob_reward": -0.04864751175045967, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864751175045967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10872194170951843, + 0.09398043900728226, + 0.04100935906171799, + 0.14900419116020203, + 0.06232397258281708, + 0.12475813180208206, + 0.05353890359401703, + 0.055695101618766785, + 0.05160099267959595, + 0.259366899728775 + ], + "credit_map": [ + 0.020352859050035477, + 0.01094209123402834, + 0.01784353516995907, + 0.013927314430475235, + 0.010918819345533848, + 0.011717786081135273, + 0.01222778856754303, + 0.010791752487421036, + 0.010122017003595829, + 0.010794240050017834, + 0.011048536747694016, + 0.021704303100705147, + 0.010190983302891254, + 0.010892028920352459, + 0.011111998930573463, + 0.00811633188277483, + 0.0070619587786495686, + 0.006880661007016897, + 0.0051209283992648125, + 0.0066042449325323105, + 0.004181342199444771, + 0.0032960509415715933, + 0.0040731849148869514, + 0.0037909909151494503, + 0.03091757372021675, + 0.012640717439353466, + 0.011899645440280437, + 0.017397385090589523, + 0.02362065762281418, + 0.009150725789368153, + 0.012612087652087212, + 0.030765412375330925, + 0.006016390863806009, + 0.003717780811712146, + 0.006041525397449732, + 0.00918593443930149, + 0.010665777139365673, + 0.004484183620661497, + 0.008242784067988396, + 0.013969595544040203, + 0.013305097818374634, + 0.009128532372415066, + 0.0147083830088377, + 0.01767481118440628, + 0.027003547176718712, + 0.009155627340078354, + 0.018026523292064667, + 0.015755606815218925, + 0.006261081900447607, + 0.005340236704796553, + 0.011512702330946922, + 0.009347023442387581, + 0.0047529530711472034, + 0.005187839735299349, + 0.006176248658448458, + 0.004960817750543356, + 0.00600062170997262, + 0.0043483455665409565, + 0.005504599306732416, + 0.016829481348395348, + 0.007546828594058752, + 0.005238607991486788, + 0.005101121962070465, + 0.005125494208186865, + 0.0056591336615383625, + 0.004637600854039192, + 0.005998741835355759, + 0.004883626941591501, + 0.008182796649634838, + 0.005815043579787016, + 0.006082776468247175, + 0.010341276414692402, + 0.01811346784234047, + 0.024434037506580353, + 0.015268788672983646, + 0.09556867927312851, + 0.023050108924508095, + 0.04005242884159088, + 0.01751670427620411, + 0.025362683460116386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.04864751175045967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731215864419937 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731215864419937, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731215864419937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15580357611179352, + 0.12838593125343323, + 0.2247426062822342, + 0.10805349051952362, + 0.0764433741569519, + 0.02630828134715557, + 0.033014144748449326, + 0.0782419964671135, + 0.06287260353565216, + 0.10613401979207993 + ], + "credit_map": [ + 0.02925870567560196, + 0.013072606176137924, + 0.023051582276821136, + 0.02551952935755253, + 0.011677393689751625, + 0.011283055879175663, + 0.025876760482788086, + 0.016063949093222618, + 0.017717765644192696, + 0.020562762394547462, + 0.016696827486157417, + 0.021385151892900467, + 0.011673377826809883, + 0.01348816603422165, + 0.015183351002633572, + 0.01167852059006691, + 0.0709390938282013, + 0.019820529967546463, + 0.03432976081967354, + 0.03338421881198883, + 0.018493831157684326, + 0.01343883853405714, + 0.01750737428665161, + 0.016828957945108414, + 0.01963246613740921, + 0.00982324592769146, + 0.011271010152995586, + 0.011422467418015003, + 0.012645740993320942, + 0.008053163066506386, + 0.008491394110023975, + 0.026714013889431953, + 0.008386355824768543, + 0.0043202294036746025, + 0.0063187177293002605, + 0.016756458207964897, + 0.006654893979430199, + 0.0037313171196728945, + 0.007120353635400534, + 0.023155050352215767, + 0.002729913918301463, + 0.001963629387319088, + 0.004091763868927956, + 0.0038666848558932543, + 0.005503897555172443, + 0.002040361287072301, + 0.0035127741284668446, + 0.002599256346002221, + 0.003472095588222146, + 0.0034290971234440804, + 0.0053365109488368034, + 0.004549193661659956, + 0.004524290096014738, + 0.0028206086717545986, + 0.004326703026890755, + 0.004555647261440754, + 0.005517362151294947, + 0.006095474120229483, + 0.01126447319984436, + 0.01324611809104681, + 0.012916788458824158, + 0.010600535199046135, + 0.008140704594552517, + 0.010460535995662212, + 0.005194392055273056, + 0.004624569322913885, + 0.006635430734604597, + 0.007771143689751625, + 0.01158901397138834, + 0.01032734289765358, + 0.0061653051525354385, + 0.010565404780209064, + 0.0049830107018351555, + 0.008208469487726688, + 0.00903939176350832, + 0.02990889549255371, + 0.012152066454291344, + 0.023084213957190514, + 0.00924148689955473, + 0.009516486898064613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006731215864419937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203349679708481 + ], + "opposite_logprob_tokens": [ + -2.8120334148406982 + ], + "logprob_reward": -0.06203349679708481, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203349679708481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1762147843837738, + 0.1012919619679451, + 0.03606279566884041, + 0.12672722339630127, + 0.2855222821235657, + 0.07344767451286316, + 0.021887578070163727, + 0.074295274913311, + 0.04426835477352142, + 0.060282059013843536 + ], + "credit_map": [ + 0.02626362442970276, + 0.02318420261144638, + 0.02077246457338333, + 0.02987297810614109, + 0.019090447574853897, + 0.015662087127566338, + 0.020262328907847404, + 0.021106671541929245, + 0.010976514779031277, + 0.011618085205554962, + 0.012243756093084812, + 0.026720590889453888, + 0.01167314127087593, + 0.00963631272315979, + 0.010659349150955677, + 0.007764220703393221, + 0.008032187819480896, + 0.0031280778348445892, + 0.006171433720737696, + 0.005306166596710682, + 0.0035930899903178215, + 0.0030482665169984102, + 0.0037593168672174215, + 0.0030242588836699724, + 0.027298420667648315, + 0.008217638358473778, + 0.018634095788002014, + 0.014434834942221642, + 0.008876495063304901, + 0.007091550622135401, + 0.00734786968678236, + 0.034826312214136124, + 0.03544143959879875, + 0.014032791368663311, + 0.04375878721475601, + 0.05775171518325806, + 0.01658892072737217, + 0.013724187389016151, + 0.0225350521504879, + 0.08168939501047134, + 0.007732398342341185, + 0.00787599477916956, + 0.011711164377629757, + 0.010654127225279808, + 0.010867033153772354, + 0.005192731041461229, + 0.009815609082579613, + 0.009598621167242527, + 0.002091562608256936, + 0.0020370574202388525, + 0.0025897794403135777, + 0.002959014382213354, + 0.004301507491618395, + 0.002099965000525117, + 0.003119664965197444, + 0.0026890290901064873, + 0.005924236495047808, + 0.006723703350871801, + 0.008282231166958809, + 0.021576300263404846, + 0.009508639574050903, + 0.007219862192869186, + 0.007081362418830395, + 0.00797894224524498, + 0.0032177662942558527, + 0.00410515908151865, + 0.004933085758239031, + 0.006447539199143648, + 0.007187493145465851, + 0.006051163654774427, + 0.004773824010044336, + 0.007552325259894133, + 0.0029901270754635334, + 0.006340119056403637, + 0.004475256893783808, + 0.018585016950964928, + 0.0073143369518220425, + 0.01030033640563488, + 0.004402817226946354, + 0.005874053109437227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.06203349679708481 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle tracking shot follows a chariot, its wheels tumbling along a bumpy ancient road, creating a subtle yet authentic jolt of movement. The magnificent chariot, crafted from dark lacquered wood, its body intricately carved with cloud patterns and relief carvings of mythical beasts, gleams with a deep patina etched by time, its bronze components reflecting a cold, hard gleam against the gloomy sky. Within the carriage, a stern-faced, middle-aged emperor sits poised, clad in a dragon robe of deep black silk. A five-clawed golden dragon, delicately embroidered with gold thread, appears faintly visible in the dim light, its scales seemingly undulating. The emperor's brow furrows, his lips pressed into a tight line. His deep gaze gazes into the void ahead, revealing a complex blend of worry and determination. His hands clasp the jade belt at his knees, his knuckles white with the strain. A chariot travels along an ancient loess road deeply carved by ruts. Its massive wooden wheels roll over the dry ground, stirring up clouds of fine dust. The sky is covered by thick, leaden clouds, and the diffuse, oppressive light casts a cold, desaturated hue throughout the painting. Only the dark gold threads of the dragon robes provide the sole warm accent, creating a tense and austere atmosphere of impending storm.", + "seed": 29260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29260_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": -0.05506726913154125, + "vlm_logprob": -0.05506726913154125, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.05506726913154125, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09349353611469269 + ], + "opposite_logprob_tokens": [ + -2.4684934616088867 + ], + "logprob_reward": -0.09349353611469269, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09349353611469269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08217663317918777, + 0.055544413626194, + 0.05673444643616676, + 0.14448726177215576, + 0.056757986545562744, + 0.06661780178546906, + 0.2502457797527313, + 0.026340005919337273, + 0.1966913789510727, + 0.06440425664186478 + ], + "credit_map": [ + 0.006434478331357241, + 0.0076938653364777565, + 0.013691229745745659, + 0.014782794751226902, + 0.01210078876465559, + 0.008409786969423294, + 0.010043726302683353, + 0.00901996623724699, + 0.005538719240576029, + 0.008773693814873695, + 0.005697937682271004, + 0.006814127787947655, + 0.009243068285286427, + 0.005520410370081663, + 0.005673382896929979, + 0.008283071219921112, + 0.00410806480795145, + 0.004409282002598047, + 0.008860908448696136, + 0.007360799238085747, + 0.004258066415786743, + 0.006224340293556452, + 0.014330544508993626, + 0.0071824416518211365, + 0.01295240968465805, + 0.010913340374827385, + 0.022237267345190048, + 0.03455740958452225, + 0.02156519889831543, + 0.009962143376469612, + 0.017000719904899597, + 0.015298780053853989, + 0.005640932824462652, + 0.003849359229207039, + 0.00803659949451685, + 0.008285995572805405, + 0.008478871546685696, + 0.004279597662389278, + 0.006929861381649971, + 0.011256772093474865, + 0.004269828088581562, + 0.005545265041291714, + 0.010632671415805817, + 0.009249797090888023, + 0.010644041001796722, + 0.00524702575057745, + 0.011958055198192596, + 0.009071122854948044, + 0.014111200347542763, + 0.030353467911481857, + 0.042064324021339417, + 0.030942250043153763, + 0.03793014585971832, + 0.017478113994002342, + 0.027019567787647247, + 0.050346702337265015, + 0.002522615483030677, + 0.004997021518647671, + 0.0032712602987885475, + 0.0036913633812218904, + 0.004028800409287214, + 0.001710322918370366, + 0.0037935704458504915, + 0.002325051696971059, + 0.01588689535856247, + 0.025538070127367973, + 0.021695932373404503, + 0.03304743021726608, + 0.021879103034734726, + 0.018873149529099464, + 0.02606802247464657, + 0.033702775835990906, + 0.005334125831723213, + 0.006579190026968718, + 0.007532502990216017, + 0.010410028509795666, + 0.010010438971221447, + 0.0037942996714264154, + 0.013995828106999397, + 0.006747844163328409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.09349353611469269 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062078870832920074 + ], + "opposite_logprob_tokens": [ + -2.8120789527893066 + ], + "logprob_reward": -0.062078870832920074, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062078870832920074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13812243938446045, + 0.04744955524802208, + 0.08141188323497772, + 0.05500388145446777, + 0.09238684177398682, + 0.04416196048259735, + 0.24050791561603546, + 0.07718168199062347, + 0.12113209813833237, + 0.10264164954423904 + ], + "credit_map": [ + 0.012245547957718372, + 0.011671660467982292, + 0.018618721514940262, + 0.021363332867622375, + 0.026086607947945595, + 0.0164777971804142, + 0.012962928041815758, + 0.018695874139666557, + 0.005921815987676382, + 0.010612904094159603, + 0.005217927508056164, + 0.0054549043998122215, + 0.005412769503891468, + 0.005224665626883507, + 0.004441775381565094, + 0.005162800196558237, + 0.008234462700784206, + 0.0071791997179389, + 0.015557626262307167, + 0.010360085405409336, + 0.006939858663827181, + 0.006924879737198353, + 0.0162795577198267, + 0.009936229325830936, + 0.004649645648896694, + 0.004578887484967709, + 0.008675451390445232, + 0.00952170416712761, + 0.007716763764619827, + 0.005323952529579401, + 0.006694087292999029, + 0.007843400351703167, + 0.009372066706418991, + 0.005744923371821642, + 0.01905699074268341, + 0.017850009724497795, + 0.012264635413885117, + 0.007812210358679295, + 0.009188033640384674, + 0.011097986251115799, + 0.003181854961439967, + 0.0037732566706836224, + 0.006410875823348761, + 0.0060850451700389385, + 0.007825090549886227, + 0.0032785655930638313, + 0.008360441774129868, + 0.005246836692094803, + 0.011932951398193836, + 0.02800290659070015, + 0.04736746847629547, + 0.04390261322259903, + 0.02698252536356449, + 0.014703688211739063, + 0.036872297525405884, + 0.03074348345398903, + 0.010163870640099049, + 0.010936192236840725, + 0.008426420390605927, + 0.011349257081747055, + 0.010863246396183968, + 0.005569636821746826, + 0.009037401527166367, + 0.010835676454007626, + 0.011103677563369274, + 0.01361646968871355, + 0.010453543625772, + 0.025252018123865128, + 0.015428354032337666, + 0.009252107702195644, + 0.01800033263862133, + 0.018025614321231842, + 0.008708199486136436, + 0.011901416815817356, + 0.01417352445423603, + 0.01524888165295124, + 0.018004510551691055, + 0.0065924194641411304, + 0.014777767471969128, + 0.013234945014119148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.062078870832920074 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369487449526787 + ], + "opposite_logprob_tokens": [ + -3.4086949825286865 + ], + "logprob_reward": -0.03369487449526787, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369487449526787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18734769523143768, + 0.04723241552710533, + 0.042439114302396774, + 0.062258031219244, + 0.05058744549751282, + 0.0470275804400444, + 0.09569118171930313, + 0.08013518899679184, + 0.2866899073123932, + 0.10059141367673874 + ], + "credit_map": [ + 0.024214958772063255, + 0.022465983405709267, + 0.027375314384698868, + 0.02327212691307068, + 0.03741078823804855, + 0.02504109963774681, + 0.01390037126839161, + 0.01366705447435379, + 0.0056211198680102825, + 0.00551279541105032, + 0.006265115924179554, + 0.004127357620745897, + 0.008788890205323696, + 0.005159973632544279, + 0.004751075524836779, + 0.007006086874753237, + 0.0029870662838220596, + 0.0035344220232218504, + 0.010190857574343681, + 0.006479834672063589, + 0.002635406330227852, + 0.003840159624814987, + 0.008133216761052608, + 0.0046381535939872265, + 0.006227371282875538, + 0.0050081112422049046, + 0.009658358991146088, + 0.012545829638838768, + 0.0070462836883962154, + 0.005067044403403997, + 0.009390061721205711, + 0.007314969785511494, + 0.004081029444932938, + 0.003586795646697283, + 0.010655814781785011, + 0.009437279775738716, + 0.007211015094071627, + 0.0034881229512393475, + 0.0046887025237083435, + 0.007438688538968563, + 0.0032229938078671694, + 0.0037573061417788267, + 0.005598367191851139, + 0.005589446984231472, + 0.009686768054962158, + 0.0028227989096194506, + 0.010167011991143227, + 0.0061828880570828915, + 0.006356869824230671, + 0.012495279312133789, + 0.0124833257868886, + 0.013071887195110321, + 0.017894884571433067, + 0.005465998779982328, + 0.014947880059480667, + 0.012975051999092102, + 0.004689454101026058, + 0.011495628394186497, + 0.008622867055237293, + 0.00981743074953556, + 0.02152622677385807, + 0.004268629942089319, + 0.01047799363732338, + 0.009236958809196949, + 0.01733262464404106, + 0.028324995189905167, + 0.02218039333820343, + 0.06661712378263474, + 0.03141707554459572, + 0.02379949763417244, + 0.033930547535419464, + 0.06308764219284058, + 0.007507701404392719, + 0.011084585450589657, + 0.010194568894803524, + 0.013307696208357811, + 0.014895780012011528, + 0.005816084798425436, + 0.02479531243443489, + 0.012989684008061886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03369487449526787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064027324318886 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064027324318886, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064027324318886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4472667872905731, + 0.08463926613330841, + 0.037427328526973724, + 0.03434019163250923, + 0.022242875769734383, + 0.04225294291973114, + 0.05645526945590973, + 0.019860535860061646, + 0.23540690541267395, + 0.02010784111917019 + ], + "credit_map": [ + 0.038680851459503174, + 0.04526309669017792, + 0.036595650017261505, + 0.05585075542330742, + 0.1109110563993454, + 0.06509508192539215, + 0.04800955578684807, + 0.04686080664396286, + 0.0066505600698292255, + 0.0095106465741992, + 0.009087039157748222, + 0.01057629194110632, + 0.020540505647659302, + 0.009320564568042755, + 0.008551472797989845, + 0.010402187705039978, + 0.002122563309967518, + 0.0025128223933279514, + 0.0042467121966183186, + 0.004852505400776863, + 0.002877462189644575, + 0.005054769571870565, + 0.010418313555419445, + 0.005342183634638786, + 0.003849967150017619, + 0.003212404204532504, + 0.005422686692327261, + 0.005902585573494434, + 0.004754654597491026, + 0.0036396626383066177, + 0.0042923856526613235, + 0.0032658474519848824, + 0.0024172610137611628, + 0.0015097128925845027, + 0.004307982977479696, + 0.0035688895732164383, + 0.003502986626699567, + 0.0017364496598020196, + 0.002684383187443018, + 0.002515213331207633, + 0.0030748704448342323, + 0.0035579900722950697, + 0.00897185131907463, + 0.004756152629852295, + 0.005993535742163658, + 0.006302986294031143, + 0.004708023741841316, + 0.004887536633759737, + 0.0033035380765795708, + 0.008095216937363148, + 0.008596726693212986, + 0.007593791466206312, + 0.009439277462661266, + 0.003758203238248825, + 0.006019847933202982, + 0.009648670442402363, + 0.0018520860467106104, + 0.0029747444204986095, + 0.002186728408560157, + 0.0024780340027064085, + 0.0043865349143743515, + 0.0014699752209708095, + 0.0020233234390616417, + 0.0024891102220863104, + 0.014858350157737732, + 0.023468706756830215, + 0.020665761083364487, + 0.04249195009469986, + 0.03892377018928528, + 0.03843938931822777, + 0.02655499428510666, + 0.03000400774180889, + 0.0015160006005316973, + 0.0022299685515463352, + 0.002400899538770318, + 0.0021224082447588444, + 0.003559095086529851, + 0.0015826496528461576, + 0.004278204403817654, + 0.002418618183583021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.016064027324318886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000503689050674 + ], + "opposite_logprob_tokens": [ + -2.695004940032959 + ], + "logprob_reward": -0.07000503689050674, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000503689050674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16565442085266113, + 0.06092092767357826, + 0.04459446296095848, + 0.08753085881471634, + 0.0416363961994648, + 0.11044694483280182, + 0.2541743516921997, + 0.02560640685260296, + 0.1528896540403366, + 0.05654558911919594 + ], + "credit_map": [ + 0.02391332946717739, + 0.022020617499947548, + 0.025344330817461014, + 0.018689291551709175, + 0.026806266978383064, + 0.01916401833295822, + 0.014096355997025967, + 0.015620212070643902, + 0.00688836257904768, + 0.010446729138493538, + 0.008098413236439228, + 0.007078223861753941, + 0.007894088514149189, + 0.006381949409842491, + 0.004799426067620516, + 0.009333735331892967, + 0.003494107164442539, + 0.0038120022509247065, + 0.010695700533688068, + 0.006138944998383522, + 0.003777860663831234, + 0.0033109041396528482, + 0.008645018562674522, + 0.004719921853393316, + 0.010025017894804478, + 0.006252303719520569, + 0.015400177799165249, + 0.010556100867688656, + 0.015324236825108528, + 0.006165400613099337, + 0.011643883772194386, + 0.012163734994828701, + 0.0033982947934418917, + 0.003947519231587648, + 0.004818062297999859, + 0.008367041125893593, + 0.0061285956762731075, + 0.002862881403416395, + 0.005877634510397911, + 0.0062363664619624615, + 0.0075758458115160465, + 0.012453696690499783, + 0.02124631404876709, + 0.01390866655856371, + 0.017034346237778664, + 0.008555438369512558, + 0.015947429463267326, + 0.013725205324590206, + 0.01719001680612564, + 0.03235550969839096, + 0.033108048141002655, + 0.024986932054162025, + 0.042560700327157974, + 0.02467080019414425, + 0.027388237416744232, + 0.051914118230342865, + 0.001978390384465456, + 0.003445221344009042, + 0.003588030580431223, + 0.00513415364548564, + 0.0022618300281465054, + 0.0031756567768752575, + 0.0025141797959804535, + 0.0035089440643787384, + 0.010180090554058552, + 0.015010509639978409, + 0.018508316949009895, + 0.030187435448169708, + 0.018120847642421722, + 0.023659218102693558, + 0.01319354958832264, + 0.024029698222875595, + 0.004252874292433262, + 0.005440303590148687, + 0.004921879153698683, + 0.0073754326440393925, + 0.011866659857332706, + 0.004243603441864252, + 0.012220553122460842, + 0.0062242792919278145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07000503689050674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, low-angle tracking shot follows a chariot, its wheels tumbling along a bumpy ancient road, creating a subtle yet authentic jolt of movement. The magnificent chariot, crafted from dark lacquered wood, its body intricately carved with cloud patterns and relief carvings of mythical beasts, gleams with a deep patina etched by time, its bronze components reflecting a cold, hard gleam against the gloomy sky. Within the carriage, a stern-faced, middle-aged emperor sits poised, clad in a dragon robe of deep black silk. A five-clawed golden dragon, delicately embroidered with gold thread, appears faintly visible in the dim light, its scales seemingly undulating. The emperor's brow furrows, his lips pressed into a tight line. His deep gaze gazes into the void ahead, revealing a complex blend of worry and determination. His hands clasp the jade belt at his knees, his knuckles white with the strain. A chariot travels along an ancient loess road deeply carved by ruts. Its massive wooden wheels roll over the dry ground, stirring up clouds of fine dust. The sky is covered by thick, leaden clouds, and the diffuse, oppressive light casts a cold, desaturated hue throughout the painting. Only the dark gold threads of the dragon robes provide the sole warm accent, creating a tense and austere atmosphere of impending storm.", + "seed": 29261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_7/index_6874_seed_29261_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": -0.11048106476664543, + "vlm_logprob": -0.11048106476664543, + "vlm_yes_no_margin": 2.674999976158142, + "vlm_token_credit": -0.11048106476664543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17053161561489105 + ], + "opposite_logprob_tokens": [ + -2.0455315113067627 + ], + "logprob_reward": -0.17053161561489105, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.17053161561489105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17417572438716888, + 0.06313180923461914, + 0.051196206361055374, + 0.2451390027999878, + 0.047344572842121124, + 0.036237046122550964, + 0.017316482961177826, + 0.020999830216169357, + 0.053930629044771194, + 0.29052865505218506 + ], + "credit_map": [ + 0.022923270240426064, + 0.02553129941225052, + 0.019609974697232246, + 0.01159198209643364, + 0.03141721710562706, + 0.03380044922232628, + 0.010797949507832527, + 0.01850358210504055, + 0.005312647204846144, + 0.012242485769093037, + 0.009206995368003845, + 0.0069256918504834175, + 0.009048759937286377, + 0.008293048478662968, + 0.005139465443789959, + 0.006962718442082405, + 0.003708183066919446, + 0.005609740037471056, + 0.0073949117213487625, + 0.0057875425554811954, + 0.008297216147184372, + 0.006769062951207161, + 0.003740232903510332, + 0.00988931767642498, + 0.04378115385770798, + 0.030784647911787033, + 0.02261347509920597, + 0.02361961454153061, + 0.0367286279797554, + 0.03491324186325073, + 0.032349321991205215, + 0.020348934456706047, + 0.009630583226680756, + 0.004198506008833647, + 0.00783886294811964, + 0.005381683353334665, + 0.006063982378691435, + 0.00477325776591897, + 0.004219586029648781, + 0.005238112062215805, + 0.0028858035802841187, + 0.0029483791440725327, + 0.0041420115157961845, + 0.005988043732941151, + 0.006524160970002413, + 0.00444292975589633, + 0.003626123070716858, + 0.0056795948185026646, + 0.0025805833283811808, + 0.0017951298505067825, + 0.002793509280309081, + 0.00183800735976547, + 0.002550107426941395, + 0.0021215847227722406, + 0.0015025021275505424, + 0.0021350577007979155, + 0.0024958413559943438, + 0.0023986015003174543, + 0.0026577680837363005, + 0.0024142570327967405, + 0.003326024394482374, + 0.0025416596326977015, + 0.003135116072371602, + 0.002030562376603484, + 0.007367311976850033, + 0.004739623516798019, + 0.007992593571543694, + 0.00897462759166956, + 0.00541814835742116, + 0.0046285358257591724, + 0.008956048637628555, + 0.005853736773133278, + 0.05336593836545944, + 0.024301547557115555, + 0.045234616845846176, + 0.07258789986371994, + 0.022029167041182518, + 0.019165409728884697, + 0.022354016080498695, + 0.03149006888270378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.17053161561489105 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2016587108373642 + ], + "opposite_logprob_tokens": [ + -1.7016587257385254 + ], + "logprob_reward": -0.2016587108373642, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2016587108373642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11590118706226349, + 0.08745600283145905, + 0.06618642807006836, + 0.038591835647821426, + 0.028226926922798157, + 0.029157549142837524, + 0.021361907944083214, + 0.017063472419977188, + 0.07328389585018158, + 0.5227708220481873 + ], + "credit_map": [ + 0.015361285768449306, + 0.020778557285666466, + 0.013738833367824554, + 0.010770347900688648, + 0.014618580229580402, + 0.016531487926840782, + 0.010855413973331451, + 0.013246677815914154, + 0.006039739120751619, + 0.015027588233351707, + 0.009493902325630188, + 0.0054701389744877815, + 0.018431466072797775, + 0.018730273470282555, + 0.005879448261111975, + 0.008383437059819698, + 0.006560091394931078, + 0.016586797311902046, + 0.008455033414065838, + 0.007461328525096178, + 0.009020834229886532, + 0.007421063259243965, + 0.004803093615919352, + 0.005878184922039509, + 0.004740839824080467, + 0.007524214684963226, + 0.005646776407957077, + 0.004252125043421984, + 0.004069528542459011, + 0.004842683207243681, + 0.00362213421612978, + 0.003893532557412982, + 0.005614934489130974, + 0.002669306006282568, + 0.005288445856422186, + 0.0034884659107774496, + 0.0031184444669634104, + 0.0027380611281841993, + 0.002150217304006219, + 0.0031590498983860016, + 0.0047910562716424465, + 0.003238817909732461, + 0.0039224824868142605, + 0.002778499387204647, + 0.00438228202983737, + 0.0035724181216210127, + 0.002635487588122487, + 0.0038365027867257595, + 0.002160389441996813, + 0.0019087905529886484, + 0.0038044375833123922, + 0.0024867334868758917, + 0.0017755249282345176, + 0.0018792819464579225, + 0.002198878675699234, + 0.005147868301719427, + 0.0010224151192232966, + 0.001850558677688241, + 0.002022980246692896, + 0.0028296830132603645, + 0.003088540630415082, + 0.0022611019667237997, + 0.0019892307464033365, + 0.0019989623688161373, + 0.008116284385323524, + 0.007096815388649702, + 0.010868173092603683, + 0.01083174254745245, + 0.006426475942134857, + 0.006043019704520702, + 0.012209279462695122, + 0.011692099273204803, + 0.09278470277786255, + 0.044720955193042755, + 0.08601987361907959, + 0.13867932558059692, + 0.03054109215736389, + 0.0328412689268589, + 0.03831052780151367, + 0.05887302756309509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.2016587108373642 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263423640280962 + ], + "opposite_logprob_tokens": [ + -3.6513423919677734 + ], + "logprob_reward": -0.0263423640280962, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263423640280962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13042128086090088, + 0.08323828876018524, + 0.14006425440311432, + 0.06582503765821457, + 0.03714683651924133, + 0.14594782888889313, + 0.047739427536726, + 0.09057740867137909, + 0.08225148916244507, + 0.17678821086883545 + ], + "credit_map": [ + 0.01871037296950817, + 0.023571766912937164, + 0.016122307628393173, + 0.010243779979646206, + 0.018667597323656082, + 0.01982915960252285, + 0.01101644616574049, + 0.012259842827916145, + 0.00807457510381937, + 0.01762075535953045, + 0.010304723866283894, + 0.0064828647300601006, + 0.013518369756639004, + 0.011849673464894295, + 0.006564178038388491, + 0.0088231535628438, + 0.0135234035551548, + 0.024324586614966393, + 0.014251256361603737, + 0.015426850877702236, + 0.02560754492878914, + 0.019102374091744423, + 0.014665178954601288, + 0.013163045048713684, + 0.007888379506766796, + 0.010896318592131138, + 0.008153285831212997, + 0.008210520260035992, + 0.00934261828660965, + 0.009456782601773739, + 0.007110833190381527, + 0.004766302648931742, + 0.0037323106080293655, + 0.0046958597376942635, + 0.005510938353836536, + 0.004073606338351965, + 0.007747091352939606, + 0.005388088058680296, + 0.0026324663776904345, + 0.0033664763905107975, + 0.010121103376150131, + 0.013075331225991249, + 0.028449274599552155, + 0.02010851725935936, + 0.02080303244292736, + 0.01643812097609043, + 0.013274313881993294, + 0.02367813140153885, + 0.008723936975002289, + 0.0053751785308122635, + 0.0071891723200678825, + 0.004507903940975666, + 0.00496662687510252, + 0.005406263750046492, + 0.003517458215355873, + 0.008052884601056576, + 0.00614119041711092, + 0.007469032425433397, + 0.01025349646806717, + 0.018406521528959274, + 0.017449242994189262, + 0.012386185117065907, + 0.010427664965391159, + 0.008044076152145863, + 0.008879068307578564, + 0.009008788503706455, + 0.011904682964086533, + 0.011319858022034168, + 0.009011386893689632, + 0.007874579168856144, + 0.014118701219558716, + 0.010134425945580006, + 0.027885453775525093, + 0.015768837183713913, + 0.02887382172048092, + 0.04687970504164696, + 0.011419164948165417, + 0.011163095012307167, + 0.012683052569627762, + 0.022115079686045647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.0263423640280962 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32024866342544556, + 0.30399057269096375, + 0.054529912769794464, + 0.034362416714429855, + 0.030986052006483078, + 0.0748472511768341, + 0.05694068595767021, + 0.01760314591228962, + 0.02119199000298977, + 0.0852992907166481 + ], + "credit_map": [ + 0.015361679717898369, + 0.0213618166744709, + 0.011270778253674507, + 0.011120788753032684, + 0.1213906779885292, + 0.11810752749443054, + 0.010490783490240574, + 0.011144593358039856, + 0.027727607637643814, + 0.029458828270435333, + 0.039303094148635864, + 0.019797198474407196, + 0.0660039409995079, + 0.052362434566020966, + 0.023327047005295753, + 0.04601043835282326, + 0.004784451797604561, + 0.011898651719093323, + 0.007580123841762543, + 0.005768290255218744, + 0.008387057110667229, + 0.006661452353000641, + 0.003952307626605034, + 0.005497576668858528, + 0.0035681670997291803, + 0.005963020958006382, + 0.004424378275871277, + 0.0038211671635508537, + 0.0046615153551101685, + 0.004651679657399654, + 0.0032544785644859076, + 0.004018012434244156, + 0.005927616264671087, + 0.0035658462438732386, + 0.004889674484729767, + 0.003440270433202386, + 0.004541520029306412, + 0.0034555094316601753, + 0.0023172318469733, + 0.002848382107913494, + 0.0060369353741407394, + 0.00836495216935873, + 0.011910634115338326, + 0.008419216610491276, + 0.011942334473133087, + 0.008477810770273209, + 0.006594665348529816, + 0.013100700452923775, + 0.009290973655879498, + 0.006997460499405861, + 0.009222086519002914, + 0.006131262984126806, + 0.0038459838833659887, + 0.006503509357571602, + 0.005406079348176718, + 0.009543332271277905, + 0.0018786709988489747, + 0.0017963004065677524, + 0.0026086727157235146, + 0.0026102338451892138, + 0.002582757268100977, + 0.002025968860834837, + 0.00208059954456985, + 0.0020199441350996494, + 0.002959836507216096, + 0.0028587020933628082, + 0.0025585161056369543, + 0.0037563578225672245, + 0.002050526672974229, + 0.001965623814612627, + 0.0025637187063694, + 0.0024787080474197865, + 0.014726071618497372, + 0.007160892244428396, + 0.014481369405984879, + 0.023055486381053925, + 0.00517390388995409, + 0.0052790860645473, + 0.006092898081988096, + 0.00932958722114563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.011075684800744057 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a chariot traveling along an ancient road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the chariot crafted from dark lacquered wood with intricate carvings of cloud patterns and mythical beasts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a stern-faced, middle-aged emperor sitting inside the chariot wearing a deep black silk dragon robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky covered by thick, leaden clouds with a cold, desaturated hue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot tight and low-angle, tracking the chariot as it moves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14279694855213165 + ], + "opposite_logprob_tokens": [ + -2.0177969932556152 + ], + "logprob_reward": -0.14279694855213165, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14279694855213165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09305700659751892, + 0.07559435069561005, + 0.05073288455605507, + 0.11986859887838364, + 0.26991236209869385, + 0.06344183534383774, + 0.05856376886367798, + 0.05554280802607536, + 0.09490213543176651, + 0.1183842271566391 + ], + "credit_map": [ + 0.006609746254980564, + 0.008683053776621819, + 0.01475922018289566, + 0.006978420075029135, + 0.013978932984173298, + 0.01820722036063671, + 0.008799591101706028, + 0.015040823258459568, + 0.006085845176130533, + 0.007690488360822201, + 0.01791464537382126, + 0.007510886061936617, + 0.011445549316704273, + 0.009504340589046478, + 0.006677199620753527, + 0.008765396662056446, + 0.005259775090962648, + 0.010638147592544556, + 0.00600852444767952, + 0.005828595254570246, + 0.0071656242944300175, + 0.005617591552436352, + 0.004802013747394085, + 0.005412611179053783, + 0.010894352570176125, + 0.024454481899738312, + 0.01800447143614292, + 0.013959264382719994, + 0.014125368557870388, + 0.01630941964685917, + 0.011049951426684856, + 0.01107128243893385, + 0.0706518292427063, + 0.024428343400359154, + 0.047155626118183136, + 0.029843434691429138, + 0.028328346088528633, + 0.03007473610341549, + 0.015571033582091331, + 0.023858999833464622, + 0.005695799365639687, + 0.006527014542371035, + 0.009064576588571072, + 0.008808991871774197, + 0.013291886076331139, + 0.008182279765605927, + 0.006011521443724632, + 0.005859756842255592, + 0.006660462357103825, + 0.009697671979665756, + 0.006596159655600786, + 0.006290629040449858, + 0.006245988421142101, + 0.00647583557292819, + 0.00504604447633028, + 0.011550978757441044, + 0.0062249889597296715, + 0.007698033470660448, + 0.007812310941517353, + 0.007226909510791302, + 0.007048446219414473, + 0.006818149704486132, + 0.008427366614341736, + 0.004286600276827812, + 0.014070216566324234, + 0.010565968230366707, + 0.013127400539815426, + 0.017479633912444115, + 0.010328210890293121, + 0.008027597330510616, + 0.011658228002488613, + 0.009644880890846252, + 0.028216321021318436, + 0.012786167673766613, + 0.017906153574585915, + 0.021961327642202377, + 0.009442407637834549, + 0.007963675074279308, + 0.009206038899719715, + 0.010902137495577335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.14279694855213165 + } + ], + "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/0000055/group_8/index_tensor([4579, 4579])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_tensor([4579, 4579])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..37112ffb8c98c4d2222ae47f589da64cb9c83a73 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_tensor([4579, 4579])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic low-angle tracking shot, the camera parallel to a galloping horse, moving close to the ground to emphasize its thunderous speed. In the frame, a dusty, ancient scholar leans over his horse, his hair bun blown loose by the wind, leaving strands of black hair, mingled with sweat, plastered to his dusty cheeks. His eyes reveal an anxious yet resolute expression. He wears a robe of coarse cloth, once a bluish-gray color, now mottled with yellow dust. The wide cuffs and hem whip wildly backward in the wind. He grips the reins tightly, his body heaving and swaying with the rhythm of the muscular, sweat-drenched dark brown steed beneath him. Each swift and powerful thump of the horse's hooves on the dry, cracked, winding mountain path stirs up a thick cloud of yellow sand. Illuminated by the golden evening sun, this dust lurches like a swirling, luminous mist, trailing them closely behind. The background is rolling desolate mountains and sparse dead trees. The whole scene is dominated by warm yellow tones, creating a vast, urgent and epic atmosphere.", + "seed": 19362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_4579_seed_19362_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": -0.02933658061083406, + "vlm_logprob": -0.02933658061083406, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02933658061083406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09499121457338333 + ], + "opposite_logprob_tokens": [ + -2.8449912071228027 + ], + "logprob_reward": -0.09499121457338333, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09499121457338333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09818567335605621, + 0.037949301302433014, + 0.041411980986595154, + 0.11430355906486511, + 0.11024051159620285, + 0.269523561000824, + 0.020527686923742294, + 0.0888749361038208, + 0.1502005159854889, + 0.06878232210874557 + ], + "credit_map": [ + 0.015853017568588257, + 0.007722564972937107, + 0.018497033044695854, + 0.006822802126407623, + 0.008647539652884007, + 0.015905581414699554, + 0.008730178698897362, + 0.01600695587694645, + 0.005830198992043734, + 0.003403594484552741, + 0.0036721203941851854, + 0.005088441539555788, + 0.004917093086987734, + 0.00507928803563118, + 0.006575221661478281, + 0.0033833442721515894, + 0.0073037478141486645, + 0.004170330241322517, + 0.004190570674836636, + 0.0032333391718566418, + 0.006591122597455978, + 0.005177742335945368, + 0.005123559385538101, + 0.005621569696813822, + 0.015065058134496212, + 0.018564976751804352, + 0.025104830041527748, + 0.007011446636170149, + 0.014249789528548717, + 0.014272386208176613, + 0.00940459780395031, + 0.010630476288497448, + 0.02005009539425373, + 0.013397570699453354, + 0.011484585702419281, + 0.00762430764734745, + 0.012115521356463432, + 0.01741364412009716, + 0.011025681160390377, + 0.01712910458445549, + 0.04449807107448578, + 0.058547575026750565, + 0.03351374715566635, + 0.021835286170244217, + 0.027898048982024193, + 0.028072122484445572, + 0.026817314326763153, + 0.028341393917798996, + 0.002348516369238496, + 0.002042369917035103, + 0.0020199627615511417, + 0.002497562440112233, + 0.0032080805394798517, + 0.002618448343127966, + 0.0035626355092972517, + 0.0022301110439002514, + 0.007152050733566284, + 0.011064316146075726, + 0.012753181159496307, + 0.006259994115680456, + 0.013980414718389511, + 0.016294488683342934, + 0.012351659126579762, + 0.009018833748996258, + 0.011834971606731415, + 0.027223696932196617, + 0.025040898472070694, + 0.01298750564455986, + 0.034344132989645004, + 0.014633004553616047, + 0.014151360839605331, + 0.009984939359128475, + 0.007749220356345177, + 0.008610378950834274, + 0.0067550791427493095, + 0.008368728682398796, + 0.00846127513796091, + 0.01076536811888218, + 0.008992535062134266, + 0.009079739451408386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.09499121457338333 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226679690182209 + ], + "opposite_logprob_tokens": [ + -4.264226913452148 + ], + "logprob_reward": -0.014226679690182209, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226679690182209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1850218027830124, + 0.07089340686798096, + 0.12381597608327866, + 0.16282665729522705, + 0.07484770566225052, + 0.14207866787910461, + 0.06841692328453064, + 0.05399489775300026, + 0.045539893209934235, + 0.07256411015987396 + ], + "credit_map": [ + 0.016952749341726303, + 0.01402165275067091, + 0.018651384860277176, + 0.0100715896114707, + 0.018732786178588867, + 0.04371270164847374, + 0.015224416740238667, + 0.047654516994953156, + 0.009273545816540718, + 0.007936564274132252, + 0.008678760379552841, + 0.005148774012923241, + 0.008342108689248562, + 0.014839115552604198, + 0.005228902213275433, + 0.011445634998381138, + 0.014818191528320312, + 0.012272602878510952, + 0.009614392183721066, + 0.00709143141284585, + 0.012118428014218807, + 0.030560221523046494, + 0.012023502960801125, + 0.025317197665572166, + 0.021293118596076965, + 0.01995006762444973, + 0.03032701089978218, + 0.009066571481525898, + 0.026961399242281914, + 0.017993248999118805, + 0.015510106459259987, + 0.021725133061408997, + 0.01707533933222294, + 0.009917063638567924, + 0.0077655380591750145, + 0.005598800722509623, + 0.008721825666725636, + 0.009508769027888775, + 0.008347172290086746, + 0.007913192734122276, + 0.018906088545918465, + 0.017418272793293, + 0.01601831428706646, + 0.009920675307512283, + 0.02288096956908703, + 0.019559018313884735, + 0.01567688398063183, + 0.02169843576848507, + 0.009080630727112293, + 0.008939400315284729, + 0.008181395940482616, + 0.009631199762225151, + 0.008822017349302769, + 0.007657030131667852, + 0.01004409696906805, + 0.006061155814677477, + 0.004266681615263224, + 0.013279927894473076, + 0.011274455115199089, + 0.0029500676319003105, + 0.005426174961030483, + 0.00727670406922698, + 0.005144850350916386, + 0.004376033786684275, + 0.0027160942554473877, + 0.008556309156119823, + 0.007477897685021162, + 0.0031762057915329933, + 0.00574047164991498, + 0.007981306873261929, + 0.005374982953071594, + 0.0045166281051933765, + 0.008202442899346352, + 0.010000567883253098, + 0.006701076403260231, + 0.00869221892207861, + 0.0077387383207678795, + 0.011064752005040646, + 0.011924093589186668, + 0.00824022851884365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.014226679690182209 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958294961601496 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958294961601496, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958294961601496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11045557260513306, + 0.05960497632622719, + 0.10780761390924454, + 0.1006283089518547, + 0.07267118245363235, + 0.09645356982946396, + 0.07204387336969376, + 0.19947552680969238, + 0.12766705453395844, + 0.05319233611226082 + ], + "credit_map": [ + 0.01754201389849186, + 0.009017286822199821, + 0.017708610743284225, + 0.009409474208950996, + 0.012629562057554722, + 0.016021372750401497, + 0.011708664707839489, + 0.01641859859228134, + 0.0060693430714309216, + 0.004625029396265745, + 0.007409412879496813, + 0.0032445925753563643, + 0.005955156404525042, + 0.013450275175273418, + 0.005057302303612232, + 0.013793863356113434, + 0.011719613336026669, + 0.009685940109193325, + 0.007664910517632961, + 0.00532328337430954, + 0.008363292552530766, + 0.032263029366731644, + 0.006515886168926954, + 0.026271650567650795, + 0.012146900407969952, + 0.014844108372926712, + 0.02040351741015911, + 0.005449284799396992, + 0.01692265458405018, + 0.011464162729680538, + 0.007971505634486675, + 0.011426174081861973, + 0.012976093217730522, + 0.009054824709892273, + 0.006464364472776651, + 0.007944201119244099, + 0.008322234265506268, + 0.00930142030119896, + 0.005789889022707939, + 0.012818151153624058, + 0.017327284440398216, + 0.015385319478809834, + 0.010370499454438686, + 0.009865943342447281, + 0.01552683487534523, + 0.012978099286556244, + 0.007412962149828672, + 0.007586624938994646, + 0.010067129507660866, + 0.007507830858230591, + 0.007563706487417221, + 0.008815518580377102, + 0.012434050440788269, + 0.010407710447907448, + 0.009221477434039116, + 0.006026451475918293, + 0.016815150156617165, + 0.024151423946022987, + 0.019028112292289734, + 0.01586259715259075, + 0.03572713956236839, + 0.03778328746557236, + 0.02851528860628605, + 0.021592527627944946, + 0.009396176785230637, + 0.017478374764323235, + 0.02058471366763115, + 0.013773525133728981, + 0.02300955541431904, + 0.020900819450616837, + 0.010544360615313053, + 0.01197951752692461, + 0.006167717278003693, + 0.007541855331510305, + 0.005176878534257412, + 0.00609457166865468, + 0.0050353179685771465, + 0.008322054520249367, + 0.007368525955826044, + 0.007485414855182171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.005958294961601496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979299984872341 + ], + "opposite_logprob_tokens": [ + -3.5297930240631104 + ], + "logprob_reward": -0.02979299984872341, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979299984872341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1134679913520813, + 0.09566538035869598, + 0.3490777313709259, + 0.05064864084124565, + 0.12574994564056396, + 0.017006630077958107, + 0.03584399074316025, + 0.07745619118213654, + 0.08604906499385834, + 0.049034398049116135 + ], + "credit_map": [ + 0.015023833140730858, + 0.011098251678049564, + 0.017580360174179077, + 0.008519493974745274, + 0.01029947865754366, + 0.02125585451722145, + 0.009408184327185154, + 0.020282533019781113, + 0.014382462948560715, + 0.009208749048411846, + 0.017102209851145744, + 0.007026799023151398, + 0.008879848755896091, + 0.016202060505747795, + 0.007150568068027496, + 0.015712691470980644, + 0.04163269326090813, + 0.02820374257862568, + 0.0224944856017828, + 0.013172398321330547, + 0.02060774713754654, + 0.11181946843862534, + 0.014265787787735462, + 0.0968814268708229, + 0.007514491677284241, + 0.007598798722028732, + 0.011505971662700176, + 0.0028664085548371077, + 0.006581531837582588, + 0.005782220512628555, + 0.0034544665832072496, + 0.005344751290977001, + 0.025084855034947395, + 0.022780533879995346, + 0.015166015364229679, + 0.008174162358045578, + 0.01278773508965969, + 0.011725073680281639, + 0.011290488764643669, + 0.018741076812148094, + 0.0035007644910365343, + 0.0032363974023610353, + 0.002333746524527669, + 0.0010635157814249396, + 0.0021054327953606844, + 0.001957634463906288, + 0.0015545226633548737, + 0.0012546166544780135, + 0.005188416689634323, + 0.0038641279097646475, + 0.003785293083637953, + 0.0033300386276096106, + 0.0054473914206027985, + 0.005388279911130667, + 0.005452824756503105, + 0.0033876183442771435, + 0.004949223715811968, + 0.023595327511429787, + 0.02260972559452057, + 0.003989699296653271, + 0.006445438601076603, + 0.005746353417634964, + 0.006317176856100559, + 0.003803243860602379, + 0.0067522116005420685, + 0.014449561014771461, + 0.014941111207008362, + 0.0077890963293612, + 0.01761586032807827, + 0.00835413672029972, + 0.009286291897296906, + 0.006860788911581039, + 0.006595652084797621, + 0.006494412664324045, + 0.005141132976859808, + 0.007011942099779844, + 0.006312608253210783, + 0.003793841926380992, + 0.009411991573870182, + 0.0042728157714009285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.02979299984872341 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017137139802798629 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017137139802798629, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017137139802798629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23233084380626678, + 0.08608179539442062, + 0.07757449895143509, + 0.10808155685663223, + 0.07762087136507034, + 0.03534834831953049, + 0.07851757854223251, + 0.08091431111097336, + 0.1850821077823639, + 0.03844815492630005 + ], + "credit_map": [ + 0.038047440350055695, + 0.01698998361825943, + 0.040703415870666504, + 0.01765415631234646, + 0.019543852657079697, + 0.0435129813849926, + 0.01786520704627037, + 0.03801377862691879, + 0.011929556727409363, + 0.007440455723553896, + 0.010617030784487724, + 0.004449347499758005, + 0.006638774648308754, + 0.019613998010754585, + 0.004660488106310368, + 0.02073213830590248, + 0.009749510325491428, + 0.0053189597092568874, + 0.005748755764216185, + 0.003036999376490712, + 0.008021090179681778, + 0.022575125098228455, + 0.005525585729628801, + 0.0175984725356102, + 0.012918300926685333, + 0.015877917408943176, + 0.025870712473988533, + 0.005699490197002888, + 0.019703377038240433, + 0.01131128054112196, + 0.007892802357673645, + 0.008807661011815071, + 0.012916725128889084, + 0.01370554231107235, + 0.009894097223877907, + 0.004405597224831581, + 0.00746937794610858, + 0.006627501454204321, + 0.006551740691065788, + 0.016050275415182114, + 0.007955816574394703, + 0.005069864448159933, + 0.0044780573807656765, + 0.00185297557618469, + 0.004654117859899998, + 0.002870784606784582, + 0.004765135236084461, + 0.003701594425365329, + 0.00994813535362482, + 0.009052535519003868, + 0.007777479011565447, + 0.010111447423696518, + 0.010080866515636444, + 0.010044209659099579, + 0.013718943111598492, + 0.007783953100442886, + 0.006636158097535372, + 0.014269586652517319, + 0.01273691188544035, + 0.010063552297651768, + 0.010271771810948849, + 0.00813239999115467, + 0.012341284193098545, + 0.006462638732045889, + 0.014471414498984814, + 0.02910727635025978, + 0.03210755065083504, + 0.01714414730668068, + 0.04000803455710411, + 0.018760662525892258, + 0.018767189234495163, + 0.01471584290266037, + 0.004726235754787922, + 0.0062284814193844795, + 0.0038640645798295736, + 0.003332821186631918, + 0.006190132349729538, + 0.005788429640233517, + 0.004163988400250673, + 0.00415399856865406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0017137139802798629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, the camera parallel to a galloping horse, moving close to the ground to emphasize its thunderous speed. In the frame, a dusty, ancient scholar leans over his horse, his hair bun blown loose by the wind, leaving strands of black hair, mingled with sweat, plastered to his dusty cheeks. His eyes reveal an anxious yet resolute expression. He wears a robe of coarse cloth, once a bluish-gray color, now mottled with yellow dust. The wide cuffs and hem whip wildly backward in the wind. He grips the reins tightly, his body heaving and swaying with the rhythm of the muscular, sweat-drenched dark brown steed beneath him. Each swift and powerful thump of the horse's hooves on the dry, cracked, winding mountain path stirs up a thick cloud of yellow sand. Illuminated by the golden evening sun, this dust lurches like a swirling, luminous mist, trailing them closely behind. The background is rolling desolate mountains and sparse dead trees. The whole scene is dominated by warm yellow tones, creating a vast, urgent and epic atmosphere.", + "seed": 19363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_4579_seed_19363_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": -0.04241390486713499, + "vlm_logprob": -0.04241390486713499, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04241390486713499, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1564330756664276 + ], + "opposite_logprob_tokens": [ + -2.03143310546875 + ], + "logprob_reward": -0.1564330756664276, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1564330756664276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10221506655216217, + 0.052154116332530975, + 0.1422375738620758, + 0.1212763711810112, + 0.027281293645501137, + 0.0955546647310257, + 0.24595455825328827, + 0.06941793113946915, + 0.06996040046215057, + 0.07394799590110779 + ], + "credit_map": [ + 0.007258994039148092, + 0.027429159730672836, + 0.022198352962732315, + 0.006630274001508951, + 0.013806706294417381, + 0.007920662872493267, + 0.009281709790229797, + 0.007689211517572403, + 0.00720583600923419, + 0.007027209736406803, + 0.005880970507860184, + 0.007539590820670128, + 0.005578297656029463, + 0.006634233053773642, + 0.0077136242762207985, + 0.0045743524096906185, + 0.018727565184235573, + 0.016580011695623398, + 0.015583967790007591, + 0.017077047377824783, + 0.017004383727908134, + 0.026977142319083214, + 0.014759517274796963, + 0.015527945943176746, + 0.04141605272889137, + 0.005103304982185364, + 0.006850087083876133, + 0.0030987204518169165, + 0.003523429622873664, + 0.011045442894101143, + 0.004502431023865938, + 0.045736901462078094, + 0.0023766246158629656, + 0.0049110157415270805, + 0.001945005962625146, + 0.004271951038390398, + 0.0028782570734620094, + 0.00327291339635849, + 0.004103685263544321, + 0.003521841252222657, + 0.02083667367696762, + 0.009542846120893955, + 0.02367468923330307, + 0.005871944595128298, + 0.005870897322893143, + 0.00714966980740428, + 0.005003577563911676, + 0.017604367807507515, + 0.014763424172997475, + 0.018773533403873444, + 0.07707498967647552, + 0.010537391528487206, + 0.012541352771222591, + 0.09214324504137039, + 0.009306411258876324, + 0.010814209468662739, + 0.00596938282251358, + 0.008419971913099289, + 0.00984933227300644, + 0.009581796824932098, + 0.00879942998290062, + 0.00952373631298542, + 0.01190772745758295, + 0.0053665516898036, + 0.008701835758984089, + 0.00795216578990221, + 0.007662329822778702, + 0.007742220535874367, + 0.0076768649742007256, + 0.015503454022109509, + 0.007126014679670334, + 0.007595517672598362, + 0.005879027768969536, + 0.008426487445831299, + 0.010095950216054916, + 0.010099091567099094, + 0.012338138185441494, + 0.012282651849091053, + 0.0077457306906580925, + 0.007080917712301016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.1564330756664276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020612085238099098 + ], + "opposite_logprob_tokens": [ + -3.8956120014190674 + ], + "logprob_reward": -0.020612085238099098, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020612085238099098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1530008316040039, + 0.08267339318990707, + 0.10482567548751831, + 0.11121820658445358, + 0.04808175191283226, + 0.11192268878221512, + 0.12591028213500977, + 0.09554131329059601, + 0.1087636947631836, + 0.05806218087673187 + ], + "credit_map": [ + 0.01214126218110323, + 0.03321123868227005, + 0.03317093104124069, + 0.013799792155623436, + 0.012951623648405075, + 0.014529183506965637, + 0.018183626234531403, + 0.015013184398412704, + 0.012264059856534004, + 0.012076213955879211, + 0.009495170786976814, + 0.008789725601673126, + 0.009180765599012375, + 0.012314246967434883, + 0.009945228695869446, + 0.008607982657849789, + 0.009160290472209454, + 0.009598570875823498, + 0.01263797003775835, + 0.015506666153669357, + 0.015215666964650154, + 0.01813787966966629, + 0.013051087036728859, + 0.01151754055172205, + 0.03476813808083534, + 0.006541880313307047, + 0.0054048532620072365, + 0.004198350943624973, + 0.003946519922465086, + 0.012016385793685913, + 0.004658772610127926, + 0.039683301001787186, + 0.004840191453695297, + 0.008288506418466568, + 0.0025786159094423056, + 0.007606722414493561, + 0.004265253897756338, + 0.004835750907659531, + 0.009238270111382008, + 0.0064284405671060085, + 0.02670256420969963, + 0.01397111639380455, + 0.017890674993395805, + 0.012210206128656864, + 0.010665660724043846, + 0.011122317984700203, + 0.009181937202811241, + 0.010178206488490105, + 0.006695480551570654, + 0.011913074180483818, + 0.03692929819226265, + 0.006436193827539682, + 0.008758896961808205, + 0.04376194626092911, + 0.005563552025705576, + 0.0058518340811133385, + 0.008016529493033886, + 0.011417482048273087, + 0.013694539666175842, + 0.012778433039784431, + 0.011410355567932129, + 0.013869515620172024, + 0.01662292331457138, + 0.007731529418379068, + 0.012786605395376682, + 0.022244146093726158, + 0.01187477633357048, + 0.012492415495216846, + 0.01398503314703703, + 0.014329980127513409, + 0.012540504336357117, + 0.008510230109095573, + 0.0038047372363507748, + 0.01403542049229145, + 0.005934190470725298, + 0.007158196065574884, + 0.007711910177022219, + 0.008106484077870846, + 0.006022730842232704, + 0.0052885073237121105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.020612085238099098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19704779982566833, + 0.07311955094337463, + 0.12357409298419952, + 0.04875098541378975, + 0.025153718888759613, + 0.03104608878493309, + 0.06998565047979355, + 0.06505636125802994, + 0.3214144706726074, + 0.04485123232007027 + ], + "credit_map": [ + 0.014423605985939503, + 0.04854869842529297, + 0.04687199369072914, + 0.01814669370651245, + 0.024171223863959312, + 0.015583968721330166, + 0.019149525091052055, + 0.010152093134820461, + 0.009776529856026173, + 0.009836899116635323, + 0.008895051665604115, + 0.01123314630240202, + 0.007821331731975079, + 0.00850716046988964, + 0.010318727232515812, + 0.006730700843036175, + 0.01286795362830162, + 0.013213815167546272, + 0.01375063881278038, + 0.017489975318312645, + 0.011421239003539085, + 0.024588411673903465, + 0.01572907343506813, + 0.014513001777231693, + 0.012742887251079082, + 0.0050403159111738205, + 0.0033286218531429768, + 0.002624200889840722, + 0.0029796152375638485, + 0.004684037528932095, + 0.003065934870392084, + 0.014285370707511902, + 0.00375782186165452, + 0.002916761441156268, + 0.0029313883278518915, + 0.0026953830383718014, + 0.0024569856468588114, + 0.003942567855119705, + 0.002462215255945921, + 0.003990597557276487, + 0.003934485837817192, + 0.0038768763188272715, + 0.008256658911705017, + 0.002192295389249921, + 0.00263479957357049, + 0.0023987246677279472, + 0.002455197973176837, + 0.005297053139656782, + 0.004636334255337715, + 0.0045797075144946575, + 0.01931704767048359, + 0.004904187750071287, + 0.00648884940892458, + 0.02151581645011902, + 0.004323561210185289, + 0.004220152273774147, + 0.01121350284665823, + 0.011562667787075043, + 0.011982094496488571, + 0.007170712575316429, + 0.004169515799731016, + 0.008837418630719185, + 0.004884112626314163, + 0.005236336495727301, + 0.03020455315709114, + 0.05422840267419815, + 0.04523482546210289, + 0.037727318704128265, + 0.03472883254289627, + 0.04176697880029678, + 0.05369911342859268, + 0.023824471980333328, + 0.003081011353060603, + 0.00777784176170826, + 0.0030438865069299936, + 0.003138434374704957, + 0.010241195559501648, + 0.007953832857310772, + 0.004248348064720631, + 0.005366684403270483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0067449514754116535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338066905736923 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338066905736923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338066905736923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2365565448999405, + 0.09993962198495865, + 0.14464637637138367, + 0.0987960547208786, + 0.02227967232465744, + 0.04324079677462578, + 0.05303189903497696, + 0.135905459523201, + 0.05204269289970398, + 0.11356092244386673 + ], + "credit_map": [ + 0.022456519305706024, + 0.03077547810971737, + 0.027078216895461082, + 0.026941418647766113, + 0.032333120703697205, + 0.031110616400837898, + 0.03557886555790901, + 0.030282311141490936, + 0.017010919749736786, + 0.016957813873887062, + 0.012622576206922531, + 0.012580309063196182, + 0.011269031092524529, + 0.010314318351447582, + 0.011585379019379616, + 0.00759928161278367, + 0.019641868770122528, + 0.01943177357316017, + 0.019171517342329025, + 0.020310677587985992, + 0.010832433588802814, + 0.02740945853292942, + 0.010792369022965431, + 0.017056262120604515, + 0.031684376299381256, + 0.004178897477686405, + 0.004366821143776178, + 0.0026714028790593147, + 0.0027895704843103886, + 0.010729657486081123, + 0.003516551572829485, + 0.03885877877473831, + 0.002874707104638219, + 0.00254692486487329, + 0.0015560878673568368, + 0.0023426746483892202, + 0.002195112407207489, + 0.003513433039188385, + 0.002421380253508687, + 0.004829352255910635, + 0.011025146581232548, + 0.004411216359585524, + 0.007638267241418362, + 0.0031017877627164125, + 0.004499351605772972, + 0.0037422887980937958, + 0.002945525338873267, + 0.005877213552594185, + 0.00456496886909008, + 0.006476671434938908, + 0.01053157914429903, + 0.004417335614562035, + 0.007843323983252048, + 0.008933637291193008, + 0.005846799351274967, + 0.004417583346366882, + 0.016370411962270737, + 0.019769813865423203, + 0.03597992658615112, + 0.011438291519880295, + 0.009078576229512691, + 0.021432295441627502, + 0.009702883660793304, + 0.012133253738284111, + 0.007151584606617689, + 0.009259103797376156, + 0.0050657037645578384, + 0.005384236574172974, + 0.00576534541323781, + 0.009353275410830975, + 0.005190120078623295, + 0.0048733241856098175, + 0.011345223523676395, + 0.010420823469758034, + 0.01263198722153902, + 0.011368501000106335, + 0.01682969182729721, + 0.023981638252735138, + 0.013763700611889362, + 0.013219351880252361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.026338066905736923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019413450499996543 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.0019413450499996543, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019413450499996543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25482526421546936, + 0.07575055211782455, + 0.10682118684053421, + 0.13566523790359497, + 0.058932699263095856, + 0.1262100785970688, + 0.0727405697107315, + 0.043989818543195724, + 0.04743284359574318, + 0.07763170450925827 + ], + "credit_map": [ + 0.01770438812673092, + 0.061276353895664215, + 0.05158110335469246, + 0.019157210364937782, + 0.02868146076798439, + 0.02327822893857956, + 0.028770286589860916, + 0.024376241490244865, + 0.01035238429903984, + 0.011562230065464973, + 0.007907221093773842, + 0.009651925414800644, + 0.010076357051730156, + 0.010452457703649998, + 0.009375464171171188, + 0.0063725183717906475, + 0.015381775796413422, + 0.013449201360344887, + 0.013015040196478367, + 0.011614173650741577, + 0.011970563791692257, + 0.019154276698827744, + 0.011346744373440742, + 0.010889421217143536, + 0.040694501250982285, + 0.007565799169242382, + 0.008539484813809395, + 0.004539744928479195, + 0.007308950647711754, + 0.014298657886683941, + 0.006609424017369747, + 0.046108681708574295, + 0.00820552371442318, + 0.005581755191087723, + 0.0073111397214233875, + 0.003965226002037525, + 0.004909110721200705, + 0.01162220723927021, + 0.004408933222293854, + 0.01292880903929472, + 0.015820130705833435, + 0.017138460651040077, + 0.026252776384353638, + 0.007766126189380884, + 0.015848545357584953, + 0.015126206912100315, + 0.009653170593082905, + 0.01860465668141842, + 0.0071168201975524426, + 0.010864261537790298, + 0.02037644200026989, + 0.005091957747936249, + 0.005228872876614332, + 0.01726326160132885, + 0.003027164377272129, + 0.003771798452362418, + 0.004214039538055658, + 0.005442346911877394, + 0.009166900999844074, + 0.005434515420347452, + 0.005035357549786568, + 0.005542812403291464, + 0.005278368014842272, + 0.0038754784036427736, + 0.0060455454513430595, + 0.007595663890242577, + 0.006981417071074247, + 0.004685927648097277, + 0.005385879892855883, + 0.006215377245098352, + 0.006610441487282515, + 0.003912597429007292, + 0.005618819035589695, + 0.008862132206559181, + 0.006363753229379654, + 0.007044522557407618, + 0.012630214914679527, + 0.018292931839823723, + 0.008732972666621208, + 0.010086365975439548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0019413450499996543 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, the camera parallel to a galloping horse, moving close to the ground to emphasize its thunderous speed. In the frame, a dusty, ancient scholar leans over his horse, his hair bun blown loose by the wind, leaving strands of black hair, mingled with sweat, plastered to his dusty cheeks. His eyes reveal an anxious yet resolute expression. He wears a robe of coarse cloth, once a bluish-gray color, now mottled with yellow dust. The wide cuffs and hem whip wildly backward in the wind. He grips the reins tightly, his body heaving and swaying with the rhythm of the muscular, sweat-drenched dark brown steed beneath him. Each swift and powerful thump of the horse's hooves on the dry, cracked, winding mountain path stirs up a thick cloud of yellow sand. Illuminated by the golden evening sun, this dust lurches like a swirling, luminous mist, trailing them closely behind. The background is rolling desolate mountains and sparse dead trees. The whole scene is dominated by warm yellow tones, creating a vast, urgent and epic atmosphere.", + "seed": 19364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_4579_seed_19364_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": -0.033721465966664255, + "vlm_logprob": -0.033721465966664255, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.033721465966664255, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05290672928094864 + ], + "opposite_logprob_tokens": [ + -3.8029067516326904 + ], + "logprob_reward": -0.05290672928094864, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.05290672928094864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07470029592514038, + 0.0459325835108757, + 0.43600380420684814, + 0.07483701407909393, + 0.06327306479215622, + 0.05232708901166916, + 0.07958284765481949, + 0.04833560436964035, + 0.05362329259514809, + 0.07138441503047943 + ], + "credit_map": [ + 0.006500605493783951, + 0.012967005372047424, + 0.008449552580714226, + 0.007539423182606697, + 0.012413278222084045, + 0.008036039769649506, + 0.007359829265624285, + 0.011434564366936684, + 0.005102128256112337, + 0.006798208225518465, + 0.004995420575141907, + 0.004881008993834257, + 0.005817210767418146, + 0.006299091037362814, + 0.004250402096658945, + 0.007789112627506256, + 0.04197492450475693, + 0.03352007269859314, + 0.09377853572368622, + 0.04076960310339928, + 0.02583961933851242, + 0.05792497843503952, + 0.028935914859175682, + 0.11326014995574951, + 0.010336413979530334, + 0.0075410460121929646, + 0.00892565120011568, + 0.007123406045138836, + 0.012562921270728111, + 0.011081354692578316, + 0.010255140252411366, + 0.007011082023382187, + 0.006831290666013956, + 0.0070854970254004, + 0.008449275977909565, + 0.00600194837898016, + 0.013538714498281479, + 0.006534682586789131, + 0.007315447553992271, + 0.007516208104789257, + 0.006813491228967905, + 0.009133602492511272, + 0.005326804239302874, + 0.00715223141014576, + 0.00902463961392641, + 0.006074514705687761, + 0.004692174959927797, + 0.004109628032892942, + 0.009421425871551037, + 0.009515673853456974, + 0.009834242053329945, + 0.010708384215831757, + 0.015803424641489983, + 0.008478709496557713, + 0.009678511880338192, + 0.006142475176602602, + 0.007150130812078714, + 0.007229495327919722, + 0.007121974602341652, + 0.004319292493164539, + 0.00566373486071825, + 0.005483913701027632, + 0.00527360662817955, + 0.0060934582725167274, + 0.006725449580699205, + 0.007737083360552788, + 0.009558477438986301, + 0.006503288168460131, + 0.005751009564846754, + 0.004554123152047396, + 0.007081981748342514, + 0.005711875855922699, + 0.007240904960781336, + 0.013561326079070568, + 0.015118497423827648, + 0.009418061934411526, + 0.005273402202874422, + 0.009591743350028992, + 0.005874783731997013, + 0.005305696744471788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.05290672928094864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026377543807029724 + ], + "opposite_logprob_tokens": [ + -3.6513774394989014 + ], + "logprob_reward": -0.026377543807029724, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026377543807029724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17936573922634125, + 0.09720201790332794, + 0.3046000003814697, + 0.0589730367064476, + 0.09977476298809052, + 0.02988159842789173, + 0.06575588136911392, + 0.035714246332645416, + 0.09631146490573883, + 0.032421279698610306 + ], + "credit_map": [ + 0.013859664089977741, + 0.022530270740389824, + 0.012635404244065285, + 0.025302991271018982, + 0.027045903727412224, + 0.031264811754226685, + 0.02153657376766205, + 0.025190124288201332, + 0.009010235778987408, + 0.016735266894102097, + 0.008707452565431595, + 0.009588845074176788, + 0.017673155292868614, + 0.01498049683868885, + 0.011003023013472557, + 0.00950353778898716, + 0.03077397681772709, + 0.024444937705993652, + 0.06611843407154083, + 0.025317246094346046, + 0.020498206838965416, + 0.038409944623708725, + 0.021890301257371902, + 0.07714695483446121, + 0.005917942617088556, + 0.005223249550908804, + 0.010153871029615402, + 0.00574179133400321, + 0.008686867542564869, + 0.007865887135267258, + 0.00838052574545145, + 0.0070029026828706264, + 0.010186076164245605, + 0.010092411190271378, + 0.01013258658349514, + 0.010115295648574829, + 0.02953208051621914, + 0.009430235251784325, + 0.011555087752640247, + 0.00873099360615015, + 0.003631396684795618, + 0.004908478353172541, + 0.003873149398714304, + 0.004173468332737684, + 0.004658597055822611, + 0.0034647027496248484, + 0.002645873697474599, + 0.0025259326212108135, + 0.006873271893709898, + 0.0061680348590016365, + 0.006737199611961842, + 0.010013382881879807, + 0.01651799865067005, + 0.00781305879354477, + 0.006699895020574331, + 0.004933041520416737, + 0.004714912734925747, + 0.005267081316560507, + 0.004228441044688225, + 0.004062957130372524, + 0.004958257544785738, + 0.00444812374189496, + 0.004842351656407118, + 0.003192119300365448, + 0.011466690339148045, + 0.008985190652310848, + 0.015600936487317085, + 0.014486719854176044, + 0.00923614390194416, + 0.011013192124664783, + 0.015245220623910427, + 0.010277366265654564, + 0.0030690981075167656, + 0.005516040604561567, + 0.006455396767705679, + 0.0038352569099515676, + 0.003267486346885562, + 0.004784427583217621, + 0.0031872738618403673, + 0.0023062978871166706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.026377543807029724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648558519780636 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648558519780636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648558519780636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26059645414352417, + 0.0672142505645752, + 0.06979575008153915, + 0.07399754971265793, + 0.13051994144916534, + 0.05205509066581726, + 0.09488388895988464, + 0.0691893994808197, + 0.14231710135936737, + 0.039430536329746246 + ], + "credit_map": [ + 0.02498001791536808, + 0.029829449951648712, + 0.019020073115825653, + 0.035755667835474014, + 0.035416025668382645, + 0.03441251814365387, + 0.028624102473258972, + 0.05255858972668648, + 0.007049440871924162, + 0.00873328372836113, + 0.007669959217309952, + 0.006239735521376133, + 0.008004548028111458, + 0.0073539866134524345, + 0.007214993238449097, + 0.01494830846786499, + 0.010283325798809528, + 0.00933375209569931, + 0.009892445988953114, + 0.0069321743212640285, + 0.009369092993438244, + 0.00737397838383913, + 0.007014003116637468, + 0.009596980176866055, + 0.009691390208899975, + 0.007269431836903095, + 0.011619295924901962, + 0.007037500850856304, + 0.013893330469727516, + 0.010059671476483345, + 0.007982814684510231, + 0.006444115191698074, + 0.018391747027635574, + 0.01843583956360817, + 0.019222883507609367, + 0.011145681142807007, + 0.024098437279462814, + 0.012344008311629295, + 0.012298312969505787, + 0.014583028852939606, + 0.007886764593422413, + 0.011805377900600433, + 0.007147941272705793, + 0.004401429556310177, + 0.007804193068295717, + 0.00579605670645833, + 0.003914074972271919, + 0.0032992518972605467, + 0.010530977509915829, + 0.008882224559783936, + 0.009762542322278023, + 0.012689606286585331, + 0.018167419359087944, + 0.01284974068403244, + 0.013153096660971642, + 0.0088482815772295, + 0.00859456043690443, + 0.00844151247292757, + 0.014173803851008415, + 0.0067542316392064095, + 0.006587450858205557, + 0.00775227602571249, + 0.008719866164028645, + 0.008165697567164898, + 0.01912602037191391, + 0.015758957713842392, + 0.02456672489643097, + 0.019409986212849617, + 0.013903685845434666, + 0.012509734369814396, + 0.02235572040081024, + 0.014686278998851776, + 0.005656367167830467, + 0.007081807591021061, + 0.006910049822181463, + 0.004659394733607769, + 0.004435169976204634, + 0.00398085406050086, + 0.003713278565555811, + 0.0029936134815216064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.008648558519780636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896102219820023 + ], + "opposite_logprob_tokens": [ + -2.578961133956909 + ], + "logprob_reward": -0.07896102219820023, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896102219820023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11311978101730347, + 0.06337936967611313, + 0.4384635090827942, + 0.05224495008587837, + 0.057982660830020905, + 0.021584410220384598, + 0.05653222277760506, + 0.02331400290131569, + 0.14524734020233154, + 0.028131769970059395 + ], + "credit_map": [ + 0.01079313363879919, + 0.011793280951678753, + 0.007770009338855743, + 0.008716858923435211, + 0.012761533260345459, + 0.02868589572608471, + 0.007881863974034786, + 0.02471720241010189, + 0.005591394379734993, + 0.007989655248820782, + 0.007355806417763233, + 0.006197106093168259, + 0.0064582256600260735, + 0.00917776394635439, + 0.005939299240708351, + 0.014670119620859623, + 0.0429081954061985, + 0.03182332217693329, + 0.09439477324485779, + 0.04054552689194679, + 0.025154652073979378, + 0.06129773333668709, + 0.030663372948765755, + 0.1116759330034256, + 0.010711354203522205, + 0.005097809713333845, + 0.0076787895523011684, + 0.0041025858372449875, + 0.006911374628543854, + 0.008403348736464977, + 0.003991901408880949, + 0.005347787868231535, + 0.011757888831198215, + 0.00995508674532175, + 0.008730093948543072, + 0.004782908130437136, + 0.00823110993951559, + 0.003937076777219772, + 0.004234866704791784, + 0.006353634409606457, + 0.0036972984671592712, + 0.004295913968235254, + 0.003746662987396121, + 0.0020946338772773743, + 0.0024863160215318203, + 0.0021808987949043512, + 0.0017865421250462532, + 0.0012961442116647959, + 0.007165576796978712, + 0.006527246907353401, + 0.007386309094727039, + 0.005373566877096891, + 0.007329155225306749, + 0.007026884704828262, + 0.0069519467651844025, + 0.008771536871790886, + 0.003526281798258424, + 0.004314087331295013, + 0.0024026986211538315, + 0.002214623847976327, + 0.0024672599975019693, + 0.0025036197621375322, + 0.0028015561401844025, + 0.0030838747043162584, + 0.023772811517119408, + 0.02682388760149479, + 0.029840076342225075, + 0.014715074561536312, + 0.01055071409791708, + 0.01615435630083084, + 0.00983023177832365, + 0.013560188934206963, + 0.0034145950339734554, + 0.0053763240575790405, + 0.004625354427844286, + 0.003768119728192687, + 0.0029099679086357355, + 0.002413395093753934, + 0.0033688670955598354, + 0.00225514592602849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.07896102219820023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017134760273620486 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017134760273620486, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017134760273620486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11009733378887177, + 0.11031181365251541, + 0.16419604420661926, + 0.056855637580156326, + 0.051244691014289856, + 0.2318115085363388, + 0.059759471565485, + 0.04298832640051842, + 0.07769275456666946, + 0.09504246711730957 + ], + "credit_map": [ + 0.010060098953545094, + 0.010541698895394802, + 0.010101547464728355, + 0.008801326155662537, + 0.01508798636496067, + 0.028696274384856224, + 0.01020081713795662, + 0.016607576981186867, + 0.011611873283982277, + 0.01790457032620907, + 0.013807657174766064, + 0.010428228415548801, + 0.014015852473676205, + 0.011250995099544525, + 0.009690998122096062, + 0.021601635962724686, + 0.016040222719311714, + 0.012202396057546139, + 0.03582364320755005, + 0.014560976065695286, + 0.009097517468035221, + 0.01945432834327221, + 0.010573646984994411, + 0.04644331708550453, + 0.007068383973091841, + 0.00650624418631196, + 0.007161192130297422, + 0.004723590798676014, + 0.007101892493665218, + 0.008877291344106197, + 0.0071085477247834206, + 0.008308494463562965, + 0.007209490053355694, + 0.004420094192028046, + 0.006030128803104162, + 0.005409606266766787, + 0.012376709841191769, + 0.004512281622737646, + 0.006254786625504494, + 0.005031593609601259, + 0.037905775010585785, + 0.045751359313726425, + 0.044436052441596985, + 0.021943487226963043, + 0.022851116955280304, + 0.024418706074357033, + 0.016211962327361107, + 0.018293051049113274, + 0.00977434404194355, + 0.006232887972146273, + 0.00802119355648756, + 0.004981265403330326, + 0.009405375458300114, + 0.006591834593564272, + 0.007476030383259058, + 0.007276542019098997, + 0.005856470670551062, + 0.0060618906281888485, + 0.008469930849969387, + 0.00488393846899271, + 0.0031200984958559275, + 0.004880665801465511, + 0.00563399400562048, + 0.004081339109688997, + 0.010710256174206734, + 0.015029145404696465, + 0.020969592034816742, + 0.009713700041174889, + 0.005223358981311321, + 0.004253843333572149, + 0.006395351141691208, + 0.005397503729909658, + 0.010501223616302013, + 0.01851160265505314, + 0.02097218669950962, + 0.015996649861335754, + 0.007397351786494255, + 0.00696120411157608, + 0.00748803187161684, + 0.007214211393147707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.0017134760273620486 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic low-angle tracking shot, the camera parallel to a galloping horse, moving close to the ground to emphasize its thunderous speed. In the frame, a dusty, ancient scholar leans over his horse, his hair bun blown loose by the wind, leaving strands of black hair, mingled with sweat, plastered to his dusty cheeks. His eyes reveal an anxious yet resolute expression. He wears a robe of coarse cloth, once a bluish-gray color, now mottled with yellow dust. The wide cuffs and hem whip wildly backward in the wind. He grips the reins tightly, his body heaving and swaying with the rhythm of the muscular, sweat-drenched dark brown steed beneath him. Each swift and powerful thump of the horse's hooves on the dry, cracked, winding mountain path stirs up a thick cloud of yellow sand. Illuminated by the golden evening sun, this dust lurches like a swirling, luminous mist, trailing them closely behind. The background is rolling desolate mountains and sparse dead trees. The whole scene is dominated by warm yellow tones, creating a vast, urgent and epic atmosphere.", + "seed": 19365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_8/index_4579_seed_19365_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": -0.032173782121390104, + "vlm_logprob": -0.032173782121390104, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.032173782121390104, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09517277777194977 + ], + "opposite_logprob_tokens": [ + -2.845172882080078 + ], + "logprob_reward": -0.09517277777194977, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09517277777194977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16478212177753448, + 0.12274790555238724, + 0.04010486975312233, + 0.05020296201109886, + 0.08903907239437103, + 0.13554663956165314, + 0.05976707115769386, + 0.142560675740242, + 0.12123704701662064, + 0.0740116685628891 + ], + "credit_map": [ + 0.01698279194533825, + 0.02602752111852169, + 0.029364891350269318, + 0.02599332295358181, + 0.012826626189053059, + 0.013129910454154015, + 0.019892239943146706, + 0.020564820617437363, + 0.011694961227476597, + 0.014379733242094517, + 0.014231383800506592, + 0.024922333657741547, + 0.013463178649544716, + 0.010464261285960674, + 0.01462310180068016, + 0.018968956544995308, + 0.0033506248146295547, + 0.0056948731653392315, + 0.005313064903020859, + 0.0069539425894618034, + 0.0050241765566170216, + 0.00388572015799582, + 0.00618326012045145, + 0.003699209075421095, + 0.006115914322435856, + 0.007029498927295208, + 0.009242109954357147, + 0.005572662688791752, + 0.004504283890128136, + 0.006983468774706125, + 0.005523884203284979, + 0.005231136921793222, + 0.01543663814663887, + 0.006662394385784864, + 0.017078205943107605, + 0.005607207305729389, + 0.012496797367930412, + 0.015363815240561962, + 0.005765597335994244, + 0.010628418065607548, + 0.018809529021382332, + 0.01681121252477169, + 0.018277378752827644, + 0.014760201796889305, + 0.016621457412838936, + 0.020238008350133896, + 0.012567280791699886, + 0.01746157743036747, + 0.007101394236087799, + 0.004799545742571354, + 0.008509911596775055, + 0.004239517264068127, + 0.009425181895494461, + 0.007486587390303612, + 0.00894833728671074, + 0.009256596677005291, + 0.013540636748075485, + 0.010095173493027687, + 0.010925006121397018, + 0.0075726741924881935, + 0.0496881902217865, + 0.026193775236606598, + 0.00902486126869917, + 0.0155203677713871, + 0.018679535016417503, + 0.012564638629555702, + 0.01198365818709135, + 0.012112786062061787, + 0.025843264535069466, + 0.014412845484912395, + 0.008514746092259884, + 0.017125561833381653, + 0.005752203520387411, + 0.009770415723323822, + 0.0088462820276618, + 0.008189842104911804, + 0.011421062983572483, + 0.008933006785809994, + 0.009689564816653728, + 0.011409291066229343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.09517277777194977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012570316903293133 + ], + "opposite_logprob_tokens": [ + -4.387570381164551 + ], + "logprob_reward": -0.012570316903293133, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012570316903293133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09546457231044769, + 0.07914449274539948, + 0.037075575441122055, + 0.036137137562036514, + 0.40513885021209717, + 0.042662620544433594, + 0.17273345589637756, + 0.054777104407548904, + 0.04901896044611931, + 0.02784721925854683 + ], + "credit_map": [ + 0.009842506609857082, + 0.013992576859891415, + 0.023388415575027466, + 0.017261704429984093, + 0.006915376987308264, + 0.010768232867121696, + 0.006585892289876938, + 0.006709868088364601, + 0.018460381776094437, + 0.015685245394706726, + 0.009070097468793392, + 0.010575952008366585, + 0.006459406577050686, + 0.00822358950972557, + 0.005448326934129, + 0.005221494007855654, + 0.003960987087339163, + 0.005327016580849886, + 0.007169421762228012, + 0.005160266533493996, + 0.0035013388842344284, + 0.00559627078473568, + 0.003433434758335352, + 0.002926839282736182, + 0.00411998713389039, + 0.00539124570786953, + 0.005677437409758568, + 0.005688000004738569, + 0.0037208653520792723, + 0.004221846349537373, + 0.003991410601884127, + 0.0033263463992625475, + 0.013008693233132362, + 0.010884481482207775, + 0.027056338265538216, + 0.01003768015652895, + 0.06564874202013016, + 0.17444279789924622, + 0.015546860173344612, + 0.08851328492164612, + 0.005347537342458963, + 0.006506723817437887, + 0.004279991611838341, + 0.005657540634274483, + 0.005601453594863415, + 0.006295482628047466, + 0.0036661485210061073, + 0.005307741463184357, + 0.01717832125723362, + 0.014304734766483307, + 0.03844229504466057, + 0.013337420299649239, + 0.03162245824933052, + 0.023662660270929337, + 0.016210174188017845, + 0.017975391820073128, + 0.006637302227318287, + 0.004896537866443396, + 0.004115172196179628, + 0.0032213390804827213, + 0.016325537115335464, + 0.009459088556468487, + 0.0036400356329977512, + 0.0064820898696780205, + 0.007288884837180376, + 0.00533952284604311, + 0.00773618696257472, + 0.005872772075235844, + 0.00813471432775259, + 0.005052817054092884, + 0.003365479875355959, + 0.006228580605238676, + 0.003082974348217249, + 0.00434823427349329, + 0.003113279351964593, + 0.0035410013515502214, + 0.003777329111471772, + 0.0038030201103538275, + 0.0033011878840625286, + 0.002880192594602704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.012570316903293133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550421990454197 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550421990454197, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550421990454197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5143257975578308, + 0.11197119951248169, + 0.056214477866888046, + 0.028143685311079025, + 0.05170443654060364, + 0.04184148088097572, + 0.0606546476483345, + 0.03958435356616974, + 0.06098148971796036, + 0.034578487277030945 + ], + "credit_map": [ + 0.05854266136884689, + 0.09876754134893417, + 0.13240933418273926, + 0.09551852941513062, + 0.028361238539218903, + 0.05168457329273224, + 0.019761312752962112, + 0.02928055264055729, + 0.02709737792611122, + 0.020541375502943993, + 0.014034518040716648, + 0.017400439828634262, + 0.007570499554276466, + 0.012285208329558372, + 0.005179848056286573, + 0.007861926220357418, + 0.005548703018575907, + 0.00734923779964447, + 0.010545003227889538, + 0.007295598275959492, + 0.005235313903540373, + 0.008730142377316952, + 0.006526913493871689, + 0.004983561113476753, + 0.0037067534867674112, + 0.004440564662218094, + 0.0043464903719723225, + 0.0032136801164597273, + 0.0029542038682848215, + 0.0040104384534060955, + 0.0029932851903140545, + 0.0024782666005194187, + 0.003475318429991603, + 0.00508387153968215, + 0.0042823548428714275, + 0.0033412270713597536, + 0.004696153104305267, + 0.015187973156571388, + 0.003607695922255516, + 0.012029839679598808, + 0.007410086691379547, + 0.005851224530488253, + 0.00517452135682106, + 0.005295212846249342, + 0.004076942801475525, + 0.004378563724458218, + 0.003313283436000347, + 0.006341641303151846, + 0.006112169474363327, + 0.009404975920915604, + 0.010715550743043423, + 0.006286171730607748, + 0.008882924914360046, + 0.007775550242513418, + 0.006869317963719368, + 0.004607985261827707, + 0.0031717573292553425, + 0.005762470420449972, + 0.004645447712391615, + 0.004636183846741915, + 0.00792898703366518, + 0.0053219981491565704, + 0.0046348609030246735, + 0.0034826439805328846, + 0.006145076360553503, + 0.01057187095284462, + 0.007536163087934256, + 0.006642797030508518, + 0.0107361925765872, + 0.006830341648310423, + 0.006159013602882624, + 0.0063600256107747555, + 0.003066138830035925, + 0.004515310749411583, + 0.0035681987646967173, + 0.0025496622547507286, + 0.0055028945207595825, + 0.007343919947743416, + 0.004025704693049192, + 0.0040066540241241455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012550421990454197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038085710257291794 + ], + "opposite_logprob_tokens": [ + -3.288085699081421 + ], + "logprob_reward": -0.038085710257291794, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038085710257291794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35471251606941223, + 0.15108509361743927, + 0.0899500846862793, + 0.10484303534030914, + 0.19118842482566833, + 0.019020039588212967, + 0.025896482169628143, + 0.02579524554312229, + 0.029412545263767242, + 0.008096582256257534 + ], + "credit_map": [ + 0.0434441938996315, + 0.05674158036708832, + 0.09296165406703949, + 0.06985339522361755, + 0.019322559237480164, + 0.03298851475119591, + 0.01666763238608837, + 0.022732974961400032, + 0.041736725717782974, + 0.018033966422080994, + 0.022749168798327446, + 0.021315110847353935, + 0.009492620825767517, + 0.019907508045434952, + 0.006919880863279104, + 0.010930120944976807, + 0.012463424354791641, + 0.007099587470293045, + 0.01891295425593853, + 0.008329278789460659, + 0.009495604783296585, + 0.016971006989479065, + 0.006315032951533794, + 0.010363196022808552, + 0.015587209723889828, + 0.01207375805824995, + 0.02071860246360302, + 0.009707451798021793, + 0.008915304206311703, + 0.01840616576373577, + 0.008385497145354748, + 0.011049051769077778, + 0.007229256443679333, + 0.003875581780448556, + 0.009445365518331528, + 0.002996305236592889, + 0.028869928792119026, + 0.08359842002391815, + 0.004447554238140583, + 0.05072600021958351, + 0.0024216584861278534, + 0.0018827741732820868, + 0.002284477697685361, + 0.0015232567675411701, + 0.0026120042894035578, + 0.0037813717499375343, + 0.0013704204466193914, + 0.0031440737657248974, + 0.0035609756596386433, + 0.002789916703477502, + 0.003504215506836772, + 0.0020229462534189224, + 0.003276749746873975, + 0.00293327821418643, + 0.004263998940587044, + 0.003544399980455637, + 0.001894856453873217, + 0.0034577599726617336, + 0.00339259160682559, + 0.003564596176147461, + 0.004029006697237492, + 0.003593899542465806, + 0.00335636711679399, + 0.002506167395040393, + 0.0027917518746107817, + 0.00425688037648797, + 0.003449531039223075, + 0.004227960016578436, + 0.004952878225594759, + 0.003259936347603798, + 0.002873597200959921, + 0.003600009484216571, + 0.0007057025213725865, + 0.001508863060735166, + 0.0010088951094076037, + 0.0010704947635531425, + 0.001295618829317391, + 0.0008013292099349201, + 0.0009273274918086827, + 0.0007783508626744151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.038085710257291794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle parallel to a galloping horse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a dusty ancient scholar leaning over a dark brown steed visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scholar wear a coarse bluish-gray robe mottled with yellow dust with wide cuffs whipping backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set on a dry cracked winding mountain path with rolling desolate mountains and sparse dead trees in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by golden evening sun creating a dominant warm yellow tone with a thick cloud of yellow sand?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248968368396163 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.00248968368396163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248968368396163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20046651363372803, + 0.08938659727573395, + 0.03982796519994736, + 0.04267878457903862, + 0.05883557349443436, + 0.292025625705719, + 0.05609963461756706, + 0.07966417074203491, + 0.04743694141507149, + 0.09357815980911255 + ], + "credit_map": [ + 0.02437056042253971, + 0.02508721314370632, + 0.034421879798173904, + 0.04009075462818146, + 0.017186667770147324, + 0.028908174484968185, + 0.01524330209940672, + 0.015157967805862427, + 0.011652668938040733, + 0.016707219183444977, + 0.010417410172522068, + 0.012676890008151531, + 0.009042741730809212, + 0.00994898658245802, + 0.009708626195788383, + 0.009232048876583576, + 0.004088329151272774, + 0.003983603324741125, + 0.006320658139884472, + 0.004988505970686674, + 0.00490690628066659, + 0.007263842038810253, + 0.004765546880662441, + 0.00351057481020689, + 0.0056961397640407085, + 0.007820880971848965, + 0.005247307009994984, + 0.006639513652771711, + 0.0034125130623579025, + 0.004909980576485395, + 0.00523771345615387, + 0.003714738879352808, + 0.003345621982589364, + 0.003843419486656785, + 0.0049368697218596935, + 0.003916170913726091, + 0.005028638523072004, + 0.017633618786931038, + 0.005114303901791573, + 0.015016933903098106, + 0.05490213632583618, + 0.024330349639058113, + 0.054085295647382736, + 0.018840137869119644, + 0.03274204209446907, + 0.036464814096689224, + 0.020393341779708862, + 0.050267528742551804, + 0.007184302434325218, + 0.004211174324154854, + 0.008564336225390434, + 0.0037251042667776346, + 0.01053432747721672, + 0.010090798139572144, + 0.004208628088235855, + 0.007580965291708708, + 0.005884477403014898, + 0.01010582409799099, + 0.0092234555631876, + 0.009844096377491951, + 0.015624488703906536, + 0.009815050289034843, + 0.010560741648077965, + 0.008606034331023693, + 0.00581256952136755, + 0.005143180023878813, + 0.00639085890725255, + 0.0071309455670416355, + 0.007402209099382162, + 0.005476178601384163, + 0.004865691065788269, + 0.0052153123542666435, + 0.007056080736219883, + 0.011319748125970364, + 0.008538463152945042, + 0.009443510323762894, + 0.017092430964112282, + 0.013737018220126629, + 0.015221331268548965, + 0.011169584468007088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.00248968368396163 + } + ], + "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/0000055/group_9/index_tensor([5327, 5327])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_tensor([5327, 5327])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9d479305169260f90fcf42ccbc02e5b6dcaa062e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_tensor([5327, 5327])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, side-on tracking shot, employing a shallow depth of field, locks focus on a snow-white nine-tailed fox as it swiftly streaks through a mysterious forest. The fox's fur, pure as snow, gleams softly in the dappled light filtering through the dense treetops. Its nine long, fluffy tails flow like a flowing silken cascade, gracefully undulating with its every movement. With a barely-there agility, it navigates the tangled, mossy trunks of ancient trees, its body low, showcasing the muscular tones hidden beneath its fur. The forest in the background is blurred into a dreamy expanse of dark brown bark and golden sunlight. A light, ethereal mist lingers over the woodland, leaving only a few fallen leaves, lifted by its four legs, clearly visible on the ground. The fox's sapphire eyes, sparkling with wisdom, are always looking intently ahead. The whole picture achieves a perfect balance between dynamic tracking and static focus, creating a mythical atmosphere that is both peaceful and full of vitality.", + "seed": 30549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_5327_seed_30549_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": -0.08635227894410491, + "vlm_logprob": -0.08635227894410491, + "vlm_yes_no_margin": 3.050000023841858, + "vlm_token_credit": -0.08635227894410491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1577252298593521 + ], + "opposite_logprob_tokens": [ + -2.0327253341674805 + ], + "logprob_reward": -0.1577252298593521, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1577252298593521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19028885662555695, + 0.10424395650625229, + 0.28023213148117065, + 0.08437187224626541, + 0.07315641641616821, + 0.0535975880920887, + 0.046327002346515656, + 0.05446726828813553, + 0.047970570623874664, + 0.06534429639577866 + ], + "credit_map": [ + 0.029474949464201927, + 0.02400488592684269, + 0.022016098722815514, + 0.022229643538594246, + 0.01788265071809292, + 0.032819416373968124, + 0.020774267613887787, + 0.021086934953927994, + 0.015394624322652817, + 0.009222371503710747, + 0.011247442103922367, + 0.016077280044555664, + 0.01123383454978466, + 0.010083384811878204, + 0.017060628160834312, + 0.01392439380288124, + 0.0762908086180687, + 0.023596016690135002, + 0.07249239087104797, + 0.03621499612927437, + 0.020039038732647896, + 0.01767551898956299, + 0.015941431745886803, + 0.017981931567192078, + 0.007195986807346344, + 0.015245030634105206, + 0.007053837180137634, + 0.013693294487893581, + 0.012910814024508, + 0.008343987166881561, + 0.010546335019171238, + 0.009382586926221848, + 0.018430812284350395, + 0.008179782889783382, + 0.008821878582239151, + 0.0065327961929142475, + 0.009020685218274593, + 0.005254496354609728, + 0.004383490886539221, + 0.012532470747828484, + 0.007588456384837627, + 0.005440650973469019, + 0.005196789745241404, + 0.008498307317495346, + 0.00965522788465023, + 0.006651035510003567, + 0.006708454806357622, + 0.0038586657028645277, + 0.0035778693854808807, + 0.005800279788672924, + 0.003751449752599001, + 0.004903547000139952, + 0.009955551475286484, + 0.004544426221400499, + 0.005846867803484201, + 0.007947012782096863, + 0.006744480691850185, + 0.006616177968680859, + 0.005853750742971897, + 0.005309861619025469, + 0.007069901097565889, + 0.008299868553876877, + 0.00514662079513073, + 0.009426607750356197, + 0.004464148543775082, + 0.0049108052626252174, + 0.005224006716161966, + 0.006790720392018557, + 0.007242646999657154, + 0.006221456918865442, + 0.007557984441518784, + 0.00555880181491375, + 0.012024454772472382, + 0.006962059531360865, + 0.008616038598120213, + 0.008755262941122055, + 0.009020349942147732, + 0.004542035516351461, + 0.008426363579928875, + 0.006997727323323488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.1577252298593521 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002359628677368 + ], + "opposite_logprob_tokens": [ + -2.695023536682129 + ], + "logprob_reward": -0.07002359628677368, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002359628677368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27617374062538147, + 0.0539693646132946, + 0.12329506874084473, + 0.09367026388645172, + 0.082277812063694, + 0.14285050332546234, + 0.02764388918876648, + 0.08734077215194702, + 0.07544425129890442, + 0.037334319204092026 + ], + "credit_map": [ + 0.020548557862639427, + 0.09160465002059937, + 0.017402226105332375, + 0.022254886105656624, + 0.01939997263252735, + 0.0246318019926548, + 0.013583232648670673, + 0.06674841046333313, + 0.006903514731675386, + 0.0053774574771523476, + 0.0056974212639033794, + 0.0076153636910021305, + 0.005302758421748877, + 0.005572727881371975, + 0.009507488459348679, + 0.007992631755769253, + 0.013755052350461483, + 0.011002746410667896, + 0.019731976091861725, + 0.016023807227611542, + 0.011853856965899467, + 0.01566227898001671, + 0.02121344581246376, + 0.014051908627152443, + 0.00969940610229969, + 0.023575227707624435, + 0.008088032715022564, + 0.014992384240031242, + 0.011565539985895157, + 0.005391642451286316, + 0.014106621965765953, + 0.006251412443816662, + 0.01731947995722294, + 0.009180489927530289, + 0.008774219080805779, + 0.008914032019674778, + 0.022465147078037262, + 0.004448851104825735, + 0.006829953286796808, + 0.004345638211816549, + 0.01375558041036129, + 0.016875838860869408, + 0.012130875140428543, + 0.023718593642115593, + 0.033420078456401825, + 0.01286819577217102, + 0.020028794184327126, + 0.010052536614239216, + 0.0023104806896299124, + 0.0027943700551986694, + 0.002640366554260254, + 0.003103364957496524, + 0.004368334077298641, + 0.0023803238291293383, + 0.004155918024480343, + 0.0058907330967485905, + 0.009686182253062725, + 0.011256073601543903, + 0.008897737599909306, + 0.008135298267006874, + 0.014574560336768627, + 0.01614033803343773, + 0.01056100893765688, + 0.008089571259915829, + 0.00784772913902998, + 0.008341846987605095, + 0.012266109697520733, + 0.011619338765740395, + 0.008548490703105927, + 0.0074272374622523785, + 0.010545319877564907, + 0.008848178200423717, + 0.005563225131481886, + 0.003929991740733385, + 0.0032688979990780354, + 0.0038679889403283596, + 0.006857714615762234, + 0.0035797718446701765, + 0.004780016839504242, + 0.005486710928380489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07002359628677368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005959717091172934 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.005959717091172934, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005959717091172934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25054553151130676, + 0.11257761716842651, + 0.2247263491153717, + 0.03366303816437721, + 0.16671837866306305, + 0.05198168382048607, + 0.03417545184493065, + 0.06774121522903442, + 0.0331430584192276, + 0.024727677926421165 + ], + "credit_map": [ + 0.025195011869072914, + 0.049382198601961136, + 0.020549004897475243, + 0.023963676765561104, + 0.024237751960754395, + 0.04793810844421387, + 0.02181522734463215, + 0.03746454045176506, + 0.015384227968752384, + 0.014116662554442883, + 0.012644756585359573, + 0.021651947870850563, + 0.008617158979177475, + 0.013220414519309998, + 0.017882976680994034, + 0.009059472009539604, + 0.0448995977640152, + 0.01906825602054596, + 0.04446066543459892, + 0.03347672149538994, + 0.017797231674194336, + 0.020225968211889267, + 0.029309118166565895, + 0.01548879686743021, + 0.003952356055378914, + 0.003981063142418861, + 0.005518922582268715, + 0.0039036960806697607, + 0.004087982699275017, + 0.003334950190037489, + 0.005355136934667826, + 0.0035289309453219175, + 0.04870924726128578, + 0.016257865354418755, + 0.012647608295083046, + 0.013445526361465454, + 0.04283905029296875, + 0.009886146523058414, + 0.008950355462729931, + 0.013982578180730343, + 0.0071297260001301765, + 0.006181242410093546, + 0.007558311801403761, + 0.009608373045921326, + 0.008991701528429985, + 0.002918855519965291, + 0.005683008581399918, + 0.003910464234650135, + 0.0021890930365771055, + 0.006544761825352907, + 0.002871505683287978, + 0.003670268692076206, + 0.006950175389647484, + 0.0039341081865131855, + 0.0038544880226254463, + 0.004161053337156773, + 0.006734558381140232, + 0.009704995900392532, + 0.005298297852277756, + 0.004916986916214228, + 0.014385069720447063, + 0.015769749879837036, + 0.004966482985764742, + 0.005965072195976973, + 0.0036419194657355547, + 0.004044413100928068, + 0.006074749864637852, + 0.005827943794429302, + 0.0039357515051960945, + 0.0030045730527490377, + 0.004276698920875788, + 0.0023370091803371906, + 0.0037433537654578686, + 0.0022756538819521666, + 0.001981897745281458, + 0.003139721928164363, + 0.0033735239412635565, + 0.003046807600185275, + 0.004166531842201948, + 0.003000189084559679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.005959717091172934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181779395788908 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.0181779395788908, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181779395788908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33849719166755676, + 0.060646843165159225, + 0.17474573850631714, + 0.029432130977511406, + 0.08399145305156708, + 0.14820215106010437, + 0.033366113901138306, + 0.022103967145085335, + 0.03114035353064537, + 0.0778740718960762 + ], + "credit_map": [ + 0.02960926666855812, + 0.11228705197572708, + 0.026713291183114052, + 0.023401876911520958, + 0.020932452753186226, + 0.02583734504878521, + 0.01871822401881218, + 0.08099766075611115, + 0.009453507140278816, + 0.008319676853716373, + 0.004519179929047823, + 0.0068559953942894936, + 0.008741626515984535, + 0.006023846101015806, + 0.006441130768507719, + 0.01029188185930252, + 0.032368388026952744, + 0.01588820107281208, + 0.03146909922361374, + 0.02548360452055931, + 0.015986226499080658, + 0.01449925359338522, + 0.02313736081123352, + 0.0159135889261961, + 0.003800946054980159, + 0.004605741240084171, + 0.0026555301155894995, + 0.003915128298103809, + 0.004360790364444256, + 0.00267395400442183, + 0.004265441559255123, + 0.0031545995734632015, + 0.01934143528342247, + 0.009053386747837067, + 0.00843786634504795, + 0.010693292133510113, + 0.02046165056526661, + 0.0053472998552024364, + 0.00663670152425766, + 0.004019825253635645, + 0.02313302457332611, + 0.016928192228078842, + 0.01603802852332592, + 0.017916541546583176, + 0.03497704491019249, + 0.012118170037865639, + 0.016676226630806923, + 0.010414915159344673, + 0.0024394001811742783, + 0.0038185729645192623, + 0.002293101279065013, + 0.003833264345303178, + 0.006491289474070072, + 0.0026463789399713278, + 0.004530762787908316, + 0.007313345558941364, + 0.0019636848010122776, + 0.0036106309853494167, + 0.002307623391970992, + 0.001894216169603169, + 0.004778072703629732, + 0.0025657396763563156, + 0.002094260649755597, + 0.0028897393494844437, + 0.0034690662287175655, + 0.003612056840211153, + 0.005039142444729805, + 0.0039401575922966, + 0.004472855944186449, + 0.00405157683417201, + 0.0028542380314320326, + 0.0037012617103755474, + 0.01526277419179678, + 0.007698021829128265, + 0.009400513023138046, + 0.011341570876538754, + 0.00902540609240532, + 0.005528399255126715, + 0.011144825257360935, + 0.008472553454339504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.0181779395788908 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987491190433502 + ], + "opposite_logprob_tokens": [ + -1.8048748970031738 + ], + "logprob_reward": -0.17987491190433502, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987491190433502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.556090235710144, + 0.0547587089240551, + 0.15217235684394836, + 0.025218157097697258, + 0.03924686089158058, + 0.04328310862183571, + 0.0160138588398695, + 0.06919614225625992, + 0.02276192419230938, + 0.021258743479847908 + ], + "credit_map": [ + 0.04998187720775604, + 0.15278691053390503, + 0.034927383065223694, + 0.03849174827337265, + 0.0503387451171875, + 0.09167128056287766, + 0.03186415508389473, + 0.10602808743715286, + 0.008770082145929337, + 0.006856141611933708, + 0.010522984899580479, + 0.006482341792434454, + 0.004455828573554754, + 0.004525104537606239, + 0.006911866832524538, + 0.006234357133507729, + 0.04777637869119644, + 0.012653902173042297, + 0.04448099434375763, + 0.01157008670270443, + 0.011652829125523567, + 0.0074897585436701775, + 0.010007036849856377, + 0.006541359703987837, + 0.0026521957479417324, + 0.002862646011635661, + 0.002745292615145445, + 0.003690398298203945, + 0.002825746778398752, + 0.0026917820796370506, + 0.00463911984115839, + 0.003110975492745638, + 0.009601387195289135, + 0.004155837465077639, + 0.004801665432751179, + 0.004311066120862961, + 0.007098449394106865, + 0.0022697898093611, + 0.003013517940416932, + 0.00399514427408576, + 0.010909742675721645, + 0.005260899662971497, + 0.004272564314305782, + 0.0037194073665887117, + 0.007246600463986397, + 0.00445855688303709, + 0.00450225779786706, + 0.0029130789916962385, + 0.0009836711687967181, + 0.002373295370489359, + 0.001362812938168645, + 0.002363331615924835, + 0.0027254815213382244, + 0.0016587723512202501, + 0.0025320262648165226, + 0.0020144653972238302, + 0.007018744945526123, + 0.009824022650718689, + 0.005873784888535738, + 0.005588034633547068, + 0.014091657474637032, + 0.01589818112552166, + 0.005643023177981377, + 0.005258689634501934, + 0.002247947035357356, + 0.002182312309741974, + 0.003315380308777094, + 0.0034610508009791374, + 0.0026579091791063547, + 0.0020896673668175936, + 0.004374338313937187, + 0.002433316083624959, + 0.0034606356639415026, + 0.002628809306770563, + 0.002230688463896513, + 0.002695315284654498, + 0.0036424100399017334, + 0.0019221509573981166, + 0.002941655460745096, + 0.0017370759742334485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.17987491190433502 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, employing a shallow depth of field, locks focus on a snow-white nine-tailed fox as it swiftly streaks through a mysterious forest. The fox's fur, pure as snow, gleams softly in the dappled light filtering through the dense treetops. Its nine long, fluffy tails flow like a flowing silken cascade, gracefully undulating with its every movement. With a barely-there agility, it navigates the tangled, mossy trunks of ancient trees, its body low, showcasing the muscular tones hidden beneath its fur. The forest in the background is blurred into a dreamy expanse of dark brown bark and golden sunlight. A light, ethereal mist lingers over the woodland, leaving only a few fallen leaves, lifted by its four legs, clearly visible on the ground. The fox's sapphire eyes, sparkling with wisdom, are always looking intently ahead. The whole picture achieves a perfect balance between dynamic tracking and static focus, creating a mythical atmosphere that is both peaceful and full of vitality.", + "seed": 30550, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_5327_seed_30550_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": -0.1131090335547924, + "vlm_logprob": -0.1131090335547924, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.1131090335547924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17826393246650696 + ], + "opposite_logprob_tokens": [ + -2.0532639026641846 + ], + "logprob_reward": -0.17826393246650696, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17826393246650696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1284540295600891, + 0.09307942539453506, + 0.18638630211353302, + 0.09500406682491302, + 0.1101088896393776, + 0.028157703578472137, + 0.11857718229293823, + 0.030096709728240967, + 0.1655769646167755, + 0.044558752328157425 + ], + "credit_map": [ + 0.01217524241656065, + 0.013992206193506718, + 0.010256162844598293, + 0.011892616748809814, + 0.02567240595817566, + 0.020720558241009712, + 0.026253053918480873, + 0.00749178184196353, + 0.01359455194324255, + 0.01711689680814743, + 0.01206622552126646, + 0.012700934894382954, + 0.010160066187381744, + 0.010496366769075394, + 0.010532524436712265, + 0.006411849055439234, + 0.03134030103683472, + 0.0369352325797081, + 0.017224077135324478, + 0.03025991842150688, + 0.015393734909594059, + 0.01764550432562828, + 0.017498256638646126, + 0.020089277997612953, + 0.0062125069089233875, + 0.007532920688390732, + 0.010480835102498531, + 0.01844046637415886, + 0.009575912728905678, + 0.009742502123117447, + 0.02217630296945572, + 0.010842615738511086, + 0.013289217837154865, + 0.02189045585691929, + 0.00964208971709013, + 0.013951152563095093, + 0.016329608857631683, + 0.012877898290753365, + 0.01051395758986473, + 0.011614503338932991, + 0.0032960891257971525, + 0.004810942802578211, + 0.003347195452079177, + 0.004937260877341032, + 0.0033096952829509974, + 0.0035967782605439425, + 0.0028792426455765963, + 0.0019804961048066616, + 0.011424761265516281, + 0.020313890650868416, + 0.012957962229847908, + 0.01995055004954338, + 0.01795228198170662, + 0.007861068472266197, + 0.01810196042060852, + 0.010014698840677738, + 0.002859643427655101, + 0.0034286868758499622, + 0.0024423468858003616, + 0.005563284270465374, + 0.004782242234796286, + 0.0038722578901797533, + 0.0036922215949743986, + 0.0034560246858745813, + 0.04911627620458603, + 0.0698351189494133, + 0.006356955040246248, + 0.010365927591919899, + 0.016780702397227287, + 0.003802446648478508, + 0.0065151602029800415, + 0.002804364776238799, + 0.00415768101811409, + 0.006677217781543732, + 0.006773253437131643, + 0.004274905193597078, + 0.008469752967357635, + 0.005612456239759922, + 0.004951356444507837, + 0.003642125753685832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.17826393246650696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205892935395241 + ], + "opposite_logprob_tokens": [ + -2.812058925628662 + ], + "logprob_reward": -0.06205892935395241, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205892935395241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08593959361314774, + 0.0729360356926918, + 0.08640178292989731, + 0.0765565037727356, + 0.35896769165992737, + 0.03242577984929085, + 0.07207411527633667, + 0.09743927419185638, + 0.03871005028486252, + 0.07854914665222168 + ], + "credit_map": [ + 0.009560943581163883, + 0.013034280389547348, + 0.008979950100183487, + 0.010863331146538258, + 0.01216378528624773, + 0.012293151579797268, + 0.013325151056051254, + 0.005719002336263657, + 0.01040207501500845, + 0.011593681760132313, + 0.008309861645102501, + 0.009189411997795105, + 0.007822993211448193, + 0.00797140970826149, + 0.009434333071112633, + 0.008212263695895672, + 0.008938421495258808, + 0.010957646183669567, + 0.010186350904405117, + 0.019694222137331963, + 0.007484221365302801, + 0.00856262631714344, + 0.011097102425992489, + 0.009481188841164112, + 0.004664191976189613, + 0.0052998741157352924, + 0.009682568721473217, + 0.012911717407405376, + 0.0077623347751796246, + 0.009516681544482708, + 0.01864878088235855, + 0.008070355281233788, + 0.03544564172625542, + 0.05558154731988907, + 0.0402095764875412, + 0.04905613139271736, + 0.055097587406635284, + 0.04136906936764717, + 0.04818630591034889, + 0.034021858125925064, + 0.0033808217849582434, + 0.0051188706420362, + 0.004388416651636362, + 0.005863862577825785, + 0.0037873934488743544, + 0.0030354615300893784, + 0.003351142629981041, + 0.003499808022752404, + 0.005755040794610977, + 0.006697853561490774, + 0.0065724803134799, + 0.014648856595158577, + 0.01347135379910469, + 0.0059309895150363445, + 0.011857722885906696, + 0.007139820139855146, + 0.008196012116968632, + 0.01165491621941328, + 0.009256321005523205, + 0.016474876552820206, + 0.017610501497983932, + 0.007094509433954954, + 0.012261562049388885, + 0.014890572056174278, + 0.009479504078626633, + 0.006766161881387234, + 0.006147510372102261, + 0.004023917019367218, + 0.005231280345469713, + 0.0018680708017200232, + 0.0035383158829063177, + 0.001655292697250843, + 0.009550001472234726, + 0.027128325775265694, + 0.006158586125820875, + 0.0040795146487653255, + 0.016383346170186996, + 0.006691697984933853, + 0.004523272160440683, + 0.004034406039863825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 2 + ], + "selected_reward": -0.06205892935395241 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330857142806053 + ], + "opposite_logprob_tokens": [ + -3.773308515548706 + ], + "logprob_reward": -0.02330857142806053, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330857142806053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09953945875167847, + 0.0554390512406826, + 0.43085187673568726, + 0.02959411032497883, + 0.08222899585962296, + 0.018008874729275703, + 0.03844057023525238, + 0.02268901653587818, + 0.13323546946048737, + 0.0899726152420044 + ], + "credit_map": [ + 0.015035217627882957, + 0.011656321585178375, + 0.013014920055866241, + 0.012933912687003613, + 0.01078160759061575, + 0.013924282044172287, + 0.014038396999239922, + 0.008154801093041897, + 0.010084742680191994, + 0.009811225347220898, + 0.007363444194197655, + 0.004696514457464218, + 0.005960446782410145, + 0.005903416778892279, + 0.00526092667132616, + 0.006358332000672817, + 0.05508654937148094, + 0.07402735203504562, + 0.035610828548669815, + 0.08415823429822922, + 0.041103534400463104, + 0.0401533879339695, + 0.050887320190668106, + 0.04982466995716095, + 0.002488541416823864, + 0.0028503623325377703, + 0.003916746936738491, + 0.006199799012392759, + 0.003214201657101512, + 0.0028475436847656965, + 0.005310223903506994, + 0.0027666909154504538, + 0.009184573777019978, + 0.009429114870727062, + 0.011106673628091812, + 0.014447944238781929, + 0.010370807722210884, + 0.006434520240873098, + 0.013642667792737484, + 0.007612691726535559, + 0.0021722381934523582, + 0.0030667323153465986, + 0.002083973027765751, + 0.00218691723421216, + 0.002371035050600767, + 0.0023354985751211643, + 0.0016867512604221702, + 0.002105728257447481, + 0.0032646723557263613, + 0.00699327839538455, + 0.004175154026597738, + 0.006457263138145208, + 0.005271011032164097, + 0.004124106839299202, + 0.0050927759148180485, + 0.0030623089987784624, + 0.002964907558634877, + 0.004014725796878338, + 0.0020812046714127064, + 0.003494692500680685, + 0.0031293719075620174, + 0.0019102670485153794, + 0.00290870713070035, + 0.0021851419005542994, + 0.03341493383049965, + 0.054364074021577835, + 0.006163803394883871, + 0.011457173153758049, + 0.015348139218986034, + 0.0029924700502306223, + 0.006879708729684353, + 0.00261516310274601, + 0.007106064353138208, + 0.01256199087947607, + 0.0122099369764328, + 0.013647586107254028, + 0.014712399803102016, + 0.007944883778691292, + 0.014587495476007462, + 0.007202259264886379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.02330857142806053 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057483419775963 + ], + "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": [ + 0.1235332041978836, + 0.05949671193957329, + 0.1639154702425003, + 0.1064743623137474, + 0.1915225088596344, + 0.029891787096858025, + 0.0440882109105587, + 0.05009187012910843, + 0.15627580881118774, + 0.07470997422933578 + ], + "credit_map": [ + 0.010320862755179405, + 0.015118122100830078, + 0.011915710754692554, + 0.014048853889107704, + 0.014725753106176853, + 0.021346738561987877, + 0.028121376410126686, + 0.007935812696814537, + 0.006830024067312479, + 0.0067702895030379295, + 0.006350800395011902, + 0.006369662005454302, + 0.007180007174611092, + 0.0058292001485824585, + 0.01168899703770876, + 0.008477740921080112, + 0.023021811619400978, + 0.019720731303095818, + 0.013820384629070759, + 0.028213605284690857, + 0.01660243049263954, + 0.024759093299508095, + 0.01936381310224533, + 0.018413610756397247, + 0.009923228994011879, + 0.018574554473161697, + 0.013446013443171978, + 0.017839934676885605, + 0.013132380321621895, + 0.010399354621767998, + 0.013214709237217903, + 0.00994420237839222, + 0.020106632262468338, + 0.02905661053955555, + 0.02160852774977684, + 0.025315944105386734, + 0.027436522766947746, + 0.021437080577015877, + 0.028195710852742195, + 0.018365496769547462, + 0.0035665149334818125, + 0.006042047403752804, + 0.0025320015847682953, + 0.0039758821949362755, + 0.0036302406806498766, + 0.0038602601271122694, + 0.003535953350365162, + 0.0027488942723721266, + 0.004063268192112446, + 0.008074723184108734, + 0.003957008011639118, + 0.00599252013489604, + 0.0076925796456635, + 0.003821343882009387, + 0.006607890594750643, + 0.0038788802921772003, + 0.008658797480165958, + 0.01138811744749546, + 0.003129037795588374, + 0.004505189135670662, + 0.011502129957079887, + 0.0034935269504785538, + 0.004197229165583849, + 0.0032178519759327173, + 0.051798079162836075, + 0.06258610635995865, + 0.008039917796850204, + 0.007103675045073032, + 0.012645498849451542, + 0.0038043835666030645, + 0.006735079921782017, + 0.0035630976781249046, + 0.007435246370732784, + 0.012510450556874275, + 0.012153164483606815, + 0.009820399805903435, + 0.010986042208969593, + 0.0061935423873364925, + 0.009728331118822098, + 0.005882808938622475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "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 shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813389003276825 + ], + "opposite_logprob_tokens": [ + -1.4063389301300049 + ], + "logprob_reward": -0.2813389003276825, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813389003276825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05842508375644684, + 0.048127204179763794, + 0.3871271312236786, + 0.05164255574345589, + 0.05674806982278824, + 0.016952505335211754, + 0.02642364241182804, + 0.043355248868465424, + 0.19552509486675262, + 0.11567357182502747 + ], + "credit_map": [ + 0.007216196041554213, + 0.006010915618389845, + 0.006949611008167267, + 0.007070338819175959, + 0.01042867824435234, + 0.008334921672940254, + 0.008716906420886517, + 0.003697509178891778, + 0.008412339724600315, + 0.008191053755581379, + 0.005959866568446159, + 0.005477862898260355, + 0.005324744153767824, + 0.004888302180916071, + 0.0060010626912117004, + 0.0038719656877219677, + 0.05713943392038345, + 0.0755632221698761, + 0.031303681433200836, + 0.07121645659208298, + 0.032284121960401535, + 0.03351946175098419, + 0.04098030924797058, + 0.045120418071746826, + 0.004296089988201857, + 0.005545740015804768, + 0.006374848075211048, + 0.010962366126477718, + 0.005126479081809521, + 0.004582596477121115, + 0.008952269330620766, + 0.005802164785563946, + 0.009624587371945381, + 0.006780208554118872, + 0.009727152064442635, + 0.006548216566443443, + 0.005142326932400465, + 0.00384696526452899, + 0.008789066225290298, + 0.006289543583989143, + 0.0021510720252990723, + 0.002629612572491169, + 0.002016103360801935, + 0.0036016327794641256, + 0.0018915997352451086, + 0.0014605120522901416, + 0.0016969892894849181, + 0.0015049827052280307, + 0.0027558307629078627, + 0.005869579967111349, + 0.0026208406779915094, + 0.003279230324551463, + 0.003975338768213987, + 0.0027316894847899675, + 0.003417275147512555, + 0.0017738540191203356, + 0.006093867123126984, + 0.009630375541746616, + 0.0034703961573541164, + 0.004781696014106274, + 0.010626543313264847, + 0.0022203498519957066, + 0.003993626683950424, + 0.0025383932515978813, + 0.05729211866855621, + 0.07916136831045151, + 0.00944145955145359, + 0.012447605840861797, + 0.01894989423453808, + 0.004566126503050327, + 0.008950206451117992, + 0.004716312512755394, + 0.009485490620136261, + 0.018162479624152184, + 0.01801944710314274, + 0.020839713513851166, + 0.014254127629101276, + 0.008166463114321232, + 0.01762268878519535, + 0.009123139083385468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.2813389003276825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, employing a shallow depth of field, locks focus on a snow-white nine-tailed fox as it swiftly streaks through a mysterious forest. The fox's fur, pure as snow, gleams softly in the dappled light filtering through the dense treetops. Its nine long, fluffy tails flow like a flowing silken cascade, gracefully undulating with its every movement. With a barely-there agility, it navigates the tangled, mossy trunks of ancient trees, its body low, showcasing the muscular tones hidden beneath its fur. The forest in the background is blurred into a dreamy expanse of dark brown bark and golden sunlight. A light, ethereal mist lingers over the woodland, leaving only a few fallen leaves, lifted by its four legs, clearly visible on the ground. The fox's sapphire eyes, sparkling with wisdom, are always looking intently ahead. The whole picture achieves a perfect balance between dynamic tracking and static focus, creating a mythical atmosphere that is both peaceful and full of vitality.", + "seed": 30551, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_5327_seed_30551_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": -0.1407409726642072, + "vlm_logprob": -0.1407409726642072, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1407409726642072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23206937313079834 + ], + "opposite_logprob_tokens": [ + -1.7320693731307983 + ], + "logprob_reward": -0.23206937313079834, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23206937313079834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2910727262496948, + 0.10090949386358261, + 0.041004568338394165, + 0.1061084121465683, + 0.09643841534852982, + 0.06619446724653244, + 0.09511494636535645, + 0.07055846601724625, + 0.065506212413311, + 0.06709225475788116 + ], + "credit_map": [ + 0.04624562710523605, + 0.02870274893939495, + 0.045524757355451584, + 0.05104779824614525, + 0.03879182040691376, + 0.026699941605329514, + 0.026319339871406555, + 0.027740711346268654, + 0.012246327474713326, + 0.012760575860738754, + 0.011242114938795567, + 0.022218579426407814, + 0.010967787355184555, + 0.013858712278306484, + 0.010382454842329025, + 0.007232943549752235, + 0.008809629827737808, + 0.0032339326571673155, + 0.004520237445831299, + 0.006849955301731825, + 0.004979032091796398, + 0.004734950140118599, + 0.004970755893737078, + 0.002906074747443199, + 0.02166893146932125, + 0.00879146158695221, + 0.01496891025453806, + 0.01016259379684925, + 0.012770693749189377, + 0.01084558479487896, + 0.009826896712183952, + 0.01707334630191326, + 0.015705665573477745, + 0.011358626186847687, + 0.00750333396717906, + 0.014759497717022896, + 0.009313267655670643, + 0.011002438142895699, + 0.012302063405513763, + 0.014493526890873909, + 0.006648941431194544, + 0.007861634716391563, + 0.0070308661088347435, + 0.00807974487543106, + 0.007652462460100651, + 0.007328338455408812, + 0.014383947476744652, + 0.007208535913378, + 0.010143469087779522, + 0.008718995377421379, + 0.00862976536154747, + 0.017410127446055412, + 0.010226820595562458, + 0.016745053231716156, + 0.014375390484929085, + 0.00886532012373209, + 0.0077606551349163055, + 0.006654000841081142, + 0.009714076295495033, + 0.011519823223352432, + 0.0053502279333770275, + 0.013064790517091751, + 0.009655674919486046, + 0.006839217618107796, + 0.010411357507109642, + 0.0061890664510428905, + 0.0070213801227509975, + 0.008517898619174957, + 0.008492791093885899, + 0.0077865393832325935, + 0.008336643688380718, + 0.008750529028475285, + 0.00850098766386509, + 0.005101823713630438, + 0.012422725558280945, + 0.011050957255065441, + 0.004032024182379246, + 0.0063375928439199924, + 0.007582946214824915, + 0.012063193134963512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.23206937313079834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559034824371338 + ], + "opposite_logprob_tokens": [ + -1.6005903482437134 + ], + "logprob_reward": -0.22559034824371338, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559034824371338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12448868900537491, + 0.07543623447418213, + 0.043643318116664886, + 0.18101325631141663, + 0.05875680595636368, + 0.059689465910196304, + 0.1068439930677414, + 0.1309114545583725, + 0.12556487321853638, + 0.09365198761224747 + ], + "credit_map": [ + 0.014028127305209637, + 0.013660511001944542, + 0.012195544317364693, + 0.020351940765976906, + 0.015329385176301003, + 0.01611104980111122, + 0.015320426784455776, + 0.01749170385301113, + 0.00894026830792427, + 0.007297540083527565, + 0.012291115708649158, + 0.014962974935770035, + 0.00823830347508192, + 0.010488363914191723, + 0.006532394327223301, + 0.006685267202556133, + 0.006910576485097408, + 0.00358911557123065, + 0.004049764946103096, + 0.006385089363902807, + 0.005871607456356287, + 0.005564635153859854, + 0.005307409912347794, + 0.005965118296444416, + 0.0460570864379406, + 0.010538428090512753, + 0.018948743119835854, + 0.01748356968164444, + 0.017547721043229103, + 0.016807474195957184, + 0.01823810301721096, + 0.03539213165640831, + 0.00791909545660019, + 0.010911575518548489, + 0.005170560907572508, + 0.006752484943717718, + 0.009516272693872452, + 0.007699728477746248, + 0.0058242627419531345, + 0.004962828475981951, + 0.005482816603034735, + 0.005870021414011717, + 0.006535903550684452, + 0.009101210162043571, + 0.005701170768588781, + 0.007749109994620085, + 0.012526565231382847, + 0.006722667254507542, + 0.008559061214327812, + 0.008455488830804825, + 0.009075336158275604, + 0.022853149101138115, + 0.010205800645053387, + 0.0139230415225029, + 0.018887657672166824, + 0.014884448610246181, + 0.017691422253847122, + 0.011455163359642029, + 0.01237463764846325, + 0.021834004670381546, + 0.009915434755384922, + 0.023685365915298462, + 0.019661525264382362, + 0.01429389975965023, + 0.016574718058109283, + 0.014997031539678574, + 0.013840200379490852, + 0.018908921629190445, + 0.01818620041012764, + 0.014146876521408558, + 0.015539877116680145, + 0.013371048495173454, + 0.009850619360804558, + 0.006465441547334194, + 0.014252257533371449, + 0.01373259723186493, + 0.004696334712207317, + 0.020197037607431412, + 0.009623503312468529, + 0.014834203757345676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.22559034824371338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004116158466786146 + ], + "opposite_logprob_tokens": [ + -5.504116058349609 + ], + "logprob_reward": -0.004116158466786146, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004116158466786146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13849341869354248, + 0.10013122111558914, + 0.04921507090330124, + 0.265755832195282, + 0.07521674782037735, + 0.05763482302427292, + 0.09932124614715576, + 0.04634133353829384, + 0.08057209849357605, + 0.08731826394796371 + ], + "credit_map": [ + 0.01690675877034664, + 0.012886571697890759, + 0.01681295596063137, + 0.012551574036478996, + 0.02452325075864792, + 0.022381111979484558, + 0.014085116796195507, + 0.018346084281802177, + 0.016348466277122498, + 0.013938426971435547, + 0.011888084933161736, + 0.018889404833316803, + 0.008056188002228737, + 0.012555675581097603, + 0.009517593309283257, + 0.008937381207942963, + 0.007952148094773293, + 0.005141060799360275, + 0.005751346237957478, + 0.008058684878051281, + 0.006001315079629421, + 0.005677505396306515, + 0.007570615038275719, + 0.003062392817810178, + 0.06870042532682419, + 0.0162363164126873, + 0.032865989953279495, + 0.022208580747246742, + 0.021090300753712654, + 0.025193477049469948, + 0.025466766208410263, + 0.05399394780397415, + 0.012312115170061588, + 0.011573933064937592, + 0.005268186330795288, + 0.009730760008096695, + 0.00934954546391964, + 0.005947019439190626, + 0.009099515154957771, + 0.011935669928789139, + 0.005824686028063297, + 0.007444953080266714, + 0.003367375349625945, + 0.007722879759967327, + 0.0065358723513782024, + 0.00804399885237217, + 0.013120930641889572, + 0.005574128590524197, + 0.009512027725577354, + 0.0077000451274216175, + 0.0072211772203445435, + 0.02068314328789711, + 0.008237272500991821, + 0.02047356590628624, + 0.013125724159181118, + 0.01236828975379467, + 0.007287723943591118, + 0.005096597597002983, + 0.004548162687569857, + 0.004519486799836159, + 0.0036319305654615164, + 0.008477164432406425, + 0.00859843660145998, + 0.004181829281151295, + 0.012047093361616135, + 0.0068888915702700615, + 0.009907212108373642, + 0.008643072098493576, + 0.01206253282725811, + 0.011428993195295334, + 0.006860945839434862, + 0.012733351439237595, + 0.008996250107884407, + 0.00744603481143713, + 0.008572407998144627, + 0.01699167862534523, + 0.006639766972512007, + 0.014106896705925465, + 0.010829240083694458, + 0.013735980726778507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004116158466786146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062032151967287064 + ], + "opposite_logprob_tokens": [ + -2.8120322227478027 + ], + "logprob_reward": -0.062032151967287064, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062032151967287064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0946258008480072, + 0.1104549691081047, + 0.06424012780189514, + 0.24364647269248962, + 0.11906745284795761, + 0.05839979648590088, + 0.03309614583849907, + 0.17217016220092773, + 0.06531533598899841, + 0.03898381441831589 + ], + "credit_map": [ + 0.011626724153757095, + 0.011088371276855469, + 0.009882514365017414, + 0.010529695078730583, + 0.011807105503976345, + 0.014477318152785301, + 0.012296262197196484, + 0.012917800806462765, + 0.011144296266138554, + 0.009366193786263466, + 0.018829720094799995, + 0.025986650958657265, + 0.009428740479052067, + 0.015583434142172337, + 0.009434092789888382, + 0.010681835003197193, + 0.0096824299544096, + 0.004303582012653351, + 0.008274651132524014, + 0.008575895801186562, + 0.007177598308771849, + 0.007727089338004589, + 0.006862175650894642, + 0.011636697687208652, + 0.06403519213199615, + 0.014441298320889473, + 0.026532161980867386, + 0.023121384903788567, + 0.02352539822459221, + 0.02057158760726452, + 0.022717196494340897, + 0.048702239990234375, + 0.015146980993449688, + 0.013636671006679535, + 0.01271540392190218, + 0.014732204377651215, + 0.01735181361436844, + 0.016687769442796707, + 0.014112725853919983, + 0.014683875255286694, + 0.006093002390116453, + 0.007820242084562778, + 0.004314770922064781, + 0.008080287836492062, + 0.005848478525876999, + 0.007436705753207207, + 0.012014428153634071, + 0.006791878025978804, + 0.004284728318452835, + 0.00365985045209527, + 0.004089694004505873, + 0.0073384735733270645, + 0.0027862032875418663, + 0.003005323698744178, + 0.004381401464343071, + 0.0035504656843841076, + 0.021584827452898026, + 0.014458496123552322, + 0.0158530380576849, + 0.021010763943195343, + 0.013492763042449951, + 0.03640298545360565, + 0.03044792264699936, + 0.01891934871673584, + 0.007354470435529947, + 0.00858602300286293, + 0.009815100580453873, + 0.008687588386237621, + 0.007060239091515541, + 0.008929714560508728, + 0.00770452618598938, + 0.007177672348916531, + 0.0036519502755254507, + 0.003221852006390691, + 0.005555408075451851, + 0.00752189802005887, + 0.0023758087772876024, + 0.006058278493583202, + 0.004819419700652361, + 0.005779195576906204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.062032151967287064 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989683151245117 + ], + "opposite_logprob_tokens": [ + -1.8048968315124512 + ], + "logprob_reward": -0.17989683151245117, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989683151245117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1303725391626358, + 0.1418914645910263, + 0.05293986573815346, + 0.19448359310626984, + 0.11525405198335648, + 0.08180289715528488, + 0.04639257863163948, + 0.08861586451530457, + 0.0646384134888649, + 0.0836087241768837 + ], + "credit_map": [ + 0.019246328622102737, + 0.01247026864439249, + 0.012669894844293594, + 0.017490845173597336, + 0.01538882590830326, + 0.021704036742448807, + 0.012349286116659641, + 0.019053053110837936, + 0.01792203262448311, + 0.017394522204995155, + 0.017107972875237465, + 0.01962529495358467, + 0.019287260249257088, + 0.020993752405047417, + 0.015687773004174232, + 0.013872850686311722, + 0.0073218876495957375, + 0.004284006077796221, + 0.004802323877811432, + 0.006240439135581255, + 0.007317299954593182, + 0.00650928495451808, + 0.012201703153550625, + 0.004262920003384352, + 0.02250264398753643, + 0.02695356495678425, + 0.025119150057435036, + 0.030473602935671806, + 0.03206545487046242, + 0.016540661454200745, + 0.024394363164901733, + 0.01643414981663227, + 0.02384784258902073, + 0.012246719561517239, + 0.008572949096560478, + 0.01233621221035719, + 0.012697258964180946, + 0.009036930277943611, + 0.011834691278636456, + 0.02468145079910755, + 0.008769767358899117, + 0.010471179150044918, + 0.00520001957193017, + 0.009522302076220512, + 0.009389868937432766, + 0.013860604725778103, + 0.0178682804107666, + 0.006720873527228832, + 0.005832102615386248, + 0.004335546400398016, + 0.003924788441509008, + 0.008789380080997944, + 0.0038430490531027317, + 0.00665659923106432, + 0.006051681470125914, + 0.006959435064345598, + 0.011384816840291023, + 0.00704082939773798, + 0.008316338993608952, + 0.0105913570150733, + 0.009270615875720978, + 0.020320052281022072, + 0.011707569472491741, + 0.00998428650200367, + 0.007503079250454903, + 0.005022619850933552, + 0.007187630049884319, + 0.00613176915794611, + 0.009945987723767757, + 0.010094531811773777, + 0.007022698409855366, + 0.011730100028216839, + 0.006428983528167009, + 0.006704372353851795, + 0.011017121374607086, + 0.009921698831021786, + 0.00442991778254509, + 0.022978777065873146, + 0.009688648395240307, + 0.012439200654625893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.17989683151245117 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, side-on tracking shot, employing a shallow depth of field, locks focus on a snow-white nine-tailed fox as it swiftly streaks through a mysterious forest. The fox's fur, pure as snow, gleams softly in the dappled light filtering through the dense treetops. Its nine long, fluffy tails flow like a flowing silken cascade, gracefully undulating with its every movement. With a barely-there agility, it navigates the tangled, mossy trunks of ancient trees, its body low, showcasing the muscular tones hidden beneath its fur. The forest in the background is blurred into a dreamy expanse of dark brown bark and golden sunlight. A light, ethereal mist lingers over the woodland, leaving only a few fallen leaves, lifted by its four legs, clearly visible on the ground. The fox's sapphire eyes, sparkling with wisdom, are always looking intently ahead. The whole picture achieves a perfect balance between dynamic tracking and static focus, creating a mythical atmosphere that is both peaceful and full of vitality.", + "seed": 30552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000055/group_9/index_5327_seed_30552_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": -0.14331647511571646, + "vlm_logprob": -0.14331647511571646, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.14331647511571646, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2009424865245819 + ], + "opposite_logprob_tokens": [ + -1.8259425163269043 + ], + "logprob_reward": -0.2009424865245819, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.2009424865245819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08578142523765564, + 0.03350185230374336, + 0.23521222174167633, + 0.05299832671880722, + 0.03267580270767212, + 0.11962210386991501, + 0.05023636668920517, + 0.12177202105522156, + 0.21444232761859894, + 0.05375760793685913 + ], + "credit_map": [ + 0.011372581124305725, + 0.016490353271365166, + 0.007609454449266195, + 0.012915581464767456, + 0.011151577346026897, + 0.009665724821388721, + 0.007629566825926304, + 0.008946584537625313, + 0.004010832868516445, + 0.003036960493773222, + 0.0030919918790459633, + 0.002644810127094388, + 0.004022134933620691, + 0.006584843620657921, + 0.00432115001603961, + 0.005789130926132202, + 0.015383505262434483, + 0.008776610717177391, + 0.015455433167517185, + 0.008197104558348656, + 0.019551603123545647, + 0.08916915208101273, + 0.014626017771661282, + 0.06405278295278549, + 0.010769096203148365, + 0.005978218279778957, + 0.007607476320117712, + 0.005914689041674137, + 0.006420481018722057, + 0.004028851166367531, + 0.00522684445604682, + 0.007052671629935503, + 0.0050546517595648766, + 0.0037573357112705708, + 0.004636099096387625, + 0.006141791585832834, + 0.003114585764706135, + 0.002958556404337287, + 0.0035266976337879896, + 0.003486083820462227, + 0.018221240490674973, + 0.0038164996076375246, + 0.04388181120157242, + 0.005776894744485617, + 0.0037775705568492413, + 0.008902491070330143, + 0.006697366945445538, + 0.028548231348395348, + 0.007171435747295618, + 0.004775764886289835, + 0.004405133426189423, + 0.006657053716480732, + 0.0061647240072488785, + 0.006226064171642065, + 0.006685483735054731, + 0.008150706999003887, + 0.016601892188191414, + 0.005013631656765938, + 0.007274700794368982, + 0.008666826412081718, + 0.006553362589329481, + 0.036412324756383896, + 0.01330249011516571, + 0.02794679068028927, + 0.05854000896215439, + 0.029062289744615555, + 0.014213123358786106, + 0.014785012230277061, + 0.007267965469509363, + 0.02199731580913067, + 0.018184548243880272, + 0.05039204657077789, + 0.004368689376860857, + 0.004086076747626066, + 0.011112445965409279, + 0.006740406155586243, + 0.004729466512799263, + 0.0076764593832194805, + 0.00755649246275425, + 0.007487570866942406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.2009424865245819 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032175481319427 + ], + "opposite_logprob_tokens": [ + -2.3503217697143555 + ], + "logprob_reward": -0.10032175481319427, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032175481319427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05026869475841522, + 0.06957918405532837, + 0.2212001383304596, + 0.07857704907655716, + 0.044615454971790314, + 0.08100777119398117, + 0.0670844316482544, + 0.06173659488558769, + 0.07937902957201004, + 0.24655164778232574 + ], + "credit_map": [ + 0.0066916365176439285, + 0.0052837892435491085, + 0.004360903985798359, + 0.009574760682880878, + 0.006141803227365017, + 0.007629537954926491, + 0.004544561263173819, + 0.006041697226464748, + 0.008010432124137878, + 0.01400226540863514, + 0.006907004863023758, + 0.006178905721753836, + 0.007227334193885326, + 0.010445738211274147, + 0.007845684885978699, + 0.008961819112300873, + 0.01173617783933878, + 0.009975883178412914, + 0.007597372401505709, + 0.010999955236911774, + 0.01800093613564968, + 0.07399262487888336, + 0.022481771185994148, + 0.06641540676355362, + 0.0137254036962986, + 0.012141905725002289, + 0.01313814427703619, + 0.008975760079920292, + 0.007682653144001961, + 0.007529539056122303, + 0.007725464645773172, + 0.007658185437321663, + 0.005928270053118467, + 0.004175195004791021, + 0.007223572116345167, + 0.009250902570784092, + 0.0031640471424907446, + 0.0034842146560549736, + 0.007102655712515116, + 0.004286602139472961, + 0.007853491231799126, + 0.007696507964283228, + 0.010856088250875473, + 0.01093717385083437, + 0.009383808821439743, + 0.00730226980522275, + 0.01687087118625641, + 0.010107563808560371, + 0.009227150119841099, + 0.006973847281187773, + 0.007139126770198345, + 0.006480047013610601, + 0.008133207447826862, + 0.012757363729178905, + 0.00957686547189951, + 0.0067968242801725864, + 0.004626050125807524, + 0.002228396013379097, + 0.009493416175246239, + 0.003344918368384242, + 0.00387230864726007, + 0.015168389305472374, + 0.005697974003851414, + 0.01730514131486416, + 0.009165880270302296, + 0.007800404913723469, + 0.01878432184457779, + 0.016432493925094604, + 0.0044713676907122135, + 0.0043446458876132965, + 0.011095813475549221, + 0.007284102030098438, + 0.019405363127589226, + 0.029039455577731133, + 0.06690800935029984, + 0.04724954441189766, + 0.016983946785330772, + 0.015085170045495033, + 0.036074407398700714, + 0.015805749222636223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 5 + ], + "selected_reward": -0.10032175481319427 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007645866833627224 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007645866833627224, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007645866833627224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14114020764827728, + 0.052696771919727325, + 0.27155041694641113, + 0.028707191348075867, + 0.09416715055704117, + 0.12824106216430664, + 0.03410988673567772, + 0.04090947285294533, + 0.09005707502365112, + 0.11842084676027298 + ], + "credit_map": [ + 0.014947045594453812, + 0.023658327758312225, + 0.013373038731515408, + 0.011049366556107998, + 0.016231441870331764, + 0.02484905906021595, + 0.014828597195446491, + 0.022203316912055016, + 0.006244974210858345, + 0.006363230291754007, + 0.005707757547497749, + 0.003772149793803692, + 0.0060692294500768185, + 0.011473579332232475, + 0.004339251201599836, + 0.00872659683227539, + 0.013922078534960747, + 0.008580628782510757, + 0.012846303172409534, + 0.008112309500575066, + 0.022324545308947563, + 0.10874238610267639, + 0.01225336268544197, + 0.08476878702640533, + 0.0034816667903214693, + 0.0029667105991393328, + 0.0027431328780949116, + 0.004919913597404957, + 0.0036652153357863426, + 0.003471773350611329, + 0.003196883946657181, + 0.004261896014213562, + 0.013122373260557652, + 0.009309583343565464, + 0.013534900732338428, + 0.01850471831858158, + 0.006458229385316372, + 0.007594837341457605, + 0.017062952741980553, + 0.008579553104937077, + 0.01380130648612976, + 0.00812921766191721, + 0.03425313159823418, + 0.010669367387890816, + 0.009789281524717808, + 0.009051155298948288, + 0.016569606959819794, + 0.02597799338400364, + 0.004393083043396473, + 0.003194187767803669, + 0.0032846047542989254, + 0.005452115088701248, + 0.004479255061596632, + 0.0028631307650357485, + 0.005590904038399458, + 0.004852605983614922, + 0.005609867163002491, + 0.001897593610920012, + 0.00998148787766695, + 0.004000294022262096, + 0.004302427638322115, + 0.003712412668392062, + 0.0048792497254908085, + 0.0065261381678283215, + 0.02475585788488388, + 0.012607380747795105, + 0.00863322988152504, + 0.006032126024365425, + 0.003590128617361188, + 0.008521235547959805, + 0.007102311588823795, + 0.018814796581864357, + 0.006716966163367033, + 0.009606984443962574, + 0.019242914393544197, + 0.018770968541502953, + 0.014067675918340683, + 0.01808556355535984, + 0.016375601291656494, + 0.015554171055555344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.007645866833627224 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057553455233574 + ], + "opposite_logprob_tokens": [ + -3.895575523376465 + ], + "logprob_reward": -0.02057553455233574, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057553455233574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04983619973063469, + 0.07000375539064407, + 0.3478196859359741, + 0.12031327933073044, + 0.05823206529021263, + 0.08327062427997589, + 0.09653368592262268, + 0.02452889084815979, + 0.09509751200675964, + 0.05436430871486664 + ], + "credit_map": [ + 0.0073608108796179295, + 0.0069551020860672, + 0.005989815108478069, + 0.0070119211450219154, + 0.005161253269761801, + 0.007365535944700241, + 0.004443714395165443, + 0.005548049230128527, + 0.008865587413311005, + 0.014466355554759502, + 0.007348197977989912, + 0.007272760383784771, + 0.007584905251860619, + 0.008886687457561493, + 0.005208895541727543, + 0.01037036906927824, + 0.017960507422685623, + 0.016460752114653587, + 0.01545721385627985, + 0.016848614439368248, + 0.031921159476041794, + 0.1158505380153656, + 0.03285958245396614, + 0.10046130418777466, + 0.0164957195520401, + 0.010079611092805862, + 0.013742679730057716, + 0.01169863622635603, + 0.017011750489473343, + 0.013316294178366661, + 0.0167130958288908, + 0.021255487576127052, + 0.007309057749807835, + 0.005092703737318516, + 0.00838862732052803, + 0.011967216618359089, + 0.0043286653235554695, + 0.006750664673745632, + 0.00870604906231165, + 0.005689079873263836, + 0.01106109656393528, + 0.0036614059936255217, + 0.025991735979914665, + 0.006017159204930067, + 0.006021541077643633, + 0.006180512253195047, + 0.0077629126608371735, + 0.016574259847402573, + 0.013384319841861725, + 0.01049510482698679, + 0.010981038212776184, + 0.010213088244199753, + 0.013193076476454735, + 0.015428552404046059, + 0.010430359281599522, + 0.012408144772052765, + 0.0029141264967620373, + 0.0017060830723494291, + 0.003986315336078405, + 0.002357427729293704, + 0.002772057196125388, + 0.0032555153593420982, + 0.002986893057823181, + 0.004550470504909754, + 0.006459036376327276, + 0.006502483505755663, + 0.02055867575109005, + 0.02136400155723095, + 0.005241699051111937, + 0.006650580558925867, + 0.012963488698005676, + 0.0153575474396348, + 0.004519758280366659, + 0.004610095173120499, + 0.011541849933564663, + 0.008615517988801003, + 0.006396122742444277, + 0.006858766544610262, + 0.007013408001512289, + 0.0048087891191244125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.02057553455233574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a side-on tracking shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the focus shallow with a snow-white nine-tailed fox as the main subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the fox moving swiftly through a forest with mossy trunks and fallen leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the fox's sapphire eyes visible and looking intently ahead?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709673285484314 + ], + "opposite_logprob_tokens": [ + -1.1370967626571655 + ], + "logprob_reward": -0.38709673285484314, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38709673285484314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1497984528541565, + 0.07646934688091278, + 0.250248521566391, + 0.06565036624670029, + 0.06541222333908081, + 0.09164095669984818, + 0.03960711508989334, + 0.10723161697387695, + 0.0345907099545002, + 0.11935068666934967 + ], + "credit_map": [ + 0.014159814454615116, + 0.024777382612228394, + 0.015913644805550575, + 0.01203053630888462, + 0.01611974835395813, + 0.02732226625084877, + 0.01599152944982052, + 0.023483531549572945, + 0.007845119573175907, + 0.013434743508696556, + 0.005964693147689104, + 0.008352086879312992, + 0.009155750274658203, + 0.010580097325146198, + 0.011414890177547932, + 0.009721959941089153, + 0.015277563594281673, + 0.008629963733255863, + 0.009327293373644352, + 0.009738696739077568, + 0.019207777455449104, + 0.09163659811019897, + 0.01812576688826084, + 0.07830483466386795, + 0.010706637986004353, + 0.00585040170699358, + 0.008311708457767963, + 0.006684121675789356, + 0.006640799809247255, + 0.005760802887380123, + 0.010641968809068203, + 0.011053930036723614, + 0.008605343289673328, + 0.008269320242106915, + 0.010400999337434769, + 0.008412090130150318, + 0.005580587778240442, + 0.0066019282676279545, + 0.010231988504529, + 0.007309966720640659, + 0.006397365126758814, + 0.010160163976252079, + 0.008438589982688427, + 0.013150576502084732, + 0.013748859986662865, + 0.010455337353050709, + 0.01891688071191311, + 0.010373184457421303, + 0.005667720455676317, + 0.005010993219912052, + 0.003227921901270747, + 0.004833488259464502, + 0.005657336674630642, + 0.004768838174641132, + 0.005387964192777872, + 0.005052851513028145, + 0.014951556921005249, + 0.004704990889877081, + 0.02422463335096836, + 0.008129451423883438, + 0.007877269759774208, + 0.01304281409829855, + 0.010714796371757984, + 0.023586109280586243, + 0.006357385776937008, + 0.005412499886006117, + 0.00522636016830802, + 0.003692283295094967, + 0.0032647056505084038, + 0.0022350687067955732, + 0.004888228606432676, + 0.0035141785629093647, + 0.008449542336165905, + 0.010336692444980145, + 0.0309921782463789, + 0.017476310953497887, + 0.011752559803426266, + 0.01300835981965065, + 0.01417908351868391, + 0.01315595954656601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.38709673285484314 + } + ], + "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/0000125/group_12/index_tensor([1914, 1914])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_tensor([1914, 1914])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..641cfaaecdc22ef0d79e6f9a9ba0ed16847e3e26 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_tensor([1914, 1914])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, captured from an extremely low angle on the forest floor, slowly advances at an almost imperceptible pace, deep into a dark, damp woodland. Atop a thick layer of decaying, dark brown leaves, a gigantic centipede crouches silently, its head a disturbing crimson, a stark contrast to its countless segmented, glossy black body. Faint light filters through gaps in the dense forest canopy, casting a vivid, wet sheen on its glossy carapace, a physically rendered, remarkably realistic effect. Each curve and subtle texture of the carapace is clearly visible through the interplay of highlights and shadows, giving it the texture of finely polished obsidian. At the tip of its crimson head, a pair of slender tentacles trembles subtly, probing the air, while hundreds of sharp legs on its sides cling to the damp leaves, motionless and poised for action. The entire picture is pure and flawless, without any impurities or interference, except for a few leaves with condensation on them and twisted roots covered with moss, which together create a primitive, quiet yet potentially threatening atmosphere.", + "seed": 31883, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_1914_seed_31883_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": -0.18779494063928723, + "vlm_logprob": -0.18779494063928723, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.18779494063928723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6552778482437134 + ], + "opposite_logprob_tokens": [ + -0.7802778482437134 + ], + "logprob_reward": -0.6552778482437134, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6552778482437134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14779029786586761, + 0.39471083879470825, + 0.043713998049497604, + 0.0687745064496994, + 0.033388473093509674, + 0.03356891870498657, + 0.1196272075176239, + 0.060532793402671814, + 0.036398690193891525, + 0.06149424985051155 + ], + "credit_map": [ + 0.022152749821543694, + 0.012809732928872108, + 0.017977803945541382, + 0.021125132218003273, + 0.015239539556205273, + 0.02157474122941494, + 0.017464419826865196, + 0.019446169957518578, + 0.05082022398710251, + 0.03500404953956604, + 0.04632455110549927, + 0.05724278837442398, + 0.03846786543726921, + 0.0711909607052803, + 0.04333292692899704, + 0.05232745781540871, + 0.00559866102412343, + 0.0033558490686118603, + 0.00544313620775938, + 0.0058521488681435585, + 0.004537191241979599, + 0.0069680786691606045, + 0.00596485985442996, + 0.005994071252644062, + 0.007731062360107899, + 0.006755236070603132, + 0.009633170440793037, + 0.011538284830749035, + 0.005780131556093693, + 0.009649809449911118, + 0.008619524538516998, + 0.009067285805940628, + 0.004350217524915934, + 0.002904704073444009, + 0.004141627345234156, + 0.005381738301366568, + 0.0031352131627500057, + 0.0051797316409647465, + 0.0036270408891141415, + 0.0046681989915668964, + 0.0030823773704469204, + 0.0026189517229795456, + 0.004521238151937723, + 0.004481549374759197, + 0.0029131791088730097, + 0.008030416443943977, + 0.0027379693929105997, + 0.005183237139135599, + 0.009430778212845325, + 0.012262328527867794, + 0.01728958822786808, + 0.013804366812109947, + 0.011038865894079208, + 0.028197932988405228, + 0.01150955818593502, + 0.016093799844384193, + 0.005854565184563398, + 0.004683993756771088, + 0.005179724656045437, + 0.013441329821944237, + 0.00468978937715292, + 0.010656235739588737, + 0.007444408256560564, + 0.008582744747400284, + 0.0036189535167068243, + 0.0021949068177491426, + 0.0069550820626318455, + 0.005502649582922459, + 0.003987445030361414, + 0.005223663989454508, + 0.004075593315064907, + 0.004840394482016563, + 0.006132316775619984, + 0.008370301686227322, + 0.009748450480401516, + 0.011176485568284988, + 0.0046937065199017525, + 0.007315365597605705, + 0.006018450018018484, + 0.008039169013500214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.6552778482437134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993785440921783 + ], + "opposite_logprob_tokens": [ + -1.8049378395080566 + ], + "logprob_reward": -0.17993785440921783, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993785440921783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2690310478210449, + 0.14905314147472382, + 0.08434339612722397, + 0.042259637266397476, + 0.04435350000858307, + 0.07486367970705032, + 0.08713967353105545, + 0.08816065639257431, + 0.07417885959148407, + 0.08661641925573349 + ], + "credit_map": [ + 0.044379688799381256, + 0.01788000948727131, + 0.03814892843365669, + 0.029259277507662773, + 0.02105686068534851, + 0.043924156576395035, + 0.03750914707779884, + 0.03687295690178871, + 0.022507525980472565, + 0.011869086883962154, + 0.01677233912050724, + 0.02878602407872677, + 0.010611006058752537, + 0.01607813686132431, + 0.02412131242454052, + 0.018307693302631378, + 0.010971280746161938, + 0.006898961961269379, + 0.009010224603116512, + 0.012757950462400913, + 0.0067580873146653175, + 0.018145715817809105, + 0.0077008819207549095, + 0.01210029050707817, + 0.005086065270006657, + 0.0035572259221225977, + 0.009139962494373322, + 0.004795888904482126, + 0.003611331107094884, + 0.006148464512079954, + 0.0038312601391226053, + 0.006089435890316963, + 0.004981846082955599, + 0.002669449895620346, + 0.005917244590818882, + 0.007441085297614336, + 0.004123375751078129, + 0.008430583402514458, + 0.0037930291146039963, + 0.00699688121676445, + 0.007244784850627184, + 0.006489359773695469, + 0.01068933866918087, + 0.008688285015523434, + 0.005644151940941811, + 0.01821405626833439, + 0.006539450027048588, + 0.011354251764714718, + 0.009676187299191952, + 0.008725341409444809, + 0.00593083631247282, + 0.007339449133723974, + 0.01612680032849312, + 0.01864522323012352, + 0.009044509381055832, + 0.011651312001049519, + 0.007641180418431759, + 0.007019710261374712, + 0.01178774330765009, + 0.01473646704107523, + 0.009066409431397915, + 0.014104287140071392, + 0.011570247821509838, + 0.012234603054821491, + 0.006298617925494909, + 0.00664159283041954, + 0.008007846772670746, + 0.009557615965604782, + 0.013911569491028786, + 0.013017597608268261, + 0.00724300229921937, + 0.009501018561422825, + 0.007350234314799309, + 0.008129527792334557, + 0.013096501119434834, + 0.020983662456274033, + 0.005806435830891132, + 0.010881707072257996, + 0.00798659585416317, + 0.012381752952933311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.17993785440921783 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078632436692715 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078632436692715, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078632436692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10672564059495926, + 0.3002334535121918, + 0.10308389365673065, + 0.0674913302063942, + 0.05891633406281471, + 0.15001587569713593, + 0.061794232577085495, + 0.057367488741874695, + 0.03659163787961006, + 0.05778014287352562 + ], + "credit_map": [ + 0.017667684704065323, + 0.008297346532344818, + 0.013159395195543766, + 0.01290938165038824, + 0.008155412040650845, + 0.016378264874219894, + 0.016225559636950493, + 0.013932585716247559, + 0.041760917752981186, + 0.020821645855903625, + 0.028337744995951653, + 0.06939183175563812, + 0.01560524757951498, + 0.03357881307601929, + 0.052492354065179825, + 0.03824487701058388, + 0.015850737690925598, + 0.00965341180562973, + 0.01245022565126419, + 0.014335010200738907, + 0.009331604465842247, + 0.017054911702871323, + 0.011982706375420094, + 0.012425275519490242, + 0.009004291146993637, + 0.007164216134697199, + 0.007321013603359461, + 0.015405724756419659, + 0.0052872877568006516, + 0.004720442462712526, + 0.008770515210926533, + 0.009817834943532944, + 0.006088223308324814, + 0.004418757278472185, + 0.009705216623842716, + 0.008962185122072697, + 0.005925643723458052, + 0.008159616962075233, + 0.00802096538245678, + 0.007635725196450949, + 0.015372077003121376, + 0.014621357433497906, + 0.01916447840631008, + 0.023219766095280647, + 0.01096003595739603, + 0.029808660969138145, + 0.01561032049357891, + 0.021259170025587082, + 0.0044389222748577595, + 0.0037713118363171816, + 0.005215686280280352, + 0.006800951436161995, + 0.013697082176804543, + 0.012068802490830421, + 0.007710842881351709, + 0.008090629242360592, + 0.004996504634618759, + 0.005120509769767523, + 0.008889411576092243, + 0.008427801541984081, + 0.00698068318888545, + 0.008712134324014187, + 0.006900194101035595, + 0.00734024727717042, + 0.0028701401315629482, + 0.003421108005568385, + 0.005925996694713831, + 0.0042427158914506435, + 0.005090989638119936, + 0.0058194478042423725, + 0.004532462451606989, + 0.004688777029514313, + 0.00803010817617178, + 0.00709156459197402, + 0.0072034806944429874, + 0.009699191898107529, + 0.004509421065449715, + 0.006025936920195818, + 0.007970290258526802, + 0.007250145543366671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.011078632436692715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906640112400055 + ], + "opposite_logprob_tokens": [ + -2.464066505432129 + ], + "logprob_reward": -0.08906640112400055, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906640112400055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12460964918136597, + 0.1386564075946808, + 0.07697395235300064, + 0.08067238330841064, + 0.08838468790054321, + 0.11179915815591812, + 0.07494889199733734, + 0.11240366101264954, + 0.04255681484937668, + 0.14899443089962006 + ], + "credit_map": [ + 0.020802266895771027, + 0.010796453803777695, + 0.015747742727398872, + 0.01498688105493784, + 0.011495181359350681, + 0.01693585142493248, + 0.017110906541347504, + 0.01673436351120472, + 0.0168283823877573, + 0.012290413491427898, + 0.018308255821466446, + 0.02464723400771618, + 0.01080317422747612, + 0.018502308055758476, + 0.018581939861178398, + 0.018694687634706497, + 0.007720543071627617, + 0.007622642442584038, + 0.008206808008253574, + 0.01133638247847557, + 0.010110240429639816, + 0.01304907537996769, + 0.00949539989233017, + 0.00943286344408989, + 0.008027156814932823, + 0.0070328121073544025, + 0.015339843928813934, + 0.015063333325088024, + 0.006126961205154657, + 0.010846797376871109, + 0.007739882916212082, + 0.010495598427951336, + 0.009284566156566143, + 0.006877067964524031, + 0.014144545421004295, + 0.016854282468557358, + 0.00781357754021883, + 0.012199626304209232, + 0.009413497522473335, + 0.011797522194683552, + 0.011040052399039268, + 0.009569703601300716, + 0.012518422678112984, + 0.018551191315054893, + 0.009496146813035011, + 0.02242879569530487, + 0.011413579806685448, + 0.016781266778707504, + 0.008199358358979225, + 0.005417301319539547, + 0.008307789452373981, + 0.009113956242799759, + 0.00914295669645071, + 0.015405155718326569, + 0.009737320244312286, + 0.009625052101910114, + 0.013549713417887688, + 0.00873098149895668, + 0.014284823089838028, + 0.018609682098031044, + 0.012923317961394787, + 0.01616539992392063, + 0.012544187717139721, + 0.015595552511513233, + 0.004779337905347347, + 0.0036421797703951597, + 0.0042860242538154125, + 0.008676458150148392, + 0.0037573787849396467, + 0.0059859976172447205, + 0.005839402787387371, + 0.005590034183114767, + 0.013443686068058014, + 0.017734207212924957, + 0.02721388451755047, + 0.02871588058769703, + 0.010884560644626617, + 0.018486851826310158, + 0.0138529809191823, + 0.018662387505173683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.08906640112400055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19850589334964752, + 0.1702023446559906, + 0.15845263004302979, + 0.09007936716079712, + 0.03388746455311775, + 0.1073242574930191, + 0.05992450192570686, + 0.03721572086215019, + 0.06906051188707352, + 0.07534736394882202 + ], + "credit_map": [ + 0.028297238051891327, + 0.011511805467307568, + 0.023485876619815826, + 0.028710730373859406, + 0.01577632501721382, + 0.034516505897045135, + 0.02984989620745182, + 0.026357509195804596, + 0.02300216443836689, + 0.010986877605319023, + 0.020361747592687607, + 0.017298787832260132, + 0.012879962101578712, + 0.04085557535290718, + 0.02003907971084118, + 0.02477813884615898, + 0.01800510287284851, + 0.012887530028820038, + 0.02660755068063736, + 0.02076217345893383, + 0.017162756994366646, + 0.02918708696961403, + 0.014677418395876884, + 0.019163010641932487, + 0.010251814499497414, + 0.007959521375596523, + 0.013460194692015648, + 0.011015498079359531, + 0.009902594611048698, + 0.015472115017473698, + 0.009851349517703056, + 0.012166278436779976, + 0.0035782046616077423, + 0.0022505237720906734, + 0.004682913888245821, + 0.004732912871986628, + 0.0036696987226605415, + 0.00648719584569335, + 0.0035318038426339626, + 0.004954208619892597, + 0.011416299268603325, + 0.009081336669623852, + 0.01459159329533577, + 0.016097256913781166, + 0.00715514225885272, + 0.023958703503012657, + 0.010204663500189781, + 0.01481926254928112, + 0.005247768480330706, + 0.003996023442596197, + 0.0065130009315907955, + 0.007770877331495285, + 0.0071842907927930355, + 0.014481241814792156, + 0.005517715122550726, + 0.009213579818606377, + 0.0028089075349271297, + 0.0037325751036405563, + 0.0029252723325043917, + 0.007949449121952057, + 0.004353365395218134, + 0.006205791141837835, + 0.004299495369195938, + 0.004940862767398357, + 0.004997866693884134, + 0.006193554028868675, + 0.00672406330704689, + 0.01180384773761034, + 0.012407826259732246, + 0.011324548162519932, + 0.006931984331458807, + 0.008676818571984768, + 0.007945834659039974, + 0.005929120350629091, + 0.011591927148401737, + 0.020137757062911987, + 0.00397837907075882, + 0.008071741089224815, + 0.006450807210057974, + 0.011241799220442772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0036139669828116894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from an extremely low angle on the forest floor, slowly advances at an almost imperceptible pace, deep into a dark, damp woodland. Atop a thick layer of decaying, dark brown leaves, a gigantic centipede crouches silently, its head a disturbing crimson, a stark contrast to its countless segmented, glossy black body. Faint light filters through gaps in the dense forest canopy, casting a vivid, wet sheen on its glossy carapace, a physically rendered, remarkably realistic effect. Each curve and subtle texture of the carapace is clearly visible through the interplay of highlights and shadows, giving it the texture of finely polished obsidian. At the tip of its crimson head, a pair of slender tentacles trembles subtly, probing the air, while hundreds of sharp legs on its sides cling to the damp leaves, motionless and poised for action. The entire picture is pure and flawless, without any impurities or interference, except for a few leaves with condensation on them and twisted roots covered with moss, which together create a primitive, quiet yet potentially threatening atmosphere.", + "seed": 31884, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_1914_seed_31884_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": -0.0573120285756886, + "vlm_logprob": -0.0573120285756886, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.0573120285756886, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05254926159977913 + ], + "opposite_logprob_tokens": [ + -3.677549362182617 + ], + "logprob_reward": -0.05254926159977913, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05254926159977913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22010736167430878, + 0.3074610233306885, + 0.12295535206794739, + 0.06965811550617218, + 0.0562334842979908, + 0.03483643755316734, + 0.032225243747234344, + 0.10979396849870682, + 0.02753208763897419, + 0.01919693313539028 + ], + "credit_map": [ + 0.03815169632434845, + 0.017088960856199265, + 0.04603554680943489, + 0.03337589651346207, + 0.021899964660406113, + 0.020906513556838036, + 0.027235127985477448, + 0.01541364099830389, + 0.10239402949810028, + 0.011011648923158646, + 0.12546026706695557, + 0.012619439512491226, + 0.007620002143085003, + 0.007908129133284092, + 0.03036371059715748, + 0.010083755478262901, + 0.015248037874698639, + 0.020103316754102707, + 0.013896283693611622, + 0.016399269923567772, + 0.010484742932021618, + 0.011658166535198689, + 0.015547450631856918, + 0.019618071615695953, + 0.008304796181619167, + 0.006751255132257938, + 0.009427456185221672, + 0.017323365435004234, + 0.006016144063323736, + 0.006144051440060139, + 0.00931040570139885, + 0.00638063158839941, + 0.00832584872841835, + 0.005397509783506393, + 0.008869373239576817, + 0.009340229444205761, + 0.004269108176231384, + 0.006268850062042475, + 0.007472383324056864, + 0.006290178745985031, + 0.0040113842114806175, + 0.004414395894855261, + 0.003964842297136784, + 0.003982263617217541, + 0.0031608238350600004, + 0.005922660231590271, + 0.0042676678858697414, + 0.005112397484481335, + 0.002988893073052168, + 0.001585975755006075, + 0.007486734073609114, + 0.0022792380768805742, + 0.0016236299416050315, + 0.0023045039270073175, + 0.011131847277283669, + 0.0028244201093912125, + 0.009744493290781975, + 0.006757604889571667, + 0.03209756687283516, + 0.007778100669384003, + 0.007460635621100664, + 0.006850019097328186, + 0.0074265459552407265, + 0.03167898952960968, + 0.004028883762657642, + 0.003389804158359766, + 0.0036564504262059927, + 0.0037821331061422825, + 0.002684694482013583, + 0.0040559363551437855, + 0.0028876354917883873, + 0.00304654729552567, + 0.004574926104396582, + 0.001246194588020444, + 0.0018397204112261534, + 0.0021887614857405424, + 0.0010014827130362391, + 0.0019077284960076213, + 0.0025973806623369455, + 0.0038407365791499615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.05254926159977913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034364461898804 + ], + "opposite_logprob_tokens": [ + -2.350343704223633 + ], + "logprob_reward": -0.10034364461898804, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034364461898804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06948037445545197, + 0.266754686832428, + 0.10737847536802292, + 0.06166365370154381, + 0.11179064214229584, + 0.09299756586551666, + 0.055465348064899445, + 0.1236412301659584, + 0.04894127696752548, + 0.06188671663403511 + ], + "credit_map": [ + 0.009731258265674114, + 0.007505891844630241, + 0.007829984650015831, + 0.01118987426161766, + 0.011483382433652878, + 0.0043928856030106544, + 0.010948586277663708, + 0.00639850739389658, + 0.08488448709249496, + 0.012905932031571865, + 0.10355361551046371, + 0.011673517525196075, + 0.008140062913298607, + 0.006911424919962883, + 0.02780740335583687, + 0.010878236964344978, + 0.020414194092154503, + 0.01378592662513256, + 0.022578435018658638, + 0.010776163078844547, + 0.00654190219938755, + 0.008952736854553223, + 0.009771314449608326, + 0.01455780491232872, + 0.006199600640684366, + 0.007287938613444567, + 0.008420967496931553, + 0.008139071986079216, + 0.008309498429298401, + 0.008864148519933224, + 0.008418296463787556, + 0.006024132017046213, + 0.013262346386909485, + 0.008067727088928223, + 0.02059805765748024, + 0.01718740351498127, + 0.01294340193271637, + 0.009715585969388485, + 0.021998532116413116, + 0.008017590269446373, + 0.010902101173996925, + 0.01053269300609827, + 0.01407573837786913, + 0.010623052716255188, + 0.007455533370375633, + 0.014247552491724491, + 0.010226787999272346, + 0.014934108592569828, + 0.00539257051423192, + 0.004396476317197084, + 0.011376905255019665, + 0.0048918393440544605, + 0.003879346651956439, + 0.0038376026786863804, + 0.01703518070280552, + 0.004655429162085056, + 0.010791264474391937, + 0.006255227141082287, + 0.03617869317531586, + 0.011788778938353062, + 0.006734298076480627, + 0.008221148513257504, + 0.008135978132486343, + 0.03553583472967148, + 0.00702498247846961, + 0.006280898116528988, + 0.006770672742277384, + 0.006518913898617029, + 0.004637029953300953, + 0.006527717225253582, + 0.005033411085605621, + 0.006147648207843304, + 0.017773721367120743, + 0.004370007198303938, + 0.005502532701939344, + 0.00734576303511858, + 0.0053002601489424706, + 0.006073559634387493, + 0.009388986974954605, + 0.006131886038929224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.10034364461898804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072474420070648 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072474420070648, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072474420070648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11487620323896408, + 0.07281198352575302, + 0.06709939986467361, + 0.10179357975721359, + 0.04029770940542221, + 0.0276198610663414, + 0.03405989706516266, + 0.24528950452804565, + 0.022812144830822945, + 0.27333971858024597 + ], + "credit_map": [ + 0.023942897096276283, + 0.013964409939944744, + 0.020620010793209076, + 0.013919460587203503, + 0.010574893094599247, + 0.011526250280439854, + 0.012854835949838161, + 0.007473451551049948, + 0.009531980380415916, + 0.0052126869559288025, + 0.009002811275422573, + 0.007238428108394146, + 0.01443510502576828, + 0.006148481275886297, + 0.0159145500510931, + 0.005327943246811628, + 0.010617664083838463, + 0.006745354738086462, + 0.012998262420296669, + 0.009217570535838604, + 0.006800634786486626, + 0.005916088353842497, + 0.00843600369989872, + 0.006367823109030724, + 0.009869711473584175, + 0.011042962782084942, + 0.01279304176568985, + 0.026829320937395096, + 0.008972035720944405, + 0.008288009092211723, + 0.01316371001303196, + 0.010834788903594017, + 0.004442598205059767, + 0.003917569760233164, + 0.005649676080793142, + 0.008102897554636002, + 0.002497514244168997, + 0.0036895542871207, + 0.008043793961405754, + 0.003954105544835329, + 0.003946247510612011, + 0.00368102197535336, + 0.0047363946214318275, + 0.003171120770275593, + 0.0019083346705883741, + 0.0032526191789656878, + 0.002710298402234912, + 0.004213823936879635, + 0.00342483539134264, + 0.002389047294855118, + 0.006368995178490877, + 0.0029895263724029064, + 0.002737876959145069, + 0.002227069577202201, + 0.010951650328934193, + 0.0029708961956202984, + 0.023771071806550026, + 0.012950758449733257, + 0.06704314798116684, + 0.028778724372386932, + 0.013264544308185577, + 0.016749199479818344, + 0.018372192978858948, + 0.06435985863208771, + 0.003249093657359481, + 0.0034354636445641518, + 0.003648186568170786, + 0.003035807516425848, + 0.002064495114609599, + 0.0023145002778619528, + 0.0024691945873200893, + 0.0025954036973416805, + 0.09399295598268509, + 0.012056524865329266, + 0.014898487366735935, + 0.022659212350845337, + 0.015165918506681919, + 0.01851339265704155, + 0.02639785036444664, + 0.06965536624193192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.016072474420070648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295779794454575 + ], + "opposite_logprob_tokens": [ + -2.237957715988159 + ], + "logprob_reward": -0.11295779794454575, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295779794454575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07373729348182678, + 0.12943029403686523, + 0.11232690513134003, + 0.07047946751117706, + 0.053563058376312256, + 0.09673746675252914, + 0.016848161816596985, + 0.300908625125885, + 0.07109647244215012, + 0.07487232238054276 + ], + "credit_map": [ + 0.008186805061995983, + 0.012336941435933113, + 0.008883786387741566, + 0.009442961774766445, + 0.007818123325705528, + 0.011123839765787125, + 0.0085065346211195, + 0.007438292261213064, + 0.020448343828320503, + 0.010373419150710106, + 0.01603620871901512, + 0.015758927911520004, + 0.022332070395350456, + 0.009714172221720219, + 0.02454739809036255, + 0.01021974254399538, + 0.02166336588561535, + 0.012065180577337742, + 0.02387881651520729, + 0.013256221078336239, + 0.007691362872719765, + 0.008403804153203964, + 0.012462284415960312, + 0.012905867770314217, + 0.008833488449454308, + 0.007912847213447094, + 0.010893971659243107, + 0.008592259138822556, + 0.006520779803395271, + 0.006852420978248119, + 0.013581447303295135, + 0.007292244583368301, + 0.006393101532012224, + 0.0037185417022556067, + 0.006887047551572323, + 0.006243361160159111, + 0.0074267638847231865, + 0.004914299119263887, + 0.012942264787852764, + 0.005037674214690924, + 0.01156283263117075, + 0.011553706601262093, + 0.0138887083157897, + 0.011671683751046658, + 0.007156903389841318, + 0.014773551374673843, + 0.009946359321475029, + 0.01618371531367302, + 0.00273236189968884, + 0.0015491436934098601, + 0.0020934781059622765, + 0.0020919821690768003, + 0.002320837927982211, + 0.0017662266036495566, + 0.0023542577400803566, + 0.0019398730946704745, + 0.0280587300658226, + 0.013131657615303993, + 0.09324033558368683, + 0.02362699806690216, + 0.0133163221180439, + 0.018297139555215836, + 0.018880588933825493, + 0.092356838285923, + 0.008342980407178402, + 0.007653483655303717, + 0.011933417990803719, + 0.00887947715818882, + 0.007053844630718231, + 0.011266014538705349, + 0.007234229706227779, + 0.00873301550745964, + 0.022511981427669525, + 0.004608298651874065, + 0.0049050347879529, + 0.00648143794387579, + 0.005582453217357397, + 0.005835303571075201, + 0.007583433762192726, + 0.01736437901854515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.11295779794454575 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13161562383174896, + 0.18758127093315125, + 0.10676891356706619, + 0.1087692379951477, + 0.04038110375404358, + 0.06558187305927277, + 0.02705359272658825, + 0.1329210102558136, + 0.036608435213565826, + 0.16271895170211792 + ], + "credit_map": [ + 0.021093418821692467, + 0.016168862581253052, + 0.023267455399036407, + 0.02120879292488098, + 0.01338451448827982, + 0.010546525940299034, + 0.014785028994083405, + 0.011161021888256073, + 0.05649668723344803, + 0.0066275643184781075, + 0.07052455842494965, + 0.010597330518066883, + 0.010200681164860725, + 0.005960481707006693, + 0.01894555427134037, + 0.008228412829339504, + 0.013149517588317394, + 0.012951839715242386, + 0.018161870539188385, + 0.013349223881959915, + 0.011279773898422718, + 0.010442535392940044, + 0.01501644030213356, + 0.012417711317539215, + 0.010996853932738304, + 0.010620568878948689, + 0.016497017815709114, + 0.02257324941456318, + 0.010177849791944027, + 0.009447358548641205, + 0.01926051266491413, + 0.009195830672979355, + 0.003926309756934643, + 0.0031460176687687635, + 0.006046266760677099, + 0.00710145803168416, + 0.002897207625210285, + 0.004809511825442314, + 0.007722278591245413, + 0.004732058383524418, + 0.00864588562399149, + 0.007700436748564243, + 0.011155324056744576, + 0.007616812363266945, + 0.005344828590750694, + 0.00965352077037096, + 0.0065628415904939175, + 0.008902219124138355, + 0.003467920236289501, + 0.0028614490292966366, + 0.0032263838220387697, + 0.002999559510499239, + 0.0026815535966306925, + 0.002282362664118409, + 0.00626175943762064, + 0.0032726034987717867, + 0.01352381706237793, + 0.010806410573422909, + 0.03117348439991474, + 0.018501851707696915, + 0.008511977270245552, + 0.010202398523688316, + 0.01121753454208374, + 0.028983525931835175, + 0.004144898150116205, + 0.004374773241579533, + 0.005003442522138357, + 0.005671788472682238, + 0.003389321267604828, + 0.005998264532536268, + 0.004176916554570198, + 0.003849032800644636, + 0.04567505419254303, + 0.007907632738351822, + 0.009333519265055656, + 0.01537343766540289, + 0.011431749910116196, + 0.013492469675838947, + 0.019354645162820816, + 0.04015045240521431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.0046369642950594425 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from an extremely low angle on the forest floor, slowly advances at an almost imperceptible pace, deep into a dark, damp woodland. Atop a thick layer of decaying, dark brown leaves, a gigantic centipede crouches silently, its head a disturbing crimson, a stark contrast to its countless segmented, glossy black body. Faint light filters through gaps in the dense forest canopy, casting a vivid, wet sheen on its glossy carapace, a physically rendered, remarkably realistic effect. Each curve and subtle texture of the carapace is clearly visible through the interplay of highlights and shadows, giving it the texture of finely polished obsidian. At the tip of its crimson head, a pair of slender tentacles trembles subtly, probing the air, while hundreds of sharp legs on its sides cling to the damp leaves, motionless and poised for action. The entire picture is pure and flawless, without any impurities or interference, except for a few leaves with condensation on them and twisted roots covered with moss, which together create a primitive, quiet yet potentially threatening atmosphere.", + "seed": 31885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_1914_seed_31885_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": -0.2121458043344319, + "vlm_logprob": -0.2121458043344319, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.2121458043344319, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5533257722854614 + ], + "opposite_logprob_tokens": [ + -0.9283257722854614 + ], + "logprob_reward": -0.5533257722854614, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5533257722854614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16197757422924042, + 0.12028724700212479, + 0.2452758550643921, + 0.10282710939645767, + 0.03818376734852791, + 0.0846596509218216, + 0.023056067526340485, + 0.11471535265445709, + 0.07944554835557938, + 0.02957185171544552 + ], + "credit_map": [ + 0.029757950454950333, + 0.01888223923742771, + 0.015914741903543472, + 0.023086484521627426, + 0.010705225169658661, + 0.019188543781638145, + 0.015861140564084053, + 0.028581222519278526, + 0.016957974061369896, + 0.011203067377209663, + 0.021510887891054153, + 0.01869114302098751, + 0.007633596658706665, + 0.010899070650339127, + 0.02059188485145569, + 0.012799615040421486, + 0.03652878478169441, + 0.01920122280716896, + 0.02125679701566696, + 0.022455763071775436, + 0.04339087754487991, + 0.03183155506849289, + 0.023291951045393944, + 0.047318872064352036, + 0.013143755495548248, + 0.012390763498842716, + 0.012929229065775871, + 0.011768532916903496, + 0.012684613466262817, + 0.012289243750274181, + 0.014296121895313263, + 0.013324839994311333, + 0.004674197174608707, + 0.003655331674963236, + 0.004577078856527805, + 0.0032871016301214695, + 0.0031257793307304382, + 0.004963758867233992, + 0.007425770629197359, + 0.006474743131548166, + 0.010334194637835026, + 0.005593298003077507, + 0.01627489924430847, + 0.008664124645292759, + 0.006028172094374895, + 0.007027113810181618, + 0.023052994161844254, + 0.0076848543249070644, + 0.003018624847754836, + 0.0021123779006302357, + 0.003478759666904807, + 0.004308112431317568, + 0.0013222999405115843, + 0.002068508183583617, + 0.003507738234475255, + 0.0032396437600255013, + 0.010620823130011559, + 0.0073696039617061615, + 0.014496601186692715, + 0.010896142572164536, + 0.007317868527024984, + 0.010877861641347408, + 0.019014965742826462, + 0.034121476113796234, + 0.005372694227844477, + 0.006460477598011494, + 0.010154522024095058, + 0.01452675461769104, + 0.007947024889290333, + 0.005597260314971209, + 0.01896623522043228, + 0.010420568287372589, + 0.0035199744161218405, + 0.0024507828056812286, + 0.004215719178318977, + 0.005547218024730682, + 0.0022222385741770267, + 0.0024355691857635975, + 0.004752623848617077, + 0.00442772451788187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.5533257722854614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871375322341919 + ], + "opposite_logprob_tokens": [ + -1.137137532234192 + ], + "logprob_reward": -0.3871375322341919, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871375322341919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1317426562309265, + 0.16648605465888977, + 0.07713834196329117, + 0.05723850056529045, + 0.05486729368567467, + 0.06617351621389389, + 0.052926838397979736, + 0.3155917525291443, + 0.047012683004140854, + 0.030822400003671646 + ], + "credit_map": [ + 0.024172596633434296, + 0.012415595352649689, + 0.01543687004595995, + 0.01773330755531788, + 0.011337701231241226, + 0.02166021801531315, + 0.011613604612648487, + 0.017372747883200645, + 0.02818971686065197, + 0.013314411975443363, + 0.036535266786813736, + 0.028423592448234558, + 0.009630003944039345, + 0.013643411919474602, + 0.020558910444378853, + 0.01619071327149868, + 0.00986204482614994, + 0.010893360711634159, + 0.013187761418521404, + 0.010223670862615108, + 0.007480665110051632, + 0.007298139855265617, + 0.008806777186691761, + 0.009385921992361546, + 0.007885935716331005, + 0.004435392562299967, + 0.007262070197612047, + 0.005594837013632059, + 0.00522730452939868, + 0.00898019876331091, + 0.009738871827721596, + 0.008113881573081017, + 0.007554376497864723, + 0.004078924655914307, + 0.008697131648659706, + 0.007245131768286228, + 0.004518595058470964, + 0.007623379118740559, + 0.009150121361017227, + 0.005999630317091942, + 0.007975911721587181, + 0.00474796025082469, + 0.012442568317055702, + 0.006101925857365131, + 0.0048318770714104176, + 0.005053410306572914, + 0.018976911902427673, + 0.006042945198714733, + 0.004566982388496399, + 0.004019112791866064, + 0.009974948130548, + 0.008340197615325451, + 0.0036545286420732737, + 0.004662206396460533, + 0.008464127779006958, + 0.009244732558727264, + 0.030505146831274033, + 0.015793094411492348, + 0.03839252144098282, + 0.03378521278500557, + 0.021013567224144936, + 0.027920853346586227, + 0.05386916548013687, + 0.09431218355894089, + 0.0043615009635686874, + 0.003677387721836567, + 0.006560632959008217, + 0.008139487355947495, + 0.003177851205691695, + 0.002060691826045513, + 0.011400928720831871, + 0.0076341996900737286, + 0.002940455451607704, + 0.0030266842804849148, + 0.004224526695907116, + 0.005821396131068468, + 0.002205437747761607, + 0.002393980510532856, + 0.005325646139681339, + 0.004884270019829273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.3871375322341919 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032010050490498543 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032010050490498543, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032010050490498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1451035887002945, + 0.3011503219604492, + 0.12676014006137848, + 0.08539856225252151, + 0.05977672338485718, + 0.037508524954319, + 0.02543899416923523, + 0.147692009806633, + 0.04423583671450615, + 0.026935288682579994 + ], + "credit_map": [ + 0.03289832919836044, + 0.011130315251648426, + 0.01873985305428505, + 0.020422618836164474, + 0.012722576968371868, + 0.01268826425075531, + 0.013018508441746235, + 0.02348313294351101, + 0.04280789941549301, + 0.022650880739092827, + 0.03556699678301811, + 0.03481961786746979, + 0.05472911521792412, + 0.036421217024326324, + 0.04204950109124184, + 0.03210509568452835, + 0.020090457051992416, + 0.011484709568321705, + 0.03001679852604866, + 0.015390870161354542, + 0.009843320585787296, + 0.011557267978787422, + 0.013617645017802715, + 0.014759068377315998, + 0.013070272281765938, + 0.006833421997725964, + 0.01174667663872242, + 0.010304112918674946, + 0.008980118669569492, + 0.007170228753238916, + 0.012493233196437359, + 0.014800495468080044, + 0.008451856672763824, + 0.004514004569500685, + 0.010152272880077362, + 0.011085039004683495, + 0.004863903857767582, + 0.0056075588800013065, + 0.008718933910131454, + 0.0063831559382379055, + 0.0037747162859886885, + 0.0033914896193891764, + 0.00671655498445034, + 0.004924757406115532, + 0.0030853927601128817, + 0.0041216518729925156, + 0.007839297875761986, + 0.0036546618212014437, + 0.0024947295896708965, + 0.001987254014238715, + 0.003571743844076991, + 0.0036518180277198553, + 0.002315379213541746, + 0.0038647600449621677, + 0.004211104474961758, + 0.003342204261571169, + 0.01408332772552967, + 0.006636856123805046, + 0.02279970422387123, + 0.01590524986386299, + 0.008262285962700844, + 0.0118442727252841, + 0.024703973904252052, + 0.04345633089542389, + 0.004506322555243969, + 0.003967911005020142, + 0.004961264319717884, + 0.008236058056354523, + 0.003998479340225458, + 0.004338445141911507, + 0.009826216846704483, + 0.00440114364027977, + 0.0021582217887043953, + 0.003352017607539892, + 0.0034359139390289783, + 0.004779814276844263, + 0.002131199697032571, + 0.0022447144147008657, + 0.0031777527183294296, + 0.005655654706060886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0032010050490498543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296897381544113 + ], + "opposite_logprob_tokens": [ + -2.237968921661377 + ], + "logprob_reward": -0.11296897381544113, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296897381544113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1646045297384262, + 0.24322466552257538, + 0.10249194502830505, + 0.05430835857987404, + 0.04667147248983383, + 0.07547757029533386, + 0.04140172898769379, + 0.12768150866031647, + 0.09480469673871994, + 0.049333542585372925 + ], + "credit_map": [ + 0.026700181886553764, + 0.010141786187887192, + 0.01840994320809841, + 0.021780407056212425, + 0.016579126939177513, + 0.028050929307937622, + 0.015024753287434578, + 0.02791743166744709, + 0.0412028506398201, + 0.022149961441755295, + 0.05295576527714729, + 0.03839293494820595, + 0.019533125683665276, + 0.010465903207659721, + 0.03697912395000458, + 0.02154502086341381, + 0.01568513549864292, + 0.00709389615803957, + 0.0184620451182127, + 0.008309373632073402, + 0.018417127430438995, + 0.009535698220133781, + 0.012628442607820034, + 0.012360231950879097, + 0.006805220618844032, + 0.004793442785739899, + 0.006572446785867214, + 0.005876838695257902, + 0.007095183711498976, + 0.006288847886025906, + 0.007422440219670534, + 0.009453946724534035, + 0.005106768105179071, + 0.0029999902471899986, + 0.0062201181426644325, + 0.006253356579691172, + 0.004275911021977663, + 0.006701429840177298, + 0.007069852668792009, + 0.00804404728114605, + 0.008935671299695969, + 0.004982470069080591, + 0.013377376832067966, + 0.0077521540224552155, + 0.0047012390568852425, + 0.006812940817326307, + 0.02120525948703289, + 0.0077104708179831505, + 0.004312875680625439, + 0.003496540477499366, + 0.005003521218895912, + 0.006571960169821978, + 0.003315764246508479, + 0.004995272029191256, + 0.0066277808509767056, + 0.007078018505126238, + 0.009378881193697453, + 0.012761557474732399, + 0.015157978981733322, + 0.013863580301404, + 0.00946031790226698, + 0.018256889656186104, + 0.018739568069577217, + 0.030062738806009293, + 0.010218638926744461, + 0.010729951784014702, + 0.01590777561068535, + 0.0160216111689806, + 0.005692025646567345, + 0.006041433196514845, + 0.017019614577293396, + 0.013173655606806278, + 0.0038865054957568645, + 0.00469695869833231, + 0.006396741606295109, + 0.009404612705111504, + 0.004009996075183153, + 0.004068286158144474, + 0.009095831774175167, + 0.007774615194648504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.11296897381544113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095738288015127 + ], + "opposite_logprob_tokens": [ + -5.504095554351807 + ], + "logprob_reward": -0.004095738288015127, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095738288015127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15129855275154114, + 0.1243680790066719, + 0.2052399218082428, + 0.10903549939393997, + 0.07715778052806854, + 0.07150009274482727, + 0.04063308984041214, + 0.06777678430080414, + 0.0946316048502922, + 0.05835860222578049 + ], + "credit_map": [ + 0.027913065627217293, + 0.009636792354285717, + 0.015254741534590721, + 0.01933688297867775, + 0.01288413256406784, + 0.02361432835459709, + 0.016104644164443016, + 0.02655397169291973, + 0.01984071545302868, + 0.008871663361787796, + 0.023914141580462456, + 0.01919408142566681, + 0.011812818236649036, + 0.01218861248344183, + 0.0168578140437603, + 0.011688227765262127, + 0.030194422230124474, + 0.011306017637252808, + 0.016372229903936386, + 0.013621160760521889, + 0.05559000000357628, + 0.022646084427833557, + 0.016067033633589745, + 0.03944297507405281, + 0.019980449229478836, + 0.009456475265324116, + 0.010607331059873104, + 0.011240635067224503, + 0.014049996621906757, + 0.008002887479960918, + 0.012935738079249859, + 0.022761981934309006, + 0.011756760999560356, + 0.004918671678751707, + 0.012347714975476265, + 0.009500979445874691, + 0.00575612485408783, + 0.009475814178586006, + 0.013323506340384483, + 0.010078215040266514, + 0.007394967135041952, + 0.006270214915275574, + 0.011044560000300407, + 0.00790796335786581, + 0.006668841931968927, + 0.010072052478790283, + 0.01476866751909256, + 0.007372830528765917, + 0.004423268139362335, + 0.003616125090047717, + 0.00810330267995596, + 0.00723812822252512, + 0.002647132147103548, + 0.004499510861933231, + 0.0067484742030501366, + 0.0033571524545550346, + 0.006827640812844038, + 0.0077522932551801205, + 0.010587598197162151, + 0.010953843593597412, + 0.00701042078435421, + 0.006422633770853281, + 0.01048612967133522, + 0.007736224681138992, + 0.008064278401434422, + 0.008378051221370697, + 0.012272242456674576, + 0.016262030228972435, + 0.012645328417420387, + 0.00761938001960516, + 0.01748020574450493, + 0.011910095810890198, + 0.005176139995455742, + 0.0034568747505545616, + 0.008716970682144165, + 0.011313335038721561, + 0.00517094973474741, + 0.005355345085263252, + 0.008893970400094986, + 0.010275019332766533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.004095738288015127 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, captured from an extremely low angle on the forest floor, slowly advances at an almost imperceptible pace, deep into a dark, damp woodland. Atop a thick layer of decaying, dark brown leaves, a gigantic centipede crouches silently, its head a disturbing crimson, a stark contrast to its countless segmented, glossy black body. Faint light filters through gaps in the dense forest canopy, casting a vivid, wet sheen on its glossy carapace, a physically rendered, remarkably realistic effect. Each curve and subtle texture of the carapace is clearly visible through the interplay of highlights and shadows, giving it the texture of finely polished obsidian. At the tip of its crimson head, a pair of slender tentacles trembles subtly, probing the air, while hundreds of sharp legs on its sides cling to the damp leaves, motionless and poised for action. The entire picture is pure and flawless, without any impurities or interference, except for a few leaves with condensation on them and twisted roots covered with moss, which together create a primitive, quiet yet potentially threatening atmosphere.", + "seed": 31886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_12/index_1914_seed_31886_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": -0.15951984585262835, + "vlm_logprob": -0.15951984585262835, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.15951984585262835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20140524208545685 + ], + "opposite_logprob_tokens": [ + -1.8264052867889404 + ], + "logprob_reward": -0.20140524208545685, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.20140524208545685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2673359513282776, + 0.05328848585486412, + 0.06279262900352478, + 0.02955988049507141, + 0.13492032885551453, + 0.08449816703796387, + 0.028475584462285042, + 0.05894026160240173, + 0.11567961424589157, + 0.1645091027021408 + ], + "credit_map": [ + 0.02874363772571087, + 0.01400792971253395, + 0.055021580308675766, + 0.02546369656920433, + 0.03211073577404022, + 0.04667114466428757, + 0.02476710081100464, + 0.040550123900175095, + 0.005691003520041704, + 0.006357480771839619, + 0.006532938219606876, + 0.0070952014066278934, + 0.004111978225409985, + 0.008360167033970356, + 0.011051957495510578, + 0.004087760578840971, + 0.003506290027871728, + 0.002986516570672393, + 0.008723885752260685, + 0.018703676760196686, + 0.004821142181754112, + 0.010703054256737232, + 0.007284205406904221, + 0.0060638608410954475, + 0.0023352305870503187, + 0.002443652367219329, + 0.002715611131861806, + 0.004479615017771721, + 0.00393819622695446, + 0.004101993981748819, + 0.004523834679275751, + 0.005021747201681137, + 0.014057666063308716, + 0.01031606737524271, + 0.016354184597730637, + 0.021012362092733383, + 0.017309416085481644, + 0.021521667018532753, + 0.01521008089184761, + 0.019138893112540245, + 0.013681436888873577, + 0.007480794098228216, + 0.013525739312171936, + 0.011851069517433643, + 0.007574282120913267, + 0.008110214956104755, + 0.009848416782915592, + 0.012426215223968029, + 0.002863779664039612, + 0.0014307963429018855, + 0.0031766307074576616, + 0.002580412896350026, + 0.0031526372767984867, + 0.006419793702661991, + 0.006085614673793316, + 0.0027659174520522356, + 0.005078843329101801, + 0.004582935478538275, + 0.007383153308182955, + 0.010101460851728916, + 0.007933810353279114, + 0.006286791525781155, + 0.008301629684865475, + 0.009271636605262756, + 0.01216696947813034, + 0.01015961728990078, + 0.014999531209468842, + 0.017438437789678574, + 0.008123704232275486, + 0.020383784547448158, + 0.013379340060055256, + 0.019028235226869583, + 0.006847777403891087, + 0.008353673852980137, + 0.009901084005832672, + 0.03949333727359772, + 0.020226305350661278, + 0.046098433434963226, + 0.019207466393709183, + 0.014381027780473232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.20140524208545685 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743521809577942 + ], + "opposite_logprob_tokens": [ + -0.9743521809577942 + ], + "logprob_reward": -0.4743521809577942, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743521809577942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34670406579971313, + 0.06581240892410278, + 0.2442961186170578, + 0.05634065344929695, + 0.0679176077246666, + 0.0413554422557354, + 0.023176804184913635, + 0.0368930883705616, + 0.03509877249598503, + 0.08240504562854767 + ], + "credit_map": [ + 0.04237784072756767, + 0.019720980897545815, + 0.0695820227265358, + 0.03411385416984558, + 0.033933550119400024, + 0.0540371835231781, + 0.04371125251054764, + 0.04922739416360855, + 0.005865919403731823, + 0.0035521371755748987, + 0.009538615122437477, + 0.008937150239944458, + 0.005018011201173067, + 0.012293960899114609, + 0.014461974613368511, + 0.006144643295556307, + 0.012550398707389832, + 0.008044864982366562, + 0.023403272032737732, + 0.08955629914999008, + 0.0077912393026053905, + 0.07483868300914764, + 0.015135628171265125, + 0.012975728139281273, + 0.004346223548054695, + 0.0032713657710701227, + 0.0056747253984212875, + 0.010439279489219189, + 0.006402689963579178, + 0.00704772025346756, + 0.007399390451610088, + 0.011759258806705475, + 0.009526129812002182, + 0.005910038016736507, + 0.007841942831873894, + 0.011443434283137321, + 0.007879718206822872, + 0.009868687950074673, + 0.006671526934951544, + 0.008776129223406315, + 0.008857733570039272, + 0.0040194629691541195, + 0.00596190057694912, + 0.004580971784889698, + 0.0028659796807914972, + 0.00506655965000391, + 0.004385077394545078, + 0.005617758259177208, + 0.0016702394932508469, + 0.0011277147568762302, + 0.0029633829835802317, + 0.0018291042651981115, + 0.003174553159624338, + 0.005746394861489534, + 0.00376748014241457, + 0.0028979340568184853, + 0.0033333473838865757, + 0.004211912397295237, + 0.004360374063253403, + 0.005230194889008999, + 0.004688143730163574, + 0.004344507586210966, + 0.005567481741309166, + 0.005157125182449818, + 0.0023721756879240274, + 0.0032423995435237885, + 0.0050218855030834675, + 0.004567025229334831, + 0.00550415925681591, + 0.0050132665783166885, + 0.004015197977423668, + 0.005362664349377155, + 0.00793237891048193, + 0.005539099220186472, + 0.011167493648827076, + 0.01435539498925209, + 0.007151162251830101, + 0.012278948910534382, + 0.011340982280671597, + 0.012639584019780159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.4743521809577942 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17407110333442688, + 0.023314502090215683, + 0.08669345825910568, + 0.026015188544988632, + 0.0293282363563776, + 0.11537782847881317, + 0.02484746277332306, + 0.04669849947094917, + 0.41874295473098755, + 0.05491078272461891 + ], + "credit_map": [ + 0.02385062724351883, + 0.010172647424042225, + 0.03265949711203575, + 0.01739376224577427, + 0.019076015800237656, + 0.029445083811879158, + 0.02090902253985405, + 0.020564429461956024, + 0.002116352552548051, + 0.0011886331485584378, + 0.0032686239574104548, + 0.0031849255319684744, + 0.0023081020917743444, + 0.003974794875830412, + 0.005043456796556711, + 0.002229610225185752, + 0.00429086247459054, + 0.002854142105206847, + 0.0080415615811944, + 0.03105592355132103, + 0.002849089214578271, + 0.027602389454841614, + 0.0055160620249807835, + 0.004483426455408335, + 0.0024051936343312263, + 0.0020213250536471605, + 0.0036619652528315783, + 0.0037226350978016853, + 0.00448235310614109, + 0.0029954316560178995, + 0.0032933882903307676, + 0.0034328943584114313, + 0.003008819418027997, + 0.002330028684809804, + 0.003944970201700926, + 0.004314700607210398, + 0.004785945639014244, + 0.003292966168373823, + 0.003737197257578373, + 0.003913603723049164, + 0.00679198931902647, + 0.004822360351681709, + 0.023228470236063004, + 0.021002134308218956, + 0.01158493384718895, + 0.011949634179472923, + 0.017338424921035767, + 0.01865987665951252, + 0.0009965389035642147, + 0.0009634828893467784, + 0.0036104326136410236, + 0.0029641524888575077, + 0.0030728268902748823, + 0.00574068259447813, + 0.004569100681692362, + 0.0029302428010851145, + 0.0053309365175664425, + 0.0031713612843304873, + 0.00465979753062129, + 0.005878213327378035, + 0.006715306080877781, + 0.006154375616461039, + 0.007414301857352257, + 0.007374201901257038, + 0.03610078990459442, + 0.029678180813789368, + 0.04211712256073952, + 0.05552832409739494, + 0.06155826523900032, + 0.08427028357982635, + 0.040608882904052734, + 0.06888105720281601, + 0.006676423363387585, + 0.005008803680539131, + 0.006606998387724161, + 0.0073733460158109665, + 0.008651993237435818, + 0.007410124875605106, + 0.005556214135140181, + 0.007626870647072792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.005258656572550535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296716332435608 + ], + "opposite_logprob_tokens": [ + -2.2379672527313232 + ], + "logprob_reward": -0.11296716332435608, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296716332435608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12017536908388138, + 0.03683920204639435, + 0.08525221049785614, + 0.02934114821255207, + 0.10200902819633484, + 0.05974636226892471, + 0.12500958144664764, + 0.073704794049263, + 0.2946496307849884, + 0.07327266037464142 + ], + "credit_map": [ + 0.014499024488031864, + 0.007237197365611792, + 0.023649489507079124, + 0.011704725213348866, + 0.015621724538505077, + 0.018970103934407234, + 0.01748582534492016, + 0.011007281951606274, + 0.0034979907795786858, + 0.0022024877835065126, + 0.005383213050663471, + 0.00543510727584362, + 0.0030836379155516624, + 0.006153632886707783, + 0.008547322824597359, + 0.0025358079001307487, + 0.004180108197033405, + 0.004006269387900829, + 0.008029313758015633, + 0.028200509026646614, + 0.0028643461409956217, + 0.023837042972445488, + 0.009328516200184822, + 0.0048061092384159565, + 0.0036380081437528133, + 0.0024259714409708977, + 0.003583265235647559, + 0.004382777959108353, + 0.003584502497687936, + 0.003463613335043192, + 0.0034145312383770943, + 0.004848478361964226, + 0.010768593288958073, + 0.007522089872509241, + 0.012611006386578083, + 0.01401170901954174, + 0.014129585586488247, + 0.017811471596360207, + 0.01128946803510189, + 0.013865099288523197, + 0.009245552122592926, + 0.008327866904437542, + 0.00929961446672678, + 0.006756563670933247, + 0.0057634455151855946, + 0.006337000988423824, + 0.006286753341555595, + 0.007729565259069204, + 0.006217495072633028, + 0.005093551706522703, + 0.01962183229625225, + 0.014897694811224937, + 0.017781518399715424, + 0.023241901770234108, + 0.02218019962310791, + 0.01597539335489273, + 0.00838722474873066, + 0.006737290881574154, + 0.010672097094357014, + 0.008930040523409843, + 0.008755515329539776, + 0.0076508671045303345, + 0.012599185109138489, + 0.009972566738724709, + 0.02600107528269291, + 0.021721096709370613, + 0.023952871561050415, + 0.041828203946352005, + 0.04055199772119522, + 0.058029912412166595, + 0.03378792107105255, + 0.04877655208110809, + 0.0038399510085582733, + 0.0036892781499773264, + 0.006229552440345287, + 0.013964688405394554, + 0.00885817501693964, + 0.016512345522642136, + 0.013097686693072319, + 0.007080982439219952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.11296716332435608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle on the forest floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly advance deep into a dark, damp woodland?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a gigantic centipede with a crimson head and glossy black segmented body visible atop decaying leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are slender tentacles trembling at the tip of the centipede's head while its legs remain motionless?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by faint light filtering through the canopy, creating a wet sheen on the centipede's carapace?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036159863229840994 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036159863229840994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036159863229840994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21905085444450378, + 0.07199938595294952, + 0.11378267407417297, + 0.05937871336936951, + 0.0709577426314354, + 0.06844404339790344, + 0.07960034906864166, + 0.0464203879237175, + 0.09700816869735718, + 0.17335766553878784 + ], + "credit_map": [ + 0.021027369424700737, + 0.010288512334227562, + 0.04505746066570282, + 0.024608178064227104, + 0.044344231486320496, + 0.02718045376241207, + 0.030473671853542328, + 0.016070984303951263, + 0.008716091513633728, + 0.006352735683321953, + 0.00934456568211317, + 0.00941360741853714, + 0.006011278368532658, + 0.009381660260260105, + 0.01608888991177082, + 0.006690551061183214, + 0.005013872869312763, + 0.004947924055159092, + 0.010419727303087711, + 0.03741561248898506, + 0.0052944025956094265, + 0.036206696182489395, + 0.00791140180081129, + 0.006573036313056946, + 0.006067861802875996, + 0.0044569894671440125, + 0.00695462292060256, + 0.00989289116114378, + 0.005725422874093056, + 0.005476613063365221, + 0.009099075570702553, + 0.011705237440764904, + 0.004944398999214172, + 0.003901134477928281, + 0.009283918887376785, + 0.011763611808419228, + 0.011040055193006992, + 0.011447925120592117, + 0.008489149622619152, + 0.010087543167173862, + 0.009136155247688293, + 0.006122875027358532, + 0.00801767222583294, + 0.01198136992752552, + 0.00986473448574543, + 0.0058842189610004425, + 0.010512802749872208, + 0.006924215704202652, + 0.006306121591478586, + 0.004676913842558861, + 0.009739824570715427, + 0.008037210442125797, + 0.007257987279444933, + 0.0184657983481884, + 0.013496912084519863, + 0.01161958184093237, + 0.002915432909503579, + 0.0038552754558622837, + 0.007179040461778641, + 0.0058782463893294334, + 0.006392437964677811, + 0.00635912548750639, + 0.007694919593632221, + 0.00614590710029006, + 0.007122161332517862, + 0.005749496165663004, + 0.012683073990046978, + 0.010984164662659168, + 0.01741372048854828, + 0.013005119748413563, + 0.01225300319492817, + 0.017797430977225304, + 0.01742936484515667, + 0.010224253870546818, + 0.02554485574364662, + 0.032382793724536896, + 0.016725657507777214, + 0.034149520099163055, + 0.016457686200737953, + 0.020443545654416084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0036159863229840994 + } + ], + "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/0000125/group_6/index_tensor([4363, 4363])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_tensor([4363, 4363])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa474b4385f532a0a89bbbec0e6fb8aa0dfd487a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_tensor([4363, 4363])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme close-up focuses intently on the 5.25-inch drive bay of a black brushed metal desktop computer tower. As the drive's front panel door carefully slides open from the inside, a quiet internal whirring begins the operation. Then, with mechanical precision, a matte black plastic tray appears, sliding out smoothly and deliberately on its internal track. As it extends into the foreground, it reveals a pristine silver disc, perfectly nestled in its circular groove. Cool, direct light from an unseen source gleams on the disc's surface, creating brilliant mirror-like highlights and casting a rainbow-like sheen that dances across its data layer. The tray slides to a soft, definitive stop, fully emerging from the computer's chassis, giving the disc a sense of finality. The atmosphere is focused, the quiet, and satisfying mechanical function capturing the device's clean, precise engineering in a moment of silent operation.", + "seed": 21269, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21269_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": -0.010728198895230889, + "vlm_logprob": -0.010728198895230889, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.010728198895230889, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01437757071107626 + ], + "opposite_logprob_tokens": [ + -4.889377593994141 + ], + "logprob_reward": -0.01437757071107626, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01437757071107626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.304170161485672, + 0.11466481536626816, + 0.05245569348335266, + 0.19941511750221252, + 0.06617963314056396, + 0.052207473665475845, + 0.032629504799842834, + 0.03361882269382477, + 0.06317459791898727, + 0.08148419111967087 + ], + "credit_map": [ + 0.024597350507974625, + 0.0178520530462265, + 0.09931322932243347, + 0.11316356807947159, + 0.01057538203895092, + 0.016183381900191307, + 0.010425332933664322, + 0.012059858068823814, + 0.014137078076601028, + 0.006365804933011532, + 0.026812486350536346, + 0.030731169506907463, + 0.006024263333529234, + 0.008939845487475395, + 0.011430714279413223, + 0.010223454795777798, + 0.008376022800803185, + 0.004255295265465975, + 0.008054308593273163, + 0.00536250090226531, + 0.0035927908029407263, + 0.005905758589506149, + 0.008070206269621849, + 0.008838806301355362, + 0.008318763226270676, + 0.0075142113491892815, + 0.006241246126592159, + 0.02456887811422348, + 0.0071958149783313274, + 0.006497528869658709, + 0.08469833433628082, + 0.05438033118844032, + 0.006340831983834505, + 0.012068449519574642, + 0.012103386223316193, + 0.008030952885746956, + 0.004343780223280191, + 0.007124279625713825, + 0.007880982011556625, + 0.008286970667541027, + 0.0046139792539179325, + 0.005439814645797014, + 0.009479082189500332, + 0.007083813194185495, + 0.0037876670248806477, + 0.006264551542699337, + 0.008069229312241077, + 0.007469329982995987, + 0.0036522590089589357, + 0.007598085794597864, + 0.004172428511083126, + 0.003242471953853965, + 0.002976418472826481, + 0.00462697958573699, + 0.002883797511458397, + 0.0034770627971738577, + 0.0030907492619007826, + 0.004179385025054216, + 0.0038055279292166233, + 0.004044755361974239, + 0.003981715068221092, + 0.005317796021699905, + 0.004316999576985836, + 0.0048818932846188545, + 0.004975265823304653, + 0.004778650589287281, + 0.02430434338748455, + 0.004571855533868074, + 0.00273657008074224, + 0.0039058392867445946, + 0.004144419450312853, + 0.013757657259702682, + 0.005803915672004223, + 0.005288740154355764, + 0.026439277455210686, + 0.024253275245428085, + 0.004344747867435217, + 0.0040720440447330475, + 0.005789988208562136, + 0.005492206662893295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01437757071107626 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004130048677325249 + ], + "opposite_logprob_tokens": [ + -5.504129886627197 + ], + "logprob_reward": -0.004130048677325249, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004130048677325249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15967939794063568, + 0.09694287180900574, + 0.06944873929023743, + 0.3187824785709381, + 0.06367530673742294, + 0.03571347892284393, + 0.04741509631276131, + 0.06834714859724045, + 0.0917133316397667, + 0.048282183706760406 + ], + "credit_map": [ + 0.016227087005972862, + 0.01571451500058174, + 0.034055210649967194, + 0.05871911346912384, + 0.009556874632835388, + 0.006436619907617569, + 0.007264439947903156, + 0.011705530807375908, + 0.013442843221127987, + 0.0071272654458880424, + 0.011337661184370518, + 0.013942224904894829, + 0.007264170795679092, + 0.016098078340291977, + 0.012129359878599644, + 0.015601268038153648, + 0.0066404701210558414, + 0.004076348152011633, + 0.0060419924557209015, + 0.021601136773824692, + 0.0030502041336148977, + 0.0039044402074068785, + 0.00753805274143815, + 0.016596093773841858, + 0.013503278605639935, + 0.013024481944739819, + 0.008831840939819813, + 0.03572199121117592, + 0.011594850569963455, + 0.017283882945775986, + 0.12998288869857788, + 0.08883925527334213, + 0.006142776925116777, + 0.007320399861782789, + 0.005389631260186434, + 0.007406407967209816, + 0.01294297631829977, + 0.0073792580515146255, + 0.008916862308979034, + 0.008176990784704685, + 0.003113700309768319, + 0.007789358031004667, + 0.005453254096210003, + 0.003944208845496178, + 0.0027313102036714554, + 0.0029058351647108793, + 0.0037301587872207165, + 0.0060456558130681515, + 0.005025114398449659, + 0.008703157305717468, + 0.004779777955263853, + 0.0049895853735506535, + 0.0036112270317971706, + 0.006953234784305096, + 0.007094577886164188, + 0.006258420180529356, + 0.005789260379970074, + 0.006289633922278881, + 0.02125236764550209, + 0.007880210876464844, + 0.003836470190435648, + 0.008871788159012794, + 0.00656635407358408, + 0.007861061953008175, + 0.011071733199059963, + 0.007704729214310646, + 0.02736332081258297, + 0.009083525277674198, + 0.004685347434133291, + 0.005991028621792793, + 0.00831903237849474, + 0.01749461703002453, + 0.006159569136798382, + 0.004385347943753004, + 0.006485660560429096, + 0.011449252255260944, + 0.0038389908149838448, + 0.0042311460711061954, + 0.005171936005353928, + 0.006560282316058874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.004130048677325249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644185960292816 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.008644185960292816, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644185960292816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11189530789852142, + 0.2638966739177704, + 0.04422618821263313, + 0.03548816591501236, + 0.0816531702876091, + 0.06200287863612175, + 0.04668740555644035, + 0.05932897701859474, + 0.23788566887378693, + 0.056935619562864304 + ], + "credit_map": [ + 0.018330812454223633, + 0.01410005148500204, + 0.027915101498365402, + 0.014845785684883595, + 0.010352245531976223, + 0.009262925945222378, + 0.005873559508472681, + 0.011214816011488438, + 0.022599369287490845, + 0.009484454058110714, + 0.08288118243217468, + 0.11068576574325562, + 0.005155773833394051, + 0.006833094172179699, + 0.012607067823410034, + 0.013649918138980865, + 0.0058695850893855095, + 0.004270823206752539, + 0.00461951456964016, + 0.012006239034235477, + 0.002522510476410389, + 0.004740417003631592, + 0.004417566582560539, + 0.005779528059065342, + 0.0028666986618191004, + 0.0018991578835994005, + 0.0035261933226138353, + 0.0032909212168306112, + 0.0018654253799468279, + 0.00425633043050766, + 0.00873540248721838, + 0.009048033505678177, + 0.008072908036410809, + 0.010006079450249672, + 0.007839679718017578, + 0.010552078485488892, + 0.01603655144572258, + 0.008507794700562954, + 0.009855901822447777, + 0.010782172903418541, + 0.005251627415418625, + 0.0067584458738565445, + 0.010765845887362957, + 0.009893813170492649, + 0.005789261311292648, + 0.008670441806316376, + 0.006772901397198439, + 0.008100539445877075, + 0.004727036692202091, + 0.011518281884491444, + 0.005357612390071154, + 0.004139637108892202, + 0.004066506866365671, + 0.005164915695786476, + 0.006303917150944471, + 0.005409496836364269, + 0.00457313098013401, + 0.00540171330794692, + 0.01563597470521927, + 0.007809366099536419, + 0.004176059737801552, + 0.010260437615215778, + 0.005265136249363422, + 0.006207152269780636, + 0.031103402376174927, + 0.01897447742521763, + 0.07402630150318146, + 0.017631910741329193, + 0.00886628869920969, + 0.014420581981539726, + 0.01588600128889084, + 0.05697667598724365, + 0.0057910289615392685, + 0.005513847339898348, + 0.020524999126791954, + 0.009773682802915573, + 0.003650279249995947, + 0.0039967698976397514, + 0.003382320050150156, + 0.004302692599594593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.008644185960292816 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00320064858533442 + ], + "opposite_logprob_tokens": [ + -5.753200531005859 + ], + "logprob_reward": -0.00320064858533442, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00320064858533442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12221255898475647, + 0.21571041643619537, + 0.11634181439876556, + 0.07586146146059036, + 0.10182712972164154, + 0.12320300936698914, + 0.05888568237423897, + 0.05195416510105133, + 0.06668919324874878, + 0.0673145279288292 + ], + "credit_map": [ + 0.021950799971818924, + 0.020306073129177094, + 0.009429408237338066, + 0.012996111996471882, + 0.014352403581142426, + 0.013721736147999763, + 0.008448146283626556, + 0.02100789174437523, + 0.023723836988210678, + 0.008768265135586262, + 0.05311519280076027, + 0.08095481991767883, + 0.006341894622892141, + 0.013261819258332253, + 0.012115690857172012, + 0.01742890477180481, + 0.011254162527620792, + 0.012788230553269386, + 0.010843712836503983, + 0.035160113126039505, + 0.007375291083008051, + 0.007797727826982737, + 0.012722685933113098, + 0.018399905413389206, + 0.00797872431576252, + 0.004509723745286465, + 0.00907582975924015, + 0.011020886711776257, + 0.003406207077205181, + 0.00966027844697237, + 0.022855611518025398, + 0.007354199420660734, + 0.011726233176887035, + 0.022461147978901863, + 0.010946565307676792, + 0.012358714826405048, + 0.013264663517475128, + 0.0077270991168916225, + 0.0124488091096282, + 0.01089390181005001, + 0.009579313918948174, + 0.013201587833464146, + 0.024932825937867165, + 0.019729578867554665, + 0.010851897299289703, + 0.012410849332809448, + 0.015291515737771988, + 0.01720544509589672, + 0.006515472661703825, + 0.013741772621870041, + 0.008059920743107796, + 0.004944339860230684, + 0.004954659845679998, + 0.00700788851827383, + 0.006317024119198322, + 0.007344610523432493, + 0.00576518289744854, + 0.0068930163979530334, + 0.013453157618641853, + 0.006471442524343729, + 0.003166483948007226, + 0.006689954549074173, + 0.005007443018257618, + 0.004507487639784813, + 0.011869034729897976, + 0.010531160980463028, + 0.008999410085380077, + 0.007111088372766972, + 0.0052926079370081425, + 0.005937766749411821, + 0.007869462482631207, + 0.00907866470515728, + 0.007858943194150925, + 0.00772588886320591, + 0.020639752969145775, + 0.00473934318870306, + 0.005326748825609684, + 0.005509388167411089, + 0.006013140082359314, + 0.009501321241259575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.00320064858533442 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288540542125702 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.023288540542125702, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288540542125702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1286677122116089, + 0.06297654658555984, + 0.08821628242731094, + 0.02040952257812023, + 0.06574438512325287, + 0.027667995542287827, + 0.08080877363681793, + 0.07758452743291855, + 0.06475017964839935, + 0.38317397236824036 + ], + "credit_map": [ + 0.014257999137043953, + 0.009088085032999516, + 0.024990342557430267, + 0.04667118191719055, + 0.009176960214972496, + 0.006380555685609579, + 0.005963776260614395, + 0.012138820253312588, + 0.008550943806767464, + 0.0057629309594631195, + 0.007254894822835922, + 0.007592128124088049, + 0.0042719184421002865, + 0.01157408393919468, + 0.006959485821425915, + 0.011010166257619858, + 0.00799765344709158, + 0.00869432557374239, + 0.007808632683008909, + 0.01888476498425007, + 0.005602809600532055, + 0.004708504304289818, + 0.009263701736927032, + 0.025255896151065826, + 0.002405181061476469, + 0.0019161994569003582, + 0.003492396790534258, + 0.0018362529808655381, + 0.0015606018714606762, + 0.0026863410603255033, + 0.0027207473758608103, + 0.0037918041925877333, + 0.006900475826114416, + 0.012914360500872135, + 0.009069052524864674, + 0.009303770959377289, + 0.004304572939872742, + 0.005435642786324024, + 0.009658207185566425, + 0.00815830659121275, + 0.0029538192320615053, + 0.0035241146106272936, + 0.004372395109385252, + 0.004213929176330566, + 0.0022509433329105377, + 0.002307695569470525, + 0.0035536354407668114, + 0.004491464700549841, + 0.009228883311152458, + 0.012628764845430851, + 0.00992412120103836, + 0.009092994965612888, + 0.006536835338920355, + 0.012457146309316158, + 0.01160803809762001, + 0.009331997483968735, + 0.006563395727425814, + 0.007032772060483694, + 0.027786262333393097, + 0.009544032625854015, + 0.003501293482258916, + 0.009196904487907887, + 0.006242783740162849, + 0.007717095781117678, + 0.007783839013427496, + 0.005486983340233564, + 0.017253771424293518, + 0.006959649734199047, + 0.0030637478921562433, + 0.004750498570501804, + 0.006928981747478247, + 0.012522715143859386, + 0.03006746433675289, + 0.034594327211380005, + 0.16505324840545654, + 0.0428876057267189, + 0.01878877356648445, + 0.013719490729272366, + 0.014884519390761852, + 0.06317856907844543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.023288540542125702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up focuses intently on the 5.25-inch drive bay of a black brushed metal desktop computer tower. As the drive's front panel door carefully slides open from the inside, a quiet internal whirring begins the operation. Then, with mechanical precision, a matte black plastic tray appears, sliding out smoothly and deliberately on its internal track. As it extends into the foreground, it reveals a pristine silver disc, perfectly nestled in its circular groove. Cool, direct light from an unseen source gleams on the disc's surface, creating brilliant mirror-like highlights and casting a rainbow-like sheen that dances across its data layer. The tray slides to a soft, definitive stop, fully emerging from the computer's chassis, giving the disc a sense of finality. The atmosphere is focused, the quiet, and satisfying mechanical function capturing the device's clean, precise engineering in a moment of silent operation.", + "seed": 21270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21270_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": -0.009418194461613894, + "vlm_logprob": -0.009418194461613894, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.009418194461613894, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025230340659618378 + ], + "opposite_logprob_tokens": [ + -3.9002304077148438 + ], + "logprob_reward": -0.025230340659618378, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.025230340659618378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2511925995349884, + 0.026122350245714188, + 0.029336515814065933, + 0.02741187810897827, + 0.045082539319992065, + 0.022079894319176674, + 0.09905894845724106, + 0.039915863424539566, + 0.4253033995628357, + 0.03449602052569389 + ], + "credit_map": [ + 0.01200326532125473, + 0.02224266156554222, + 0.011353581212460995, + 0.01061217114329338, + 0.1018676832318306, + 0.08072278648614883, + 0.0060819475911557674, + 0.006308526732027531, + 0.003620244562625885, + 0.0035695882979780436, + 0.004291345831006765, + 0.00227567320689559, + 0.003367719007655978, + 0.004108755849301815, + 0.002052923198789358, + 0.0028361009899526834, + 0.004909334238618612, + 0.003749367082491517, + 0.00287115341052413, + 0.002699140226468444, + 0.0038324855268001556, + 0.006347907241433859, + 0.002450538333505392, + 0.0024765923153609037, + 0.003502742387354374, + 0.004609507042914629, + 0.004718239419162273, + 0.0023977812379598618, + 0.004509586840867996, + 0.004186857026070356, + 0.0018951492384076118, + 0.0015920158475637436, + 0.006690241862088442, + 0.0036492531653493643, + 0.005793525371700525, + 0.004567069932818413, + 0.004950535949319601, + 0.008309378288686275, + 0.0033026724122464657, + 0.007819865830242634, + 0.0034706727601587772, + 0.002303274581208825, + 0.0024377265945076942, + 0.002918495098128915, + 0.0021384572610259056, + 0.0034435184206813574, + 0.00263382401317358, + 0.0027339260559529066, + 0.01210867054760456, + 0.018223583698272705, + 0.01068024430423975, + 0.028245102614164352, + 0.006695710588246584, + 0.003635523607954383, + 0.015458952635526657, + 0.004011166747659445, + 0.005294130649417639, + 0.005912973079830408, + 0.004216368775814772, + 0.006877327337861061, + 0.0035129254683852196, + 0.004949342459440231, + 0.006151397712528706, + 0.003001404693350196, + 0.10376405715942383, + 0.02785152569413185, + 0.04280315339565277, + 0.022415202111005783, + 0.0185779370367527, + 0.02571912109851837, + 0.16141878068447113, + 0.022753624245524406, + 0.006970239803195, + 0.005443850066512823, + 0.004369718953967094, + 0.003070279723033309, + 0.0027364713605493307, + 0.003717097220942378, + 0.002219927264377475, + 0.005968436598777771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.025230340659618378 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036532822996377945 + ], + "opposite_logprob_tokens": [ + -5.628653049468994 + ], + "logprob_reward": -0.0036532822996377945, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036532822996377945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3230988681316376, + 0.04531664401292801, + 0.02145383134484291, + 0.020164046436548233, + 0.022189732640981674, + 0.03015005961060524, + 0.058366116136312485, + 0.07315348088741302, + 0.38297930359840393, + 0.02312791347503662 + ], + "credit_map": [ + 0.02242097817361355, + 0.024523479864001274, + 0.01738431677222252, + 0.013547377660870552, + 0.13889369368553162, + 0.0894157662987709, + 0.008156570605933666, + 0.008756700903177261, + 0.005213342607021332, + 0.004373790230602026, + 0.009109066799283028, + 0.003510127542540431, + 0.005364628974348307, + 0.007960840128362179, + 0.005070240702480078, + 0.004714609123766422, + 0.0023138064425438643, + 0.001823483151383698, + 0.0027402155101299286, + 0.003295064438134432, + 0.0033640346955507994, + 0.00463284645229578, + 0.0017207979690283537, + 0.0015635826857760549, + 0.0024538759607821703, + 0.0031858657021075487, + 0.003006631974130869, + 0.0022906558588147163, + 0.0025905980728566647, + 0.002564374590292573, + 0.0023933514021337032, + 0.0016786920605227351, + 0.0025011368561536074, + 0.001814553514122963, + 0.0031532940920442343, + 0.0023659865837544203, + 0.0023878137581050396, + 0.0061515336856245995, + 0.0014727715170010924, + 0.0023426436819136143, + 0.0032944311387836933, + 0.0020404045935720205, + 0.004042098764330149, + 0.0045902361162006855, + 0.0024393489584326744, + 0.007072223350405693, + 0.00333031197078526, + 0.0033410056494176388, + 0.0075807650573551655, + 0.009107477031648159, + 0.0070636882446706295, + 0.014477278105914593, + 0.003244190476834774, + 0.002523767063394189, + 0.011221143417060375, + 0.003147807205095887, + 0.010835827328264713, + 0.006536452565342188, + 0.006975121796131134, + 0.014984351582825184, + 0.006653121672570705, + 0.00575066776946187, + 0.01570221595466137, + 0.0057157231494784355, + 0.09312092512845993, + 0.025065043941140175, + 0.03854629769921303, + 0.02045430988073349, + 0.016788428649306297, + 0.023320846259593964, + 0.14550377428531647, + 0.020179659128189087, + 0.005265761632472277, + 0.0029808334074914455, + 0.002729334169998765, + 0.0019922361243516207, + 0.0021310641895979643, + 0.0024357601068913937, + 0.001954207429662347, + 0.0036387168802320957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0036532822996377945 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002824367256835103 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.002824367256835103, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002824367256835103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2677706778049469, + 0.0406530536711216, + 0.0553397461771965, + 0.025326047092676163, + 0.023039529100060463, + 0.02851829119026661, + 0.04445015266537666, + 0.19149231910705566, + 0.2998044192790985, + 0.02360576018691063 + ], + "credit_map": [ + 0.037715133279561996, + 0.01668955199420452, + 0.01692638359963894, + 0.013540000654757023, + 0.11397969722747803, + 0.05436573177576065, + 0.007645569741725922, + 0.006908608600497246, + 0.004296392202377319, + 0.003841906785964966, + 0.007707133889198303, + 0.005483055952936411, + 0.006845913361757994, + 0.005794759839773178, + 0.003692562459036708, + 0.0029913317412137985, + 0.008681525476276875, + 0.003799088066443801, + 0.002449198393151164, + 0.004507320933043957, + 0.022006303071975708, + 0.008578823879361153, + 0.003059676382690668, + 0.0022578113712370396, + 0.0024938827846199274, + 0.004011186305433512, + 0.003786935005337, + 0.002252675127238035, + 0.003928568679839373, + 0.003774441545829177, + 0.003215204691514373, + 0.001863152370788157, + 0.002298203529790044, + 0.0012214850867167115, + 0.0034218085929751396, + 0.002470593899488449, + 0.0023571497295051813, + 0.0069734081625938416, + 0.0017140224808827043, + 0.002582857385277748, + 0.005478112027049065, + 0.002009594114497304, + 0.00345449591986835, + 0.0035629156045615673, + 0.0021977201104164124, + 0.0034019884187728167, + 0.0033897245302796364, + 0.005023741163313389, + 0.007966040633618832, + 0.008872815407812595, + 0.0035509716253727674, + 0.009323387406766415, + 0.0032576725352555513, + 0.0025845314376056194, + 0.005846361164003611, + 0.0030483740847557783, + 0.027279933914542198, + 0.020083092153072357, + 0.01682533696293831, + 0.038795892149209976, + 0.02069375291466713, + 0.014453773386776447, + 0.03890114650130272, + 0.014459400437772274, + 0.0715293288230896, + 0.022428931668400764, + 0.032077889889478683, + 0.016354896128177643, + 0.01707826741039753, + 0.017426874488592148, + 0.1065143495798111, + 0.016393886879086494, + 0.0037275298964232206, + 0.0032817746978253126, + 0.002491211285814643, + 0.0022194236516952515, + 0.0023047758731991053, + 0.003533398499712348, + 0.001978948712348938, + 0.004068695940077305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.002824367256835103 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011869773734360933 + ], + "opposite_logprob_tokens": [ + -6.751186847686768 + ], + "logprob_reward": -0.0011869773734360933, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011869773734360933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2529558539390564, + 0.05073624849319458, + 0.08674958348274231, + 0.04126369208097458, + 0.033951278775930405, + 0.06051992252469063, + 0.1113271564245224, + 0.21478797495365143, + 0.07220674306154251, + 0.07550157606601715 + ], + "credit_map": [ + 0.04264654591679573, + 0.01739797741174698, + 0.021083926782011986, + 0.017403986304998398, + 0.09280891716480255, + 0.041284892708063126, + 0.011315147392451763, + 0.00901444349437952, + 0.005236679222434759, + 0.004918962717056274, + 0.011049134656786919, + 0.004386939574033022, + 0.006715575233101845, + 0.011344272643327713, + 0.004040897823870182, + 0.0030437815003097057, + 0.006674049887806177, + 0.0062613836489617825, + 0.005622998811304569, + 0.004830607213079929, + 0.035422440618276596, + 0.020988881587982178, + 0.004025149624794722, + 0.0029240718577057123, + 0.004933437332510948, + 0.007418191526085138, + 0.005562182050198317, + 0.00257040630094707, + 0.008257432840764523, + 0.0066121360287070274, + 0.0034467619843780994, + 0.0024631409905850887, + 0.004622290842235088, + 0.002446557395160198, + 0.004679595120251179, + 0.0032148624304682016, + 0.0030492153018712997, + 0.010355949401855469, + 0.0022665588185191154, + 0.003316248767077923, + 0.005989195313304663, + 0.0065762074664235115, + 0.007603627163916826, + 0.011478538624942303, + 0.006552020087838173, + 0.009924164973199368, + 0.0063384887762367725, + 0.006057674065232277, + 0.017717191949486732, + 0.021282706409692764, + 0.026079468429088593, + 0.012595752254128456, + 0.009270556271076202, + 0.00587053457275033, + 0.009945849888026714, + 0.008565087802708149, + 0.029563728719949722, + 0.021409863606095314, + 0.02693997323513031, + 0.03057088702917099, + 0.02136952616274357, + 0.02091095596551895, + 0.03121327795088291, + 0.03280973434448242, + 0.011332082562148571, + 0.007176628336310387, + 0.005981332156807184, + 0.005134434439241886, + 0.007507386151701212, + 0.006357872858643532, + 0.023823214694857597, + 0.00489378534257412, + 0.017082806676626205, + 0.006012244615703821, + 0.00806164089590311, + 0.005051294341683388, + 0.005400658585131168, + 0.009437264874577522, + 0.006848801393061876, + 0.017606856301426888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0011869773734360933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196004718542099 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014196004718542099, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196004718542099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09905196726322174, + 0.017125099897384644, + 0.11870481073856354, + 0.014900696463882923, + 0.02631642110645771, + 0.0211845301091671, + 0.04752067103981972, + 0.05993282422423363, + 0.5554488301277161, + 0.03981422632932663 + ], + "credit_map": [ + 0.008841904811561108, + 0.008039919659495354, + 0.0063894218765199184, + 0.005984121467918158, + 0.03708933666348457, + 0.026366282254457474, + 0.003857681294903159, + 0.002483284566551447, + 0.002014963421970606, + 0.0018413953948765993, + 0.002720616990700364, + 0.001958278939127922, + 0.002653435803949833, + 0.003432956989854574, + 0.0016102581284940243, + 0.0008931941702030599, + 0.017667261883616447, + 0.0075348480604588985, + 0.0031256438232958317, + 0.006086566019803286, + 0.05217406898736954, + 0.023390408605337143, + 0.005729431286454201, + 0.0029965664725750685, + 0.0021992623805999756, + 0.002287827664986253, + 0.002298683160915971, + 0.001568991458043456, + 0.0017784726805984974, + 0.0021890185307711363, + 0.0012993409764021635, + 0.0012790983309969306, + 0.002648239489644766, + 0.001288238912820816, + 0.004005936440080404, + 0.0022006940562278032, + 0.0022669262252748013, + 0.00869643222540617, + 0.0015766549622640014, + 0.0036332954186946154, + 0.004501085728406906, + 0.0014696435537189245, + 0.0025401965249329805, + 0.002594148274511099, + 0.0015043566236272454, + 0.00231441599316895, + 0.0019796157721430063, + 0.004281064961105585, + 0.006699640769511461, + 0.005398454610258341, + 0.004392984788864851, + 0.013618336990475655, + 0.0029651359654963017, + 0.0018736739875748754, + 0.009972774423658848, + 0.0025996637996286154, + 0.007333788555115461, + 0.00642336206510663, + 0.00804874300956726, + 0.00708711938932538, + 0.007733692415058613, + 0.005383728072047234, + 0.010061871260404587, + 0.007860510610044003, + 0.13436757028102875, + 0.03637520223855972, + 0.055259041488170624, + 0.030147863551974297, + 0.02421063743531704, + 0.035257842391729355, + 0.20981888473033905, + 0.03001173958182335, + 0.00525020994246006, + 0.005218356382101774, + 0.004392205737531185, + 0.004514246713370085, + 0.005124094430357218, + 0.006276151165366173, + 0.004718408919870853, + 0.0043205516412854195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.014196004718542099 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up focuses intently on the 5.25-inch drive bay of a black brushed metal desktop computer tower. As the drive's front panel door carefully slides open from the inside, a quiet internal whirring begins the operation. Then, with mechanical precision, a matte black plastic tray appears, sliding out smoothly and deliberately on its internal track. As it extends into the foreground, it reveals a pristine silver disc, perfectly nestled in its circular groove. Cool, direct light from an unseen source gleams on the disc's surface, creating brilliant mirror-like highlights and casting a rainbow-like sheen that dances across its data layer. The tray slides to a soft, definitive stop, fully emerging from the computer's chassis, giving the disc a sense of finality. The atmosphere is focused, the quiet, and satisfying mechanical function capturing the device's clean, precise engineering in a moment of silent operation.", + "seed": 21271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21271_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": -0.01119156163185835, + "vlm_logprob": -0.01119156163185835, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.01119156163185835, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01742682233452797 + ], + "opposite_logprob_tokens": [ + -4.26742696762085 + ], + "logprob_reward": -0.01742682233452797, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01742682233452797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12297078222036362, + 0.1005839928984642, + 0.04971160739660263, + 0.11088427156209946, + 0.026165129616856575, + 0.02721414715051651, + 0.2761896252632141, + 0.055667173117399216, + 0.14282949268817902, + 0.08778376132249832 + ], + "credit_map": [ + 0.02899494394659996, + 0.016915690153837204, + 0.013227858580648899, + 0.010466551408171654, + 0.011584638617932796, + 0.01595809869468212, + 0.008864386938512325, + 0.016958627849817276, + 0.011510934680700302, + 0.009618997573852539, + 0.02249433659017086, + 0.021678514778614044, + 0.006598507985472679, + 0.01104860007762909, + 0.008221960626542568, + 0.00941215455532074, + 0.0051797968335449696, + 0.004535624757409096, + 0.006074747070670128, + 0.008540972135961056, + 0.005438066087663174, + 0.005868632346391678, + 0.0062283724546432495, + 0.007845398969948292, + 0.009317317046225071, + 0.005060315597802401, + 0.031033815816044807, + 0.03447899594902992, + 0.004412773065268993, + 0.008967915549874306, + 0.008592654950916767, + 0.009020487777888775, + 0.003086010692641139, + 0.002012584125623107, + 0.0026277604047209024, + 0.003596658119931817, + 0.0032811309210956097, + 0.006364287342876196, + 0.002096739364787936, + 0.003099959110841155, + 0.004244617186486721, + 0.0021520601585507393, + 0.004819304216653109, + 0.0032470282167196274, + 0.0030709519051015377, + 0.0039023254066705704, + 0.002828541910275817, + 0.002949318615719676, + 0.04280988499522209, + 0.017931979149580002, + 0.025146495550870895, + 0.020361175760626793, + 0.03138715773820877, + 0.04846339672803879, + 0.030345162376761436, + 0.05974438786506653, + 0.005918022245168686, + 0.005474322475492954, + 0.009267812594771385, + 0.007865295745432377, + 0.007546292617917061, + 0.008263685740530491, + 0.004988630302250385, + 0.006343110930174589, + 0.02641279622912407, + 0.03004205971956253, + 0.01588127762079239, + 0.01244107075035572, + 0.023398522287607193, + 0.019442960619926453, + 0.005608383100479841, + 0.00960242748260498, + 0.019011201336979866, + 0.007722511421889067, + 0.017431948333978653, + 0.008343481458723545, + 0.007952083833515644, + 0.013405943289399147, + 0.005979917477816343, + 0.007936683483421803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.01742682233452797 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006786511745303869 + ], + "opposite_logprob_tokens": [ + -5.006786346435547 + ], + "logprob_reward": -0.006786511745303869, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006786511745303869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15136940777301788, + 0.11513394117355347, + 0.06157909706234932, + 0.30074411630630493, + 0.04041227698326111, + 0.03211333230137825, + 0.1484282910823822, + 0.09722096472978592, + 0.022967912256717682, + 0.030030695721507072 + ], + "credit_map": [ + 0.02190084010362625, + 0.01334303617477417, + 0.02259068377315998, + 0.03190820291638374, + 0.013992540538311005, + 0.01730317808687687, + 0.013727146200835705, + 0.016603771597146988, + 0.014910433441400528, + 0.010240215808153152, + 0.02480226568877697, + 0.02499726600944996, + 0.007450765464454889, + 0.013173755258321762, + 0.009319265373051167, + 0.010239969938993454, + 0.007914559915661812, + 0.004467053804546595, + 0.007901075296103954, + 0.010874368250370026, + 0.0049724578857421875, + 0.007661232724785805, + 0.007298110518604517, + 0.010490230284631252, + 0.014639957807958126, + 0.010460610501468182, + 0.11245505511760712, + 0.11082151532173157, + 0.010006760247051716, + 0.012281046248972416, + 0.015728745609521866, + 0.01435039658099413, + 0.004071461036801338, + 0.0035821266938000917, + 0.004009902477264404, + 0.005271161440759897, + 0.003950375132262707, + 0.006929575931280851, + 0.0031177187338471413, + 0.009479952044785023, + 0.003492264309898019, + 0.002956175012513995, + 0.00559893948957324, + 0.002852972596883774, + 0.0045836386270821095, + 0.004702520091086626, + 0.003041225951164961, + 0.004885592497885227, + 0.028889739885926247, + 0.013131819665431976, + 0.014042694121599197, + 0.009141131304204464, + 0.013247520662844181, + 0.022869162261486053, + 0.014607750810682774, + 0.032498445361852646, + 0.01117404643446207, + 0.009083808399736881, + 0.016674285754561424, + 0.010401208885014057, + 0.012551328167319298, + 0.015971431508660316, + 0.007742129731923342, + 0.013622713275253773, + 0.0028702192939817905, + 0.001571073429659009, + 0.004240705631673336, + 0.0025681762490421534, + 0.0025310246273875237, + 0.003881394863128662, + 0.0012953199911862612, + 0.004009994212538004, + 0.006521447561681271, + 0.0030406767036765814, + 0.006568555254489183, + 0.002764340490102768, + 0.0026117227971553802, + 0.004423200152814388, + 0.0015955576673150063, + 0.0025051923003047705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.006786511745303869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956517532467842 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956517532467842, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956517532467842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14709804952144623, + 0.11235933005809784, + 0.05205700919032097, + 0.2321062535047531, + 0.0708317831158638, + 0.1646406650543213, + 0.06721921265125275, + 0.06567582488059998, + 0.04347998648881912, + 0.044531870633363724 + ], + "credit_map": [ + 0.030074218288064003, + 0.019591566175222397, + 0.015542218461632729, + 0.018148669973015785, + 0.01484732236713171, + 0.019384481012821198, + 0.012234614230692387, + 0.01727496273815632, + 0.014391261152923107, + 0.009878243319690228, + 0.020111659541726112, + 0.01860390231013298, + 0.010470150038599968, + 0.014841857366263866, + 0.01218489557504654, + 0.01187736727297306, + 0.006859003100544214, + 0.004813584964722395, + 0.007214835379272699, + 0.009065044112503529, + 0.0037084496580064297, + 0.007616270799189806, + 0.005938026588410139, + 0.006841791793704033, + 0.02028941921889782, + 0.011842526495456696, + 0.06725628674030304, + 0.07136442512273788, + 0.009977158159017563, + 0.01787625439465046, + 0.015768876299262047, + 0.01773129403591156, + 0.007578793913125992, + 0.006852520629763603, + 0.007566276006400585, + 0.008769143372774124, + 0.008038566447794437, + 0.01383874099701643, + 0.004652315750718117, + 0.01353542972356081, + 0.025155559182167053, + 0.016561131924390793, + 0.02883855439722538, + 0.014999393373727798, + 0.01773027516901493, + 0.02266623266041279, + 0.01888382062315941, + 0.01980569213628769, + 0.013141804374754429, + 0.0061620306223630905, + 0.00790075957775116, + 0.005850579589605331, + 0.0064139398746192455, + 0.008162781596183777, + 0.00766304275020957, + 0.011924274265766144, + 0.010242282412946224, + 0.006219583097845316, + 0.010405276902019978, + 0.005527971778064966, + 0.007100674323737621, + 0.011274359188973904, + 0.004661385901272297, + 0.010244281962513924, + 0.0034897506702691317, + 0.00527323829010129, + 0.00738708209246397, + 0.005837767384946346, + 0.005193754564970732, + 0.009012825787067413, + 0.002515799831598997, + 0.00476977089419961, + 0.011330187320709229, + 0.005171598866581917, + 0.007795081939548254, + 0.0035027347039431334, + 0.004592430777847767, + 0.005510927177965641, + 0.002747426275163889, + 0.0038814835716038942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.005956517532467842 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024986020289361477 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.0024986020289361477, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024986020289361477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1247720941901207, + 0.08583361655473709, + 0.08683766424655914, + 0.19550274312496185, + 0.14315922558307648, + 0.08512753993272781, + 0.1145768016576767, + 0.08115070313215256, + 0.046796511858701706, + 0.03624311089515686 + ], + "credit_map": [ + 0.02127082645893097, + 0.011705021373927593, + 0.0156119205057621, + 0.019435252994298935, + 0.012637120671570301, + 0.017979895696043968, + 0.0123678520321846, + 0.013764207251369953, + 0.009662670083343983, + 0.0070063285529613495, + 0.01506720669567585, + 0.015316901728510857, + 0.007432511076331139, + 0.01523390132933855, + 0.008503810502588749, + 0.007610280532389879, + 0.010178362019360065, + 0.006348190829157829, + 0.01524357683956623, + 0.014150843024253845, + 0.005825104657560587, + 0.014165576547384262, + 0.007609843742102385, + 0.013316169381141663, + 0.01252022385597229, + 0.00712267542257905, + 0.06432340294122696, + 0.06551923602819443, + 0.007399474270641804, + 0.015469267033040524, + 0.011345330625772476, + 0.011803125962615013, + 0.012317989952862263, + 0.010036181658506393, + 0.023200619965791702, + 0.021547576412558556, + 0.015474658459424973, + 0.032156236469745636, + 0.010537189431488514, + 0.017888769507408142, + 0.010400752536952496, + 0.006616645492613316, + 0.014205776154994965, + 0.0060946643352508545, + 0.00917346216738224, + 0.01445870753377676, + 0.010382166132330894, + 0.013795359991490841, + 0.018977871164679527, + 0.014195341616868973, + 0.011129933409392834, + 0.010196912102401257, + 0.01142781600356102, + 0.014877596870064735, + 0.011480344459414482, + 0.02229098603129387, + 0.01142812054604292, + 0.008486327715218067, + 0.012288911268115044, + 0.008279617875814438, + 0.009290346875786781, + 0.01244355645030737, + 0.005792733747512102, + 0.013141090981662273, + 0.009608308784663677, + 0.004590840544551611, + 0.005415917374193668, + 0.00390361319296062, + 0.008086777292191982, + 0.007490287534892559, + 0.0027402432169765234, + 0.004960525315254927, + 0.005349121056497097, + 0.002561578992754221, + 0.006450020242482424, + 0.003380940528586507, + 0.004145175218582153, + 0.0064912945963442326, + 0.0030558803118765354, + 0.004809102509170771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0024986020289361477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13662615418434143, + 0.07987650483846664, + 0.051460448652505875, + 0.2593953609466553, + 0.08506222814321518, + 0.0674012303352356, + 0.13726894557476044, + 0.07093015313148499, + 0.04459585249423981, + 0.06738318502902985 + ], + "credit_map": [ + 0.025174591690301895, + 0.019402125850319862, + 0.012544737197458744, + 0.01650545559823513, + 0.014373513869941235, + 0.01804700493812561, + 0.011799570173025131, + 0.018779151141643524, + 0.009279845282435417, + 0.005346016027033329, + 0.013038485310971737, + 0.021013090386986732, + 0.005432086065411568, + 0.007635053712874651, + 0.008953334763646126, + 0.009178581647574902, + 0.006210626102983952, + 0.004215633496642113, + 0.005887631326913834, + 0.008355485275387764, + 0.003980269655585289, + 0.008512435480952263, + 0.005423362832516432, + 0.008875001221895218, + 0.014484520070254803, + 0.007915087975561619, + 0.09767413139343262, + 0.09542565792798996, + 0.008040190674364567, + 0.010029222816228867, + 0.013948895037174225, + 0.0118776336312294, + 0.010343036614358425, + 0.006138187367469072, + 0.010417633689939976, + 0.010910410434007645, + 0.010036608204245567, + 0.015084470622241497, + 0.008204644545912743, + 0.01392722874879837, + 0.010373212397098541, + 0.0052649793215096, + 0.009102114476263523, + 0.00666392408311367, + 0.006579886190593243, + 0.011944291181862354, + 0.006512345280498266, + 0.0109604736790061, + 0.023149365559220314, + 0.019212286919355392, + 0.012801891192793846, + 0.01296141929924488, + 0.013534696772694588, + 0.017593592405319214, + 0.01273441780358553, + 0.025281263515353203, + 0.006129637360572815, + 0.006099902559071779, + 0.010778180323541164, + 0.00805004220455885, + 0.010679188184440136, + 0.01079151127487421, + 0.008455382660031319, + 0.009946302510797977, + 0.00446470407769084, + 0.004772028885781765, + 0.009019666351377964, + 0.005900956690311432, + 0.004904523957520723, + 0.008443527854979038, + 0.002606921596452594, + 0.00448352238163352, + 0.011376298032701015, + 0.005365504417568445, + 0.01162464078515768, + 0.006362502928823233, + 0.006569687742739916, + 0.012967458926141262, + 0.004961293190717697, + 0.008155793882906437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.023289354518055916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme close-up focuses intently on the 5.25-inch drive bay of a black brushed metal desktop computer tower. As the drive's front panel door carefully slides open from the inside, a quiet internal whirring begins the operation. Then, with mechanical precision, a matte black plastic tray appears, sliding out smoothly and deliberately on its internal track. As it extends into the foreground, it reveals a pristine silver disc, perfectly nestled in its circular groove. Cool, direct light from an unseen source gleams on the disc's surface, creating brilliant mirror-like highlights and casting a rainbow-like sheen that dances across its data layer. The tray slides to a soft, definitive stop, fully emerging from the computer's chassis, giving the disc a sense of finality. The atmosphere is focused, the quiet, and satisfying mechanical function capturing the device's clean, precise engineering in a moment of silent operation.", + "seed": 21272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000125/group_6/index_4363_seed_21272_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": -0.010121474065817893, + "vlm_logprob": -0.010121474065817893, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.010121474065817893, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01850905269384384 + ], + "opposite_logprob_tokens": [ + -4.3935089111328125 + ], + "logprob_reward": -0.01850905269384384, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01850905269384384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10751662403345108, + 0.11573774367570877, + 0.1164504811167717, + 0.09224075824022293, + 0.08422411233186722, + 0.08703529834747314, + 0.12062815576791763, + 0.0976380929350853, + 0.1376153975725174, + 0.040913332253694534 + ], + "credit_map": [ + 0.0071975900791585445, + 0.00672840466722846, + 0.01196153275668621, + 0.020824803039431572, + 0.010086852125823498, + 0.013396534137427807, + 0.015383184887468815, + 0.02193772792816162, + 0.005991481710225344, + 0.008321196772158146, + 0.013125088065862656, + 0.019509943202137947, + 0.008036138489842415, + 0.019029587507247925, + 0.02648978680372238, + 0.015234525315463543, + 0.014950922690331936, + 0.011754428967833519, + 0.012152434326708317, + 0.015440077520906925, + 0.00736126909032464, + 0.012849906459450722, + 0.020601404830813408, + 0.021340034902095795, + 0.008778227493166924, + 0.009794355370104313, + 0.008853623643517494, + 0.010814212262630463, + 0.01051298901438713, + 0.014159037731587887, + 0.014506821520626545, + 0.014821491204202175, + 0.008107459172606468, + 0.00706661818549037, + 0.00796394795179367, + 0.011779896914958954, + 0.008321831934154034, + 0.012146455235779285, + 0.016869578510522842, + 0.011968329548835754, + 0.007838243618607521, + 0.008729112334549427, + 0.006620657164603472, + 0.010430080816149712, + 0.016266919672489166, + 0.009744866751134396, + 0.01650046743452549, + 0.01090494729578495, + 0.01874624378979206, + 0.010253865271806717, + 0.01601950079202652, + 0.014371887780725956, + 0.020831406116485596, + 0.01918318122625351, + 0.007570834830403328, + 0.013651235960423946, + 0.010877348482608795, + 0.012934906408190727, + 0.02143312245607376, + 0.01581088826060295, + 0.009840179234743118, + 0.011692081578075886, + 0.007300129625946283, + 0.007749436888843775, + 0.040769677609205246, + 0.008926674723625183, + 0.03569062054157257, + 0.008059325627982616, + 0.0076915184035897255, + 0.016259416937828064, + 0.007838275283575058, + 0.012379881925880909, + 0.004929136019200087, + 0.005508784670382738, + 0.01168759260326624, + 0.004541534464806318, + 0.003524839412420988, + 0.004477173555642366, + 0.0032834161538630724, + 0.0029608586337417364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.01850905269384384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003240217687562108 + ], + "opposite_logprob_tokens": [ + -5.75324010848999 + ], + "logprob_reward": -0.003240217687562108, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003240217687562108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07233881950378418, + 0.3471663296222687, + 0.07657793909311295, + 0.15334099531173706, + 0.04770238697528839, + 0.07395797222852707, + 0.06297693401575089, + 0.03779888525605202, + 0.11107201129198074, + 0.017067808657884598 + ], + "credit_map": [ + 0.006618081592023373, + 0.005010080989450216, + 0.00873542483896017, + 0.013792572543025017, + 0.005826275330036879, + 0.010189376771450043, + 0.013023287057876587, + 0.009143711067736149, + 0.016163455322384834, + 0.02006876841187477, + 0.04026207700371742, + 0.06071988865733147, + 0.02435259521007538, + 0.04142655432224274, + 0.09909997135400772, + 0.045072995126247406, + 0.009848115034401417, + 0.006617622449994087, + 0.009468376636505127, + 0.00852837786078453, + 0.005367428995668888, + 0.010557251051068306, + 0.00950448028743267, + 0.01668628118932247, + 0.019290287047624588, + 0.019193101674318314, + 0.015333626419305801, + 0.01584743522107601, + 0.014717924408614635, + 0.02212652377784252, + 0.02275712601840496, + 0.024074966087937355, + 0.00835618656128645, + 0.00572924455627799, + 0.004776182118803263, + 0.006126672960817814, + 0.0031533213332295418, + 0.007851776666939259, + 0.004959461744874716, + 0.006749538704752922, + 0.009607075713574886, + 0.007893516682088375, + 0.008577571250498295, + 0.0064406064338982105, + 0.00962198805063963, + 0.012283184565603733, + 0.010935017839074135, + 0.008599001914262772, + 0.0075873988680541515, + 0.01111514214426279, + 0.0074029709212481976, + 0.006739022210240364, + 0.00823393277823925, + 0.011702743358910084, + 0.0036699401680380106, + 0.006525777745991945, + 0.005401182919740677, + 0.0034339125268161297, + 0.008994623087346554, + 0.003933657426387072, + 0.003991537727415562, + 0.005196675658226013, + 0.002930012997239828, + 0.003917280118912458, + 0.029139703139662743, + 0.008837827481329441, + 0.01693269982933998, + 0.0086955102160573, + 0.008255688473582268, + 0.017803438007831573, + 0.007448643911629915, + 0.013958488591015339, + 0.002448915271088481, + 0.0032656230032444, + 0.002943538362160325, + 0.0016425352077931166, + 0.0017090961337089539, + 0.001971415476873517, + 0.0015060233417898417, + 0.0015806615119799972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.003240217687562108 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07839785516262054, + 0.1558454930782318, + 0.06719930469989777, + 0.060293056070804596, + 0.04911550134420395, + 0.07355152070522308, + 0.08763255923986435, + 0.12262679636478424, + 0.16416332125663757, + 0.1411745548248291 + ], + "credit_map": [ + 0.006851544138044119, + 0.004696134943515062, + 0.009540152736008167, + 0.01655273512005806, + 0.006724274717271328, + 0.008139186538755894, + 0.012661132030189037, + 0.013232692144811153, + 0.00958770141005516, + 0.008519023656845093, + 0.017917199060320854, + 0.026673536747694016, + 0.009950729086995125, + 0.019883401691913605, + 0.043279051780700684, + 0.02003484219312668, + 0.00597189599648118, + 0.00405585253611207, + 0.011546923778951168, + 0.009196081198751926, + 0.0028667435981333256, + 0.008012625388801098, + 0.00799628533422947, + 0.017552901059389114, + 0.010714768432080746, + 0.006592819467186928, + 0.004292996600270271, + 0.009372488595545292, + 0.005479313433170319, + 0.008526178076863289, + 0.006347601767629385, + 0.008966886438429356, + 0.006285460200160742, + 0.0037218164652585983, + 0.003919564187526703, + 0.0072161429561674595, + 0.003692036960273981, + 0.009413511492311954, + 0.00790359452366829, + 0.006963369902223349, + 0.00736598577350378, + 0.006426522508263588, + 0.008176818490028381, + 0.008683143183588982, + 0.013205698691308498, + 0.009777083992958069, + 0.011356111615896225, + 0.00856015458703041, + 0.00986496452242136, + 0.012138931080698967, + 0.014219354838132858, + 0.011122592724859715, + 0.011774864047765732, + 0.013401687145233154, + 0.006074383854866028, + 0.009035774506628513, + 0.014351564459502697, + 0.011488780379295349, + 0.027373461052775383, + 0.02069803699851036, + 0.011065968312323093, + 0.015674665570259094, + 0.010004104115068913, + 0.011970218271017075, + 0.04278101772069931, + 0.013476444408297539, + 0.016943372786045074, + 0.013234207406640053, + 0.012600169517099857, + 0.02612185850739479, + 0.010504799894988537, + 0.02850145660340786, + 0.015001161023974419, + 0.015687787905335426, + 0.026801159605383873, + 0.02008737437427044, + 0.014424526132643223, + 0.021574532613158226, + 0.014466309919953346, + 0.013131695799529552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019481267081573606 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.0019481267081573606, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019481267081573606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11278554797172546, + 0.09717874228954315, + 0.08296085149049759, + 0.04837609454989433, + 0.031880881637334824, + 0.1209946870803833, + 0.09532755613327026, + 0.10997145622968674, + 0.1325526237487793, + 0.16797161102294922 + ], + "credit_map": [ + 0.008166678249835968, + 0.006444346625357866, + 0.015330979600548744, + 0.02950379252433777, + 0.00812455452978611, + 0.013572345487773418, + 0.01804259605705738, + 0.013600258156657219, + 0.006770523730665445, + 0.00654782447963953, + 0.01371752843260765, + 0.019935546442866325, + 0.006446063984185457, + 0.010258195921778679, + 0.02219347469508648, + 0.011309589259326458, + 0.01205262541770935, + 0.007258134428411722, + 0.0137647595256567, + 0.009592522867023945, + 0.0043850005604326725, + 0.008387953974306583, + 0.008822734467685223, + 0.018697118386626244, + 0.00485552242025733, + 0.004406575113534927, + 0.005755478981882334, + 0.005943820811808109, + 0.005484497174620628, + 0.006945269647985697, + 0.007268384099006653, + 0.007716545835137367, + 0.004872404970228672, + 0.0031671207398176193, + 0.0028644551057368517, + 0.00406153267249465, + 0.0025728221517056227, + 0.004978634417057037, + 0.004685996100306511, + 0.004677911289036274, + 0.015515316277742386, + 0.009648546576499939, + 0.017114901915192604, + 0.011270325630903244, + 0.014517595991492271, + 0.019277501851320267, + 0.01744329184293747, + 0.016207214444875717, + 0.008756151422858238, + 0.010331925936043262, + 0.01254140306264162, + 0.013632665388286114, + 0.011962403543293476, + 0.023087428882718086, + 0.005920708179473877, + 0.009094872511923313, + 0.010184039361774921, + 0.010231384076178074, + 0.020126337185502052, + 0.023923330008983612, + 0.009470106102526188, + 0.019654370844364166, + 0.00794144719839096, + 0.008440453559160233, + 0.02772732824087143, + 0.011055129580199718, + 0.019206413999199867, + 0.010246971622109413, + 0.009415694512426853, + 0.019990069791674614, + 0.009322315454483032, + 0.02558869868516922, + 0.02954675629734993, + 0.03185182437300682, + 0.031692974269390106, + 0.015990646556019783, + 0.01611301861703396, + 0.018364666029810905, + 0.010573720559477806, + 0.013838008046150208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.0019481267081573606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static extreme close-up of a black brushed metal desktop computer drive bay?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a matte black plastic tray slide out smoothly from the drive bay on an internal track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pristine silver disc visible nestled in the circular groove of the tray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the silver disc display brilliant mirror-like highlights and a rainbow-like sheen on its data layer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the tray slide to a soft, definitive stop fully emerging from the computer's chassis?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.143804132938385, + 0.1079070046544075, + 0.08550238609313965, + 0.10342546552419662, + 0.044471997767686844, + 0.17430490255355835, + 0.09404058009386063, + 0.07480040937662125, + 0.1129549890756607, + 0.05878821387887001 + ], + "credit_map": [ + 0.011218822561204433, + 0.008180155418813229, + 0.01919766701757908, + 0.038505204021930695, + 0.01057546678930521, + 0.017646444961428642, + 0.019717883318662643, + 0.01876247487962246, + 0.005647784564644098, + 0.006989074405282736, + 0.013525479473173618, + 0.01988169550895691, + 0.007379206363111734, + 0.012653220444917679, + 0.02743595279753208, + 0.01439458318054676, + 0.008237900212407112, + 0.006121044047176838, + 0.00930207222700119, + 0.012856118381023407, + 0.004568767733871937, + 0.010781961493194103, + 0.017971590161323547, + 0.015662923455238342, + 0.01914072036743164, + 0.009548782370984554, + 0.009259887039661407, + 0.017455479130148888, + 0.008043809793889523, + 0.01036681979894638, + 0.013045940548181534, + 0.016564015299081802, + 0.006343673914670944, + 0.00275817746296525, + 0.005873647518455982, + 0.006718399468809366, + 0.0023762835189700127, + 0.007845940999686718, + 0.006753080990165472, + 0.005802787374705076, + 0.017532696947455406, + 0.020689338445663452, + 0.021386153995990753, + 0.021268634125590324, + 0.02128549851477146, + 0.028673840686678886, + 0.021497933194041252, + 0.02197078987956047, + 0.010208833031356335, + 0.015423421747982502, + 0.01257224939763546, + 0.009529637172818184, + 0.008589345961809158, + 0.01908402517437935, + 0.005885286256670952, + 0.012747778557240963, + 0.01142383087426424, + 0.009421128779649734, + 0.011404219083487988, + 0.008106096647679806, + 0.008776064030826092, + 0.011558763682842255, + 0.0056746602058410645, + 0.008435637690126896, + 0.025540417060256004, + 0.010006951168179512, + 0.010467387735843658, + 0.009937146678566933, + 0.008242810145020485, + 0.016169453039765358, + 0.007178956177085638, + 0.02541186846792698, + 0.009603552520275116, + 0.01192128099501133, + 0.007418159861117601, + 0.007384380325675011, + 0.005874570459127426, + 0.0076395059004426, + 0.004618347622454166, + 0.004328411538153887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.023289354518055916 + } + ], + "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/0000152/group_0/index_tensor([2212, 2212])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_tensor([2212, 2212])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e31cc87d37f3fa2ac883c926d8efc40a0125601 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_tensor([2212, 2212])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up time-lapse shot, shot from an extremely low angle, captures a beam of warm golden sunlight filtering through an old window, slanting across the weathered, dark brown wooden floor. Motes of dust dancing in the air are clearly visible in the beam, creating a beautiful Tyndall effect. From the varying widths of the floorboards, several slender, tender green vines cautiously poke their heads, their newly formed, microscopically hairy tendrils appearing translucent in the light. As time passes, these vines ascend with a steady and leisurely rhythm, their heart-shaped leaves slowly unfurling in the light, their emerald green appearing strikingly vibrant against the dark, scratched wood flooring. Tensors tentatively dance in the air before gracefully grasping the rough surface of the wood, winding their way along the distinct grain toward the center of the spot of light. The entire scene, shrouded in profound tranquility, evokes a tenacious vitality, creating a magical realist atmosphere of time condensed and life quietly reawakening.", + "seed": 5570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_2212_seed_5570_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": -0.011132179573178292, + "vlm_logprob": -0.011132179573178292, + "vlm_yes_no_margin": 4.874999952316284, + "vlm_token_credit": -0.011132179573178292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01957409828901291 + ], + "opposite_logprob_tokens": [ + -4.644574165344238 + ], + "logprob_reward": -0.01957409828901291, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01957409828901291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23365405201911926, + 0.1076958104968071, + 0.039073411375284195, + 0.030960464850068092, + 0.1452566236257553, + 0.04391578584909439, + 0.11093258112668991, + 0.06905770301818848, + 0.15209916234016418, + 0.06735436618328094 + ], + "credit_map": [ + 0.03120599314570427, + 0.019016703590750694, + 0.0277496799826622, + 0.031571730971336365, + 0.03076702170073986, + 0.03542184457182884, + 0.023093920201063156, + 0.034827157855033875, + 0.015029625035822392, + 0.009535012766718864, + 0.0112116988748312, + 0.013896383345127106, + 0.010987324640154839, + 0.014598052017390728, + 0.013128900900483131, + 0.01930881105363369, + 0.004584854934364557, + 0.0036598513834178448, + 0.0055009713396430016, + 0.005228057969361544, + 0.004412781912833452, + 0.005201107356697321, + 0.004653518088161945, + 0.005832266993820667, + 0.0043933275155723095, + 0.0028396020643413067, + 0.0029724747873842716, + 0.003105130745097995, + 0.00509835546836257, + 0.004435933660715818, + 0.004316733218729496, + 0.0037989073898643255, + 0.027039406821131706, + 0.015035353600978851, + 0.02253412827849388, + 0.016397908329963684, + 0.017069416120648384, + 0.019831014797091484, + 0.013793076388537884, + 0.013556321151554585, + 0.007670714519917965, + 0.005944743286818266, + 0.004537382163107395, + 0.0036698004696518183, + 0.0044986107386648655, + 0.004970352165400982, + 0.006207972764968872, + 0.006416205316781998, + 0.015092038549482822, + 0.008196141570806503, + 0.012261830270290375, + 0.013566379435360432, + 0.01554542314261198, + 0.015793820843100548, + 0.01534265372902155, + 0.015134287998080254, + 0.017569445073604584, + 0.005757453385740519, + 0.0070367916487157345, + 0.007276521064341068, + 0.006663408596068621, + 0.0075920345261693, + 0.00869330670684576, + 0.008468742482364178, + 0.01990826614201069, + 0.02601967379450798, + 0.017334921285510063, + 0.02028067223727703, + 0.022719373926520348, + 0.0171168502420187, + 0.0145817119628191, + 0.014137694612145424, + 0.011074391193687916, + 0.005225614178925753, + 0.006273002363741398, + 0.006960300263017416, + 0.009223832748830318, + 0.0090893330052495, + 0.01208215020596981, + 0.007425739895552397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.01957409828901291 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807968512177467 + ], + "opposite_logprob_tokens": [ + -4.63480806350708 + ], + "logprob_reward": -0.009807968512177467, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807968512177467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3905782699584961, + 0.07589955627918243, + 0.10526622831821442, + 0.050482284277677536, + 0.13011868298053741, + 0.03988980874419212, + 0.052294280380010605, + 0.024231702089309692, + 0.10488132387399673, + 0.026357758790254593 + ], + "credit_map": [ + 0.04883389547467232, + 0.03643930330872536, + 0.048687893897295, + 0.05829282104969025, + 0.056549377739429474, + 0.0626097247004509, + 0.033994656056165695, + 0.045170653611421585, + 0.01060024369508028, + 0.009524333290755749, + 0.00825533177703619, + 0.008823554962873459, + 0.009166853502392769, + 0.011264371685683727, + 0.007809706963598728, + 0.010455169714987278, + 0.01051822118461132, + 0.009470307268202305, + 0.014513526111841202, + 0.012385078705847263, + 0.015849795192480087, + 0.015930771827697754, + 0.012022502720355988, + 0.014576032757759094, + 0.007217000238597393, + 0.003478370141237974, + 0.00632427167147398, + 0.005644951015710831, + 0.004244417417794466, + 0.007706178817898035, + 0.005978261586278677, + 0.00988883525133133, + 0.01993638649582863, + 0.015543208457529545, + 0.01800340786576271, + 0.010781227611005306, + 0.01543622836470604, + 0.01903187856078148, + 0.01733028143644333, + 0.014056075364351273, + 0.0059339553117752075, + 0.003920774906873703, + 0.004021381493657827, + 0.003372118342667818, + 0.004966518376022577, + 0.005305515602231026, + 0.006045423448085785, + 0.006324121728539467, + 0.006996974814683199, + 0.0049249520525336266, + 0.006329685915261507, + 0.0055483002215623856, + 0.008459072560071945, + 0.006530129350721836, + 0.005662590730935335, + 0.00784258171916008, + 0.003679183777421713, + 0.0018048376077786088, + 0.0022462750785052776, + 0.004461587406694889, + 0.0015846493188291788, + 0.0027377712540328503, + 0.005239520221948624, + 0.002477878937497735, + 0.016821851953864098, + 0.023847997188568115, + 0.010664149187505245, + 0.011880758218467236, + 0.012529134750366211, + 0.010368389077484608, + 0.010262099094688892, + 0.008506955578923225, + 0.004011669661849737, + 0.003081952454522252, + 0.003393806517124176, + 0.0036585440393537283, + 0.004243981558829546, + 0.002499272581189871, + 0.0027696541510522366, + 0.002698881085962057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.009807968512177467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002830191981047392 + ], + "opposite_logprob_tokens": [ + -5.8778300285339355 + ], + "logprob_reward": -0.002830191981047392, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002830191981047392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11460261046886444, + 0.07324641197919846, + 0.11254417896270752, + 0.09872641414403915, + 0.13586273789405823, + 0.08159004896879196, + 0.10174670070409775, + 0.10022301226854324, + 0.10428471118211746, + 0.07717324048280716 + ], + "credit_map": [ + 0.016451530158519745, + 0.01109558716416359, + 0.014643717557191849, + 0.012843203730881214, + 0.015969227999448776, + 0.01907404698431492, + 0.010328360833227634, + 0.014196924865245819, + 0.00853114202618599, + 0.007167678792029619, + 0.006953653413802385, + 0.007789005991071463, + 0.011667282320559025, + 0.011586476117372513, + 0.007194650825113058, + 0.012356518767774105, + 0.008900132961571217, + 0.010067719034850597, + 0.011898522265255451, + 0.010782735422253609, + 0.022053513675928116, + 0.016761548817157745, + 0.01461880560964346, + 0.017461184412240982, + 0.010893190279603004, + 0.010248851031064987, + 0.010249285027384758, + 0.011150091886520386, + 0.01317576877772808, + 0.011942383833229542, + 0.017350424081087112, + 0.013716407120227814, + 0.029278350993990898, + 0.018291838467121124, + 0.015369964763522148, + 0.01045357808470726, + 0.018712036311626434, + 0.016898397356271744, + 0.011976531706750393, + 0.01488201692700386, + 0.009916551411151886, + 0.01038266345858574, + 0.00885595753788948, + 0.008061888627707958, + 0.011333481408655643, + 0.010170542635023594, + 0.010661055333912373, + 0.012207902036607265, + 0.009870537556707859, + 0.010096149519085884, + 0.011917655356228352, + 0.010253299959003925, + 0.013777950778603554, + 0.016188936308026314, + 0.010920317843556404, + 0.018721843138337135, + 0.015583094209432602, + 0.006664345506578684, + 0.009065239690244198, + 0.01920846849679947, + 0.007351462729275227, + 0.011071481741964817, + 0.02083304524421692, + 0.010445864871144295, + 0.01675673946738243, + 0.024136167019605637, + 0.009172026999294758, + 0.009779282845556736, + 0.012269308790564537, + 0.009753752499818802, + 0.014649233780801296, + 0.0077681890688836575, + 0.01486988365650177, + 0.009213545359671116, + 0.008848401717841625, + 0.006639974657446146, + 0.015219919383525848, + 0.007755340542644262, + 0.007330555468797684, + 0.007295616436749697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.002830191981047392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00526660168543458 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.00526660168543458, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00526660168543458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22398819029331207, + 0.11439810693264008, + 0.057740386575460434, + 0.09793689101934433, + 0.06707262247800827, + 0.045031819492578506, + 0.11270400136709213, + 0.07228917628526688, + 0.123105488717556, + 0.08573324233293533 + ], + "credit_map": [ + 0.03044055588543415, + 0.022066442295908928, + 0.024283071979880333, + 0.02969716675579548, + 0.03317561373114586, + 0.03632167726755142, + 0.02331911399960518, + 0.024684565141797066, + 0.010809436440467834, + 0.008960562758147717, + 0.010321776382625103, + 0.014536475762724876, + 0.014871620573103428, + 0.01997263915836811, + 0.01349133811891079, + 0.021434275433421135, + 0.006733283866196871, + 0.005118709523230791, + 0.007131179329007864, + 0.0061837234534323215, + 0.008735958486795425, + 0.008667116053402424, + 0.006053013261407614, + 0.00911740679293871, + 0.010730167850852013, + 0.004919824190437794, + 0.008107449859380722, + 0.013137230649590492, + 0.009970367886126041, + 0.015894858166575432, + 0.021907731890678406, + 0.013269265182316303, + 0.006179948337376118, + 0.006102345883846283, + 0.012428294867277145, + 0.009653964079916477, + 0.00793781690299511, + 0.008393114432692528, + 0.007176666986197233, + 0.009200481697916985, + 0.007128280587494373, + 0.0051679653115570545, + 0.005368908401578665, + 0.0030127528589218855, + 0.006716887000948191, + 0.004870596807450056, + 0.006881051696836948, + 0.005885382182896137, + 0.012376799248158932, + 0.011161011643707752, + 0.015489454381167889, + 0.011935139074921608, + 0.017460526898503304, + 0.010076737031340599, + 0.014491260051727295, + 0.019713087007403374, + 0.007598663680255413, + 0.004931122064590454, + 0.007132909260690212, + 0.014670205302536488, + 0.004585613030940294, + 0.008818324655294418, + 0.016387032344937325, + 0.008165312930941582, + 0.020037582144141197, + 0.026736605912446976, + 0.011983579024672508, + 0.015250456519424915, + 0.016260642558336258, + 0.011990314349532127, + 0.012661697342991829, + 0.008184624835848808, + 0.015563884750008583, + 0.010202521458268166, + 0.009681331925094128, + 0.008413447998464108, + 0.014650579541921616, + 0.008569118566811085, + 0.009349656291306019, + 0.009302712976932526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00526660168543458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818203739821911 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.01818203739821911, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818203739821911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1900104284286499, + 0.0871032327413559, + 0.1339959353208542, + 0.05172095075249672, + 0.15719878673553467, + 0.08027617633342743, + 0.07375900447368622, + 0.10205075144767761, + 0.047742947936058044, + 0.07614180445671082 + ], + "credit_map": [ + 0.022443754598498344, + 0.01601838320493698, + 0.03142310678958893, + 0.016286849975585938, + 0.031492073088884354, + 0.030440477654337883, + 0.016411280259490013, + 0.025494489818811417, + 0.009428352117538452, + 0.00863821804523468, + 0.010054469108581543, + 0.009430862963199615, + 0.014668135903775692, + 0.014443724416196346, + 0.009550460614264011, + 0.010889005847275257, + 0.008858619257807732, + 0.009313728660345078, + 0.01380793284624815, + 0.013800780288875103, + 0.015229390002787113, + 0.021966341882944107, + 0.01884380914270878, + 0.03217533230781555, + 0.006250317208468914, + 0.0049333772622048855, + 0.0062371594831347466, + 0.005673115607351065, + 0.005754288751631975, + 0.0076219383627176285, + 0.006694415118545294, + 0.008556335233151913, + 0.019751740619540215, + 0.016440248116850853, + 0.024883728474378586, + 0.017356878146529198, + 0.01994161121547222, + 0.02252124808728695, + 0.02074282057583332, + 0.015560493804514408, + 0.010040337219834328, + 0.010804802179336548, + 0.007768585812300444, + 0.00507871899753809, + 0.02081982232630253, + 0.010180296376347542, + 0.006756210699677467, + 0.008827392011880875, + 0.012345811352133751, + 0.007567275315523148, + 0.008816802874207497, + 0.007365516386926174, + 0.008368043228983879, + 0.009027674794197083, + 0.009770804084837437, + 0.010497069917619228, + 0.01955057866871357, + 0.009534931741654873, + 0.011796711944043636, + 0.01466264110058546, + 0.012751893140375614, + 0.011415100656449795, + 0.011781674809753895, + 0.01055721566081047, + 0.006203091237694025, + 0.006095288321375847, + 0.006176408845931292, + 0.004892538767307997, + 0.005142695736140013, + 0.006905948277562857, + 0.006337578408420086, + 0.005989396944642067, + 0.013375422917306423, + 0.008664135821163654, + 0.007644230034202337, + 0.00657267589122057, + 0.013142879121005535, + 0.00826989021152258, + 0.01020121481269598, + 0.008271349593997002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.01818203739821911 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, shot from an extremely low angle, captures a beam of warm golden sunlight filtering through an old window, slanting across the weathered, dark brown wooden floor. Motes of dust dancing in the air are clearly visible in the beam, creating a beautiful Tyndall effect. From the varying widths of the floorboards, several slender, tender green vines cautiously poke their heads, their newly formed, microscopically hairy tendrils appearing translucent in the light. As time passes, these vines ascend with a steady and leisurely rhythm, their heart-shaped leaves slowly unfurling in the light, their emerald green appearing strikingly vibrant against the dark, scratched wood flooring. Tensors tentatively dance in the air before gracefully grasping the rough surface of the wood, winding their way along the distinct grain toward the center of the spot of light. The entire scene, shrouded in profound tranquility, evokes a tenacious vitality, creating a magical realist atmosphere of time condensed and life quietly reawakening.", + "seed": 5571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_2212_seed_5571_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": -0.013740606675855815, + "vlm_logprob": -0.013740606675855815, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.013740606675855815, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025104328989982605 + ], + "opposite_logprob_tokens": [ + -4.400104522705078 + ], + "logprob_reward": -0.025104328989982605, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.025104328989982605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2355809211730957, + 0.08806798607110977, + 0.06908953934907913, + 0.08063255250453949, + 0.07885923981666565, + 0.0987953469157219, + 0.11182980984449387, + 0.07207434624433517, + 0.07657492905855179, + 0.08849529176950455 + ], + "credit_map": [ + 0.04718609154224396, + 0.03151809796690941, + 0.019863957539200783, + 0.031622692942619324, + 0.026237282902002335, + 0.016897516325116158, + 0.024944016709923744, + 0.03731127455830574, + 0.015201481059193611, + 0.015998370945453644, + 0.00710270507261157, + 0.014635832980275154, + 0.010884943418204784, + 0.006454473827034235, + 0.008328096009790897, + 0.00946208368986845, + 0.014956780709326267, + 0.009633352980017662, + 0.005047709681093693, + 0.008895047008991241, + 0.00649263896048069, + 0.0053880284540355206, + 0.00866104569286108, + 0.010014939121901989, + 0.01353033073246479, + 0.014122157357633114, + 0.009081610478460789, + 0.009530842304229736, + 0.013213147409260273, + 0.00700446916744113, + 0.005944408476352692, + 0.008205592632293701, + 0.013659733347594738, + 0.01197490468621254, + 0.005927365738898516, + 0.0106884203851223, + 0.007848065346479416, + 0.006101259030401707, + 0.010283246636390686, + 0.012376246973872185, + 0.02756420709192753, + 0.01126913819462061, + 0.006453217472881079, + 0.005414366722106934, + 0.00681301299482584, + 0.004784349352121353, + 0.007341850083321333, + 0.029155204072594643, + 0.01660292223095894, + 0.018240252509713173, + 0.009305230341851711, + 0.014437245205044746, + 0.012946859002113342, + 0.0071997991763055325, + 0.012505115009844303, + 0.02059238590300083, + 0.01019740104675293, + 0.00955476239323616, + 0.008080597966909409, + 0.009815381839871407, + 0.012745794840157032, + 0.0048568351194262505, + 0.0051881056278944016, + 0.011635465547442436, + 0.009576651267707348, + 0.008464080281555653, + 0.009069547057151794, + 0.015312778763473034, + 0.00922552403062582, + 0.00609123520553112, + 0.00728877866640687, + 0.011546333320438862, + 0.011136413551867008, + 0.011438082903623581, + 0.009148295037448406, + 0.010396144352853298, + 0.016134187579154968, + 0.00700001697987318, + 0.010705624707043171, + 0.012536531314253807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.025104328989982605 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007655567489564419 + ], + "opposite_logprob_tokens": [ + -4.882655620574951 + ], + "logprob_reward": -0.007655567489564419, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007655567489564419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1586691290140152, + 0.06909916549921036, + 0.10351737588644028, + 0.06339580565690994, + 0.0621914267539978, + 0.12911295890808105, + 0.058936286717653275, + 0.06482849270105362, + 0.19350017607212067, + 0.0967492163181305 + ], + "credit_map": [ + 0.0324794203042984, + 0.018080158159136772, + 0.017173651605844498, + 0.012950288131833076, + 0.013685914687812328, + 0.016496866941452026, + 0.015288266353309155, + 0.03251456469297409, + 0.013653996400535107, + 0.009860345162451267, + 0.007377025205641985, + 0.00958788488060236, + 0.0061058467254042625, + 0.0053304205648601055, + 0.006567166652530432, + 0.010616477578878403, + 0.025201162323355675, + 0.014472433365881443, + 0.008412267081439495, + 0.010297316126525402, + 0.007909245789051056, + 0.010922860354185104, + 0.008194509893655777, + 0.018107593059539795, + 0.010427724570035934, + 0.010466168634593487, + 0.0034910710528492928, + 0.0056920466013252735, + 0.011745250783860683, + 0.004744508769363165, + 0.00766347162425518, + 0.009165564551949501, + 0.011528481729328632, + 0.0055171651765704155, + 0.004366391804069281, + 0.006763205397874117, + 0.006495270878076553, + 0.005593909882009029, + 0.007083387114107609, + 0.01484361570328474, + 0.037618257105350494, + 0.014025768265128136, + 0.009371922351419926, + 0.006962521467357874, + 0.008897479623556137, + 0.006170453038066626, + 0.012102829292416573, + 0.033963728696107864, + 0.009027023799717426, + 0.0050437068566679955, + 0.007391748018562794, + 0.008736017160117626, + 0.0038865685928612947, + 0.005405060946941376, + 0.008949876762926579, + 0.010496286675333977, + 0.007839486934244633, + 0.004391684662550688, + 0.011084390804171562, + 0.013552853837609291, + 0.006683710962533951, + 0.004256263375282288, + 0.005858793389052153, + 0.01116130780428648, + 0.022762546315789223, + 0.007982095703482628, + 0.056590598076581955, + 0.03187445551156998, + 0.008062873035669327, + 0.010322642512619495, + 0.008853729814291, + 0.047051239758729935, + 0.011221399530768394, + 0.009455710649490356, + 0.011366921477019787, + 0.012577428482472897, + 0.017123615369200706, + 0.00718621676787734, + 0.013222930021584034, + 0.014594990760087967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.007655567489564419 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019515770254656672 + ], + "opposite_logprob_tokens": [ + -6.251951694488525 + ], + "logprob_reward": -0.0019515770254656672, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019515770254656672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1635059416294098, + 0.06313173472881317, + 0.0759643092751503, + 0.08273962885141373, + 0.09061599522829056, + 0.15956489741802216, + 0.1665739268064499, + 0.03545822948217392, + 0.0769314393401146, + 0.08551390469074249 + ], + "credit_map": [ + 0.027980728074908257, + 0.02038642391562462, + 0.017467539757490158, + 0.014677557162940502, + 0.017757762223482132, + 0.014323352836072445, + 0.016059592366218567, + 0.034852977842092514, + 0.011424358934164047, + 0.007261771243065596, + 0.008833966217935085, + 0.008515928871929646, + 0.0063775586895644665, + 0.006026159040629864, + 0.006058759056031704, + 0.008633231744170189, + 0.016121409833431244, + 0.011712896637618542, + 0.00508709205314517, + 0.008451015688478947, + 0.00908531341701746, + 0.006410672329366207, + 0.007035515736788511, + 0.012060397304594517, + 0.014691703021526337, + 0.012423011474311352, + 0.0064917271956801414, + 0.009653129614889622, + 0.012826070189476013, + 0.007322902325540781, + 0.008080748841166496, + 0.011250331997871399, + 0.016037747263908386, + 0.011389725841581821, + 0.008516225963830948, + 0.010254595428705215, + 0.01061673741787672, + 0.006932257674634457, + 0.012943415902554989, + 0.013925290666520596, + 0.025901544839143753, + 0.02458292804658413, + 0.01073214691132307, + 0.015747083351016045, + 0.019878748804330826, + 0.012290087528526783, + 0.02606375701725483, + 0.024368587881326675, + 0.030041400343179703, + 0.015915052965283394, + 0.010829154402017593, + 0.020035041496157646, + 0.011448286473751068, + 0.014583379030227661, + 0.022352905943989754, + 0.04136871546506882, + 0.004763589706271887, + 0.005867780186235905, + 0.004081215709447861, + 0.007072200067341328, + 0.004891613032668829, + 0.0017482577823102474, + 0.003261447884142399, + 0.003772126045078039, + 0.010595215484499931, + 0.0071992999874055386, + 0.011965620331466198, + 0.01304549165070057, + 0.008495687507092953, + 0.004567774478346109, + 0.0051339189521968365, + 0.015928439795970917, + 0.012084512040019035, + 0.008766911923885345, + 0.009210399352014065, + 0.01242125779390335, + 0.012188640423119068, + 0.0062508247792720795, + 0.009828823618590832, + 0.014762531034648418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0019515770254656672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650125306099653 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007650125306099653, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650125306099653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08334793150424957, + 0.05919381231069565, + 0.07991363853216171, + 0.04817141592502594, + 0.044972002506256104, + 0.26534947752952576, + 0.07637687772512436, + 0.08586033433675766, + 0.13869164884090424, + 0.11812286078929901 + ], + "credit_map": [ + 0.015193485654890537, + 0.01064271293580532, + 0.009258083067834377, + 0.01247691735625267, + 0.00927259586751461, + 0.0059288605116307735, + 0.008257854729890823, + 0.01231742836534977, + 0.00845723319798708, + 0.006316299084573984, + 0.006602083798497915, + 0.012041764333844185, + 0.006777469534426928, + 0.004325175192207098, + 0.007410395424813032, + 0.007263391278684139, + 0.014754274860024452, + 0.011866465210914612, + 0.007138559594750404, + 0.011102370917797089, + 0.010329830460250378, + 0.007599798496812582, + 0.008758198469877243, + 0.008364141918718815, + 0.008775724098086357, + 0.0077130612917244434, + 0.0035469692666083574, + 0.004987968131899834, + 0.006685046013444662, + 0.004409464076161385, + 0.004743256606161594, + 0.007309924811124802, + 0.009803138673305511, + 0.007616132497787476, + 0.0032786084339022636, + 0.004027069080621004, + 0.006188763305544853, + 0.002855046186596155, + 0.004540469963103533, + 0.006662772037088871, + 0.08201945573091507, + 0.029939966276288033, + 0.01681738719344139, + 0.012558233924210072, + 0.01983029581606388, + 0.011727698147296906, + 0.01691616140305996, + 0.07554028183221817, + 0.012891722843050957, + 0.01354296412318945, + 0.005946140270680189, + 0.007394193205982447, + 0.009228731505572796, + 0.004501974210143089, + 0.010056354105472565, + 0.012814793735742569, + 0.012843932025134563, + 0.010148966684937477, + 0.011834678240120411, + 0.008265186101198196, + 0.011771930381655693, + 0.005280997604131699, + 0.007720175664871931, + 0.017994478344917297, + 0.019472120329737663, + 0.007734270300716162, + 0.03422658145427704, + 0.020851023495197296, + 0.007589475251734257, + 0.008310276083648205, + 0.0066218627616763115, + 0.03388604521751404, + 0.013032489456236362, + 0.012725631706416607, + 0.014335167594254017, + 0.01728292740881443, + 0.020462874323129654, + 0.008203206583857536, + 0.01402769610285759, + 0.01805286481976509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.007650125306099653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341434568166733 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341434568166733, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341434568166733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12110240757465363, + 0.06539543718099594, + 0.0697084590792656, + 0.07076551765203476, + 0.06780469417572021, + 0.18966533243656158, + 0.10604967176914215, + 0.0603751614689827, + 0.0815465971827507, + 0.16758666932582855 + ], + "credit_map": [ + 0.023095788434147835, + 0.01319964975118637, + 0.013797810301184654, + 0.01526723150163889, + 0.010793008841574192, + 0.009840283542871475, + 0.012385215610265732, + 0.022723427042365074, + 0.014166421256959438, + 0.010352824814617634, + 0.004290707875043154, + 0.00969801377505064, + 0.007261874619871378, + 0.0043223826214671135, + 0.0067349267192184925, + 0.00856828410178423, + 0.012564525008201599, + 0.006388199049979448, + 0.007938365451991558, + 0.008204811252653599, + 0.006838537752628326, + 0.005313514266163111, + 0.011332731693983078, + 0.011127781122922897, + 0.013554583303630352, + 0.010683272033929825, + 0.0051271808333694935, + 0.007362221833318472, + 0.011024708859622478, + 0.005270989146083593, + 0.006813940592110157, + 0.010928615927696228, + 0.011896591633558273, + 0.010539873503148556, + 0.005667511839419603, + 0.010301098227500916, + 0.008444849401712418, + 0.0049700019881129265, + 0.007234850432723761, + 0.008749914355576038, + 0.05230805277824402, + 0.0227012038230896, + 0.012019901536405087, + 0.01043891254812479, + 0.01399880088865757, + 0.009904765523970127, + 0.018021592870354652, + 0.05027210712432861, + 0.015050086192786694, + 0.012328830547630787, + 0.007725369185209274, + 0.01428106427192688, + 0.011139942333102226, + 0.008470329456031322, + 0.01456812210381031, + 0.02248593233525753, + 0.007048893719911575, + 0.006875211838632822, + 0.006444802042096853, + 0.013816985301673412, + 0.00682835653424263, + 0.004108807537704706, + 0.006523833144456148, + 0.008728273212909698, + 0.008614122867584229, + 0.008037015795707703, + 0.018610989674925804, + 0.016922999173402786, + 0.009358462877571583, + 0.004799579735845327, + 0.0072698574513196945, + 0.007933567278087139, + 0.024069098755717278, + 0.013027180917561054, + 0.020037321373820305, + 0.019600816071033478, + 0.02886919490993023, + 0.01196895819157362, + 0.019588688388466835, + 0.030425406992435455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.026341434568166733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, shot from an extremely low angle, captures a beam of warm golden sunlight filtering through an old window, slanting across the weathered, dark brown wooden floor. Motes of dust dancing in the air are clearly visible in the beam, creating a beautiful Tyndall effect. From the varying widths of the floorboards, several slender, tender green vines cautiously poke their heads, their newly formed, microscopically hairy tendrils appearing translucent in the light. As time passes, these vines ascend with a steady and leisurely rhythm, their heart-shaped leaves slowly unfurling in the light, their emerald green appearing strikingly vibrant against the dark, scratched wood flooring. Tensors tentatively dance in the air before gracefully grasping the rough surface of the wood, winding their way along the distinct grain toward the center of the spot of light. The entire scene, shrouded in profound tranquility, evokes a tenacious vitality, creating a magical realist atmosphere of time condensed and life quietly reawakening.", + "seed": 5572, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_2212_seed_5572_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": -0.009543658839538694, + "vlm_logprob": -0.009543658839538694, + "vlm_yes_no_margin": 5.275, + "vlm_token_credit": -0.009543658839538694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020882271230220795 + ], + "opposite_logprob_tokens": [ + -4.5208821296691895 + ], + "logprob_reward": -0.020882271230220795, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.020882271230220795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24601410329341888, + 0.07415151596069336, + 0.151311993598938, + 0.08863650262355804, + 0.08233334869146347, + 0.09318111091852188, + 0.02418169006705284, + 0.1321841925382614, + 0.0446985624730587, + 0.06330693513154984 + ], + "credit_map": [ + 0.020937558263540268, + 0.01769479177892208, + 0.01794840767979622, + 0.017729811370372772, + 0.06874404847621918, + 0.05245714262127876, + 0.032258596271276474, + 0.018243741244077682, + 0.006976802833378315, + 0.0049912044778466225, + 0.005915708839893341, + 0.006875152233988047, + 0.019962895661592484, + 0.013104613870382309, + 0.009036923758685589, + 0.007288210093975067, + 0.007831676863133907, + 0.006980735342949629, + 0.006686086300760508, + 0.008680003695189953, + 0.06223730370402336, + 0.039997223764657974, + 0.010042654350399971, + 0.008856316097080708, + 0.008605943992733955, + 0.010738043114542961, + 0.013139607384800911, + 0.012160810641944408, + 0.01592673361301422, + 0.01162631157785654, + 0.0073095993138849735, + 0.009129454381763935, + 0.006605631206184626, + 0.008636433631181717, + 0.007722044829279184, + 0.01403252687305212, + 0.01250995509326458, + 0.010551007464528084, + 0.013967969454824924, + 0.008307777345180511, + 0.009997024200856686, + 0.014168859459459782, + 0.007917037233710289, + 0.013422089628875256, + 0.01566324383020401, + 0.014699197374284267, + 0.007606636732816696, + 0.009707026183605194, + 0.0022483880165964365, + 0.0016628541052341461, + 0.0033634386491030455, + 0.00402195006608963, + 0.0036404593847692013, + 0.003546985564753413, + 0.003117347601801157, + 0.002580265048891306, + 0.013882954604923725, + 0.016533207148313522, + 0.014972042292356491, + 0.019315063953399658, + 0.015001141466200352, + 0.025683291256427765, + 0.013249024748802185, + 0.013547462411224842, + 0.0033355988562107086, + 0.003555742558091879, + 0.00553199602290988, + 0.0049249036237597466, + 0.004783296026289463, + 0.0084830978885293, + 0.005086017306894064, + 0.008997910656034946, + 0.005802830681204796, + 0.004116463474929333, + 0.013199600391089916, + 0.00617167679592967, + 0.008126902393996716, + 0.0093744657933712, + 0.006202086340636015, + 0.010312911123037338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.020882271230220795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004111172165721655 + ], + "opposite_logprob_tokens": [ + -5.504111289978027 + ], + "logprob_reward": -0.004111172165721655, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004111172165721655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22070620954036713, + 0.11918699741363525, + 0.1864704042673111, + 0.040424197912216187, + 0.03729996085166931, + 0.05453438684344292, + 0.01306235883384943, + 0.26597532629966736, + 0.047136977314949036, + 0.015203122049570084 + ], + "credit_map": [ + 0.018215205520391464, + 0.023822279646992683, + 0.021241357550024986, + 0.020159617066383362, + 0.050012923777103424, + 0.03686068579554558, + 0.031534306704998016, + 0.01885984092950821, + 0.008495488204061985, + 0.0076516177505254745, + 0.006042776629328728, + 0.006924211513251066, + 0.046889178454875946, + 0.028665510937571526, + 0.007444771938025951, + 0.007073449902236462, + 0.009516233578324318, + 0.007066330406814814, + 0.0057389456778764725, + 0.009810710325837135, + 0.08198872208595276, + 0.053130775690078735, + 0.008888749405741692, + 0.010329949669539928, + 0.005083997268229723, + 0.0031592934392392635, + 0.006122762802988291, + 0.0045874775387346745, + 0.005266393534839153, + 0.005351673346012831, + 0.005036556627601385, + 0.0058160522021353245, + 0.004173338413238525, + 0.0035901644732803106, + 0.003005842212587595, + 0.005335436202585697, + 0.004421703051775694, + 0.00521843321621418, + 0.004382337909191847, + 0.007172711659222841, + 0.006426377687603235, + 0.008589062839746475, + 0.004355366807430983, + 0.00680730352178216, + 0.01136836875230074, + 0.008149650879204273, + 0.004354291129857302, + 0.00448396522551775, + 0.0013687253231182694, + 0.0010311286896467209, + 0.0013391503598541021, + 0.0018097226275131106, + 0.0019447567174211144, + 0.001795325893908739, + 0.0016420397441834211, + 0.0021315105259418488, + 0.024654479697346687, + 0.030826639384031296, + 0.028242668136954308, + 0.04311486706137657, + 0.029194362461566925, + 0.04624694213271141, + 0.029136447235941887, + 0.03455892205238342, + 0.004000926855951548, + 0.0044380370527505875, + 0.006372249685227871, + 0.006596759893000126, + 0.0058295875787734985, + 0.005894861649721861, + 0.006259402260184288, + 0.007745157927274704, + 0.0017131221247836947, + 0.001534602022729814, + 0.0019664193969219923, + 0.0016381273744627833, + 0.002296403283253312, + 0.002127366606146097, + 0.0021295370534062386, + 0.001797545701265335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.004111172165721655 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013447299133986235 + ], + "opposite_logprob_tokens": [ + -6.626344680786133 + ], + "logprob_reward": -0.0013447299133986235, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013447299133986235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49005818367004395, + 0.06268199533224106, + 0.035961247980594635, + 0.04118911549448967, + 0.07310616970062256, + 0.04089207202196121, + 0.02006118930876255, + 0.12015733867883682, + 0.04779758304357529, + 0.06809507310390472 + ], + "credit_map": [ + 0.028326747938990593, + 0.03332539647817612, + 0.02889789268374443, + 0.02957809902727604, + 0.17435747385025024, + 0.11940748989582062, + 0.04231070354580879, + 0.03385442867875099, + 0.004270556382834911, + 0.005838767625391483, + 0.005269186105579138, + 0.007860440760850906, + 0.0153662646189332, + 0.009168308228254318, + 0.006180980242788792, + 0.008727498352527618, + 0.0032835279125720263, + 0.00233756797388196, + 0.003111928468570113, + 0.0058989026583731174, + 0.006507109384983778, + 0.005051682237535715, + 0.00634674122557044, + 0.0034237909130752087, + 0.0036970851942896843, + 0.004792499355971813, + 0.008017690852284431, + 0.00644023297354579, + 0.004218409303575754, + 0.005592208821326494, + 0.004434117581695318, + 0.003996873274445534, + 0.005739608313888311, + 0.00739095127210021, + 0.005980301648378372, + 0.01109152939170599, + 0.009122245013713837, + 0.00885779969394207, + 0.01616969145834446, + 0.008754047565162182, + 0.0031365775503218174, + 0.002262226538732648, + 0.005628728773444891, + 0.007473732810467482, + 0.004999841563403606, + 0.0061231800355017185, + 0.00727455411106348, + 0.003993231803178787, + 0.0024351885076612234, + 0.0019143630051985383, + 0.003090847749263048, + 0.002557479776442051, + 0.0025707255117595196, + 0.0028834654949605465, + 0.0018203541403636336, + 0.00278876512311399, + 0.012469746172428131, + 0.013824674300849438, + 0.014381574466824532, + 0.0213665459305048, + 0.012111084535717964, + 0.017790518701076508, + 0.012600854970514774, + 0.015612343326210976, + 0.004423942882567644, + 0.005257312674075365, + 0.005593038164079189, + 0.006147729232907295, + 0.006166188511997461, + 0.006838045548647642, + 0.005593874957412481, + 0.007777456659823656, + 0.0054292199201881886, + 0.005051033105701208, + 0.012852787971496582, + 0.005738314241170883, + 0.008108929730951786, + 0.007005323655903339, + 0.006143475882709026, + 0.017765989527106285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0013447299133986235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201361745595932 + ], + "opposite_logprob_tokens": [ + -5.753201484680176 + ], + "logprob_reward": -0.003201361745595932, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201361745595932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5395967960357666, + 0.05360587313771248, + 0.06635772436857224, + 0.04430156201124191, + 0.05099295452237129, + 0.018159443512558937, + 0.012119762599468231, + 0.13928990066051483, + 0.019148390740156174, + 0.05642759054899216 + ], + "credit_map": [ + 0.03631536290049553, + 0.051947563886642456, + 0.03921790048480034, + 0.03952271118760109, + 0.15687638521194458, + 0.11091797053813934, + 0.06323612481355667, + 0.04156281054019928, + 0.0045699081383645535, + 0.0045343306846916676, + 0.003140258602797985, + 0.0064572920091450214, + 0.01561775617301464, + 0.008837619796395302, + 0.0053050098940730095, + 0.0051437001675367355, + 0.003592138644307852, + 0.003239304292947054, + 0.004266984760761261, + 0.0055671934969723225, + 0.023950321599841118, + 0.014667690731585026, + 0.006142321042716503, + 0.004931768402457237, + 0.004992948845028877, + 0.005109108984470367, + 0.005816743243485689, + 0.006104619242250919, + 0.007298641838133335, + 0.005956122186034918, + 0.004059270024299622, + 0.004964111372828484, + 0.005736925173550844, + 0.0054250918328762054, + 0.003651499282568693, + 0.007348166778683662, + 0.009851142764091492, + 0.006018040236085653, + 0.007396778091788292, + 0.005565317813307047, + 0.0013900342164561152, + 0.0013120579533278942, + 0.0020640548318624496, + 0.002859898842871189, + 0.0033147598151117563, + 0.002676335396245122, + 0.0022719893604516983, + 0.002270313911139965, + 0.0011959613766521215, + 0.001533901784569025, + 0.0013350998051464558, + 0.0018345987191423774, + 0.0023112827911973, + 0.0015086609637364745, + 0.0013481167843565345, + 0.0010521415388211608, + 0.012980910949409008, + 0.016515467315912247, + 0.013908407650887966, + 0.02536093071103096, + 0.01499310415238142, + 0.022557836025953293, + 0.015296893194317818, + 0.017676373943686485, + 0.0015342426486313343, + 0.0022316842805594206, + 0.0024031181819736958, + 0.002572797005996108, + 0.002906269859522581, + 0.002088211476802826, + 0.002340537728741765, + 0.0030715318862348795, + 0.0035847830586135387, + 0.003961408510804176, + 0.012374434620141983, + 0.00511296559125185, + 0.006512560416013002, + 0.0054882727563381195, + 0.004959526006132364, + 0.014433639124035835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.003201361745595932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178759142756462 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018178759142756462, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178759142756462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3407117426395416, + 0.03775210306048393, + 0.15878035128116608, + 0.04973636195063591, + 0.05964802950620651, + 0.035379376262426376, + 0.02075822651386261, + 0.11055909842252731, + 0.09705723077058792, + 0.0896175354719162 + ], + "credit_map": [ + 0.022087907418608665, + 0.040554407984018326, + 0.02241678722202778, + 0.02177164889872074, + 0.10777425765991211, + 0.06361746042966843, + 0.038706935942173004, + 0.02378232218325138, + 0.002392867347225547, + 0.0030983369797468185, + 0.0031299295369535685, + 0.005607985891401768, + 0.009777514263987541, + 0.006429961882531643, + 0.00354652083478868, + 0.003768982132896781, + 0.0077612376771867275, + 0.0057647922076284885, + 0.0056032296270132065, + 0.00792754627764225, + 0.07020051777362823, + 0.04529837518930435, + 0.007644423749297857, + 0.008580226451158524, + 0.006158347241580486, + 0.007244929671287537, + 0.008313853293657303, + 0.006460289470851421, + 0.005930096842348576, + 0.006024559494107962, + 0.004873533733189106, + 0.004730750806629658, + 0.0032585638109594584, + 0.0024745629634708166, + 0.0032802538480609655, + 0.009835690259933472, + 0.006901954300701618, + 0.007430007215589285, + 0.01936621218919754, + 0.007100775837898254, + 0.003981088288128376, + 0.005971335340291262, + 0.0032252015080302954, + 0.004523288458585739, + 0.006274226121604443, + 0.004961227998137474, + 0.0034641490783542395, + 0.0029788585379719734, + 0.0016292459331452847, + 0.002080559264868498, + 0.002815525745972991, + 0.0028031091205775738, + 0.0036106619518250227, + 0.0029414857272058725, + 0.002324481960386038, + 0.0025531561113893986, + 0.010595095343887806, + 0.01060547400265932, + 0.013446983881294727, + 0.020413508638739586, + 0.008676202967762947, + 0.02058257721364498, + 0.01413505058735609, + 0.012104197405278683, + 0.0062539535574615, + 0.008195709437131882, + 0.013875618577003479, + 0.010288992896676064, + 0.009626801125705242, + 0.024166546761989594, + 0.010043292306363583, + 0.01460630726069212, + 0.007304362952709198, + 0.005216059274971485, + 0.02171715348958969, + 0.005766757298260927, + 0.009887232445180416, + 0.010894034057855606, + 0.007605685852468014, + 0.021226240321993828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.018178759142756462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up time-lapse shot, shot from an extremely low angle, captures a beam of warm golden sunlight filtering through an old window, slanting across the weathered, dark brown wooden floor. Motes of dust dancing in the air are clearly visible in the beam, creating a beautiful Tyndall effect. From the varying widths of the floorboards, several slender, tender green vines cautiously poke their heads, their newly formed, microscopically hairy tendrils appearing translucent in the light. As time passes, these vines ascend with a steady and leisurely rhythm, their heart-shaped leaves slowly unfurling in the light, their emerald green appearing strikingly vibrant against the dark, scratched wood flooring. Tensors tentatively dance in the air before gracefully grasping the rough surface of the wood, winding their way along the distinct grain toward the center of the spot of light. The entire scene, shrouded in profound tranquility, evokes a tenacious vitality, creating a magical realist atmosphere of time condensed and life quietly reawakening.", + "seed": 5573, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_0/index_2212_seed_5573_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": -0.02210102081298828, + "vlm_logprob": -0.02210102081298828, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02210102081298828, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05061383172869682 + ], + "opposite_logprob_tokens": [ + -3.3006138801574707 + ], + "logprob_reward": -0.05061383172869682, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05061383172869682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30147990584373474, + 0.12360486388206482, + 0.039174456149339676, + 0.06224574148654938, + 0.04338303580880165, + 0.023583468049764633, + 0.06882842630147934, + 0.060626398772001266, + 0.053287897258996964, + 0.22378584742546082 + ], + "credit_map": [ + 0.05461970716714859, + 0.021552560850977898, + 0.028659261763095856, + 0.03293384611606598, + 0.03159014880657196, + 0.046390753239393234, + 0.037519194185733795, + 0.048214446753263474, + 0.022390196099877357, + 0.00880466215312481, + 0.010115916840732098, + 0.028416510671377182, + 0.011150813661515713, + 0.013081817887723446, + 0.01580175757408142, + 0.013843197375535965, + 0.0057441540993750095, + 0.0029354619327932596, + 0.006400272250175476, + 0.0033125062473118305, + 0.00417546508833766, + 0.007342966739088297, + 0.003937970846891403, + 0.005325656849890947, + 0.01886097714304924, + 0.0036008127499371767, + 0.005829779431223869, + 0.0046723512932658195, + 0.0033323518000543118, + 0.0038403153885155916, + 0.006838771980255842, + 0.015270380303263664, + 0.00718140322715044, + 0.0034073838032782078, + 0.004743891768157482, + 0.007050303276628256, + 0.004567524418234825, + 0.005276327952742577, + 0.005117356777191162, + 0.006038845982402563, + 0.0034302223939448595, + 0.003545688930898905, + 0.0023952804040163755, + 0.002522918628528714, + 0.003627853700891137, + 0.0028196913190186024, + 0.002498868852853775, + 0.0027429431211203337, + 0.008513117209076881, + 0.006171088200062513, + 0.00644836388528347, + 0.017176616936922073, + 0.006290521938353777, + 0.007054157089442015, + 0.010911483317613602, + 0.006263085175305605, + 0.009761943481862545, + 0.004364547319710255, + 0.006443505175411701, + 0.01273166574537754, + 0.00611250102519989, + 0.006889878772199154, + 0.007728501223027706, + 0.006593861151486635, + 0.006345923524349928, + 0.0034404578618705273, + 0.011402345262467861, + 0.007921728305518627, + 0.004721773322671652, + 0.00688832625746727, + 0.0062025124207139015, + 0.006364827509969473, + 0.03836620971560478, + 0.034194882959127426, + 0.024529866874217987, + 0.022151794284582138, + 0.0360228605568409, + 0.020642433315515518, + 0.026446053758263588, + 0.021431751549243927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05061383172869682 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026388226076960564 + ], + "opposite_logprob_tokens": [ + -3.651388168334961 + ], + "logprob_reward": -0.026388226076960564, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026388226076960564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26340511441230774, + 0.11551562696695328, + 0.069978266954422, + 0.14016661047935486, + 0.07713460922241211, + 0.03691553696990013, + 0.08108291029930115, + 0.047538962215185165, + 0.06897293776273727, + 0.09928946197032928 + ], + "credit_map": [ + 0.0570109598338604, + 0.018618429079651833, + 0.02697950415313244, + 0.04080544039607048, + 0.025216246023774147, + 0.032550714910030365, + 0.023435836657881737, + 0.03878796845674515, + 0.02321568690240383, + 0.008898550644516945, + 0.010083585046231747, + 0.02683723345398903, + 0.010821342468261719, + 0.011393202468752861, + 0.012242111377418041, + 0.01202392391860485, + 0.00879419781267643, + 0.00476239575073123, + 0.015032781288027763, + 0.013274259865283966, + 0.004943682346493006, + 0.007281430531293154, + 0.006255393847823143, + 0.009634129703044891, + 0.042524125427007675, + 0.008260703645646572, + 0.009363341145217419, + 0.012788609601557255, + 0.009859391488134861, + 0.0094413161277771, + 0.01640622690320015, + 0.03152288869023323, + 0.011829388327896595, + 0.006405996158719063, + 0.008597144857048988, + 0.011407246813178062, + 0.00851038284599781, + 0.009547770954668522, + 0.009423741139471531, + 0.011412938125431538, + 0.004195463843643665, + 0.005597825627774, + 0.0041005490347743034, + 0.005232628900557756, + 0.00523240165784955, + 0.004638256039470434, + 0.003867894411087036, + 0.004050517920404673, + 0.011054126545786858, + 0.004008625168353319, + 0.010713787749409676, + 0.020481042563915253, + 0.004970633424818516, + 0.008202295750379562, + 0.012286330573260784, + 0.00936606340110302, + 0.006752585060894489, + 0.002674003131687641, + 0.006311518605798483, + 0.009697431698441505, + 0.003316455287858844, + 0.0052410210482776165, + 0.006233440712094307, + 0.007312505505979061, + 0.009691188111901283, + 0.004572418984025717, + 0.014763571321964264, + 0.010347428731620312, + 0.0056092203594744205, + 0.008509381674230099, + 0.007692092563956976, + 0.007787641137838364, + 0.015150997787714005, + 0.011859282851219177, + 0.015172173269093037, + 0.012788406573235989, + 0.009031519293785095, + 0.006601541303098202, + 0.016468284651637077, + 0.01221726555377245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.026388226076960564 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270277615636587 + ], + "opposite_logprob_tokens": [ + -5.255270481109619 + ], + "logprob_reward": -0.005270277615636587, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270277615636587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1270783394575119, + 0.1455557942390442, + 0.09113837778568268, + 0.05912882834672928, + 0.0716066062450409, + 0.04327749088406563, + 0.1333162784576416, + 0.05472056567668915, + 0.09608388692140579, + 0.17809388041496277 + ], + "credit_map": [ + 0.0270026046782732, + 0.009839817881584167, + 0.014209316112101078, + 0.0203035119920969, + 0.012745114043354988, + 0.012257902882993221, + 0.010018983855843544, + 0.020701076835393906, + 0.03546495735645294, + 0.008709543384611607, + 0.010067465715110302, + 0.04502775892615318, + 0.007452886085957289, + 0.01042870245873928, + 0.014060036279261112, + 0.014344443567097187, + 0.0142562510445714, + 0.0078027197159826756, + 0.01897781528532505, + 0.013988614082336426, + 0.007586965803056955, + 0.009449323639273643, + 0.009265638887882233, + 0.009811046533286572, + 0.011370415799319744, + 0.005087679717689753, + 0.006160095799714327, + 0.007386407814919949, + 0.00528681231662631, + 0.006471916567534208, + 0.006739272736012936, + 0.010626222938299179, + 0.012036786414682865, + 0.005715380888432264, + 0.007375355809926987, + 0.009174673818051815, + 0.007957777939736843, + 0.008564388379454613, + 0.009867480956017971, + 0.0109147559851408, + 0.006157038267701864, + 0.006465851329267025, + 0.004911134019494057, + 0.00500484136864543, + 0.004811828024685383, + 0.0058257863856852055, + 0.004260914400219917, + 0.0058400933630764484, + 0.023614486679434776, + 0.009490367956459522, + 0.013342611491680145, + 0.025888407602906227, + 0.017917023971676826, + 0.01141553558409214, + 0.017706967890262604, + 0.013940869830548763, + 0.007788268383592367, + 0.004629845730960369, + 0.004765398800373077, + 0.010664725676178932, + 0.006318285595625639, + 0.007365827914327383, + 0.007087420206516981, + 0.006100797094404697, + 0.009060864336788654, + 0.010055621154606342, + 0.021897578611969948, + 0.023358836770057678, + 0.007112909108400345, + 0.006927722133696079, + 0.010513371787965298, + 0.007156974636018276, + 0.023166397586464882, + 0.03180670738220215, + 0.020325452089309692, + 0.015572095289826393, + 0.03598910570144653, + 0.017858155071735382, + 0.014438988640904427, + 0.018936971202492714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.005270277615636587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649060804396868 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007649060804396868, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649060804396868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2593470513820648, + 0.07599043846130371, + 0.0877583771944046, + 0.04225044324994087, + 0.08576637506484985, + 0.02871745452284813, + 0.11784694343805313, + 0.058176398277282715, + 0.1195189356803894, + 0.12462759017944336 + ], + "credit_map": [ + 0.07267466932535172, + 0.016652246937155724, + 0.025868594646453857, + 0.02917538583278656, + 0.01119806244969368, + 0.01776748150587082, + 0.02757379785180092, + 0.058436837047338486, + 0.014192785136401653, + 0.00749492272734642, + 0.009153257124125957, + 0.018501894548535347, + 0.005950691178441048, + 0.00655214162543416, + 0.007178503554314375, + 0.006966244429349899, + 0.009056607261300087, + 0.006956766359508038, + 0.01729242317378521, + 0.015124390833079815, + 0.007301920559257269, + 0.010977051220834255, + 0.011284570209681988, + 0.009764647111296654, + 0.010356379672884941, + 0.004707478918135166, + 0.0032463972456753254, + 0.004155754577368498, + 0.003300054930150509, + 0.002949125599116087, + 0.004910839721560478, + 0.008624413050711155, + 0.012584795244038105, + 0.007012183777987957, + 0.010033143684267998, + 0.009348204359412193, + 0.01050904393196106, + 0.011433825828135014, + 0.01115626934915781, + 0.013688912615180016, + 0.003573982510715723, + 0.00453100586310029, + 0.0034863280598074198, + 0.003701288951560855, + 0.003925687167793512, + 0.0031008529476821423, + 0.0031597972847521305, + 0.003238509176298976, + 0.017869384959340096, + 0.007278434466570616, + 0.014112548902630806, + 0.028849763795733452, + 0.007278678938746452, + 0.012407295405864716, + 0.017676716670393944, + 0.012374119833111763, + 0.011341569013893604, + 0.003582366043701768, + 0.006418013479560614, + 0.014109345152974129, + 0.004651342052966356, + 0.004634093027561903, + 0.007743444759398699, + 0.005696222186088562, + 0.012370201759040356, + 0.009905618615448475, + 0.027727283537387848, + 0.028289450332522392, + 0.008321230299770832, + 0.00847607757896185, + 0.013696359470486641, + 0.010732706636190414, + 0.015358909964561462, + 0.014624158851802349, + 0.01486967783421278, + 0.014991851523518562, + 0.015230925753712654, + 0.014023362658917904, + 0.017290698364377022, + 0.01823800429701805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.007649060804396868 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up time-lapse captured from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a beam of warm golden sunlight filtering through an old window and slanting across a weathered, dark brown wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are slender green vines with heart-shaped leaves visible ascending and unfurling on the floorboards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are motes of dust visible dancing in the light beam, and do translucent tendrils appear to grasp the wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the vines and tendrils move with a steady, leisurely rhythm as they wind along the wood grain toward the center of the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583707839250565 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583707839250565, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583707839250565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1951407790184021, + 0.15318496525287628, + 0.04364470764994621, + 0.17741279304027557, + 0.07156555354595184, + 0.04416407644748688, + 0.05198638513684273, + 0.03208254277706146, + 0.04988706484436989, + 0.1809311956167221 + ], + "credit_map": [ + 0.04776797071099281, + 0.011682325042784214, + 0.027581412345170975, + 0.037058547139167786, + 0.014293202199041843, + 0.017684821039438248, + 0.01659240387380123, + 0.022480066865682602, + 0.023662200197577477, + 0.013827787712216377, + 0.0127825066447258, + 0.0429048128426075, + 0.011624420993030071, + 0.010684354230761528, + 0.02083459496498108, + 0.01686428114771843, + 0.007019109092652798, + 0.002883583540096879, + 0.006594325881451368, + 0.010013241320848465, + 0.004612436518073082, + 0.0034432297106832266, + 0.004072950221598148, + 0.005005827639251947, + 0.04754455387592316, + 0.012792428024113178, + 0.01193066593259573, + 0.018929263576865196, + 0.015341770835220814, + 0.01357486005872488, + 0.021341385319828987, + 0.03595784679055214, + 0.010392784141004086, + 0.005994826555252075, + 0.007726576644927263, + 0.011603874154388905, + 0.00759159866720438, + 0.007150929421186447, + 0.010250328108668327, + 0.010854627005755901, + 0.005819712299853563, + 0.006921788211911917, + 0.004805470816791058, + 0.00353027880191803, + 0.0053343139588832855, + 0.007318463642150164, + 0.005643801297992468, + 0.004790243227034807, + 0.006649315822869539, + 0.0048830024898052216, + 0.0038940375670790672, + 0.010562406852841377, + 0.006343487650156021, + 0.005462664645165205, + 0.007662867195904255, + 0.0065286001190543175, + 0.00687792943790555, + 0.0022726054303348064, + 0.0040944768115878105, + 0.005583612713962793, + 0.0028502915520220995, + 0.0029679194558411837, + 0.0036560287699103355, + 0.003779676742851734, + 0.006155330687761307, + 0.003126879222691059, + 0.007292811293154955, + 0.010911960154771805, + 0.003952715080231428, + 0.0064207520335912704, + 0.0052998969331383705, + 0.00672671664506197, + 0.023904243484139442, + 0.029434455558657646, + 0.01847018674015999, + 0.02185123600065708, + 0.03237096965312958, + 0.02030247449874878, + 0.018083151429891586, + 0.01651446335017681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.020583707839250565 + } + ], + "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/0000152/group_1/index_tensor([848, 848])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_tensor([848, 848])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e3d14587661d162990297560c6e7d1fc4e201e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_tensor([848, 848])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a breathtakingly high-speed phantom camera style, a medium shot gracefully circles a high school boy with short, unkempt hair, standing on the rough surface of a city rooftop. The camera's movement is smooth and deliberate, circling him while his own movements unfold in extreme slow motion. He wears a standard white button-down school shirt, the thin fabric fluttering gently in an invisible breeze, and dark trousers. His pose is contemplative, his hands resting lightly on the cold metal safety railing. Illuminated by the cool, ethereal light above and the warm, ambient light of the sprawling city below, his face is etched with a profound sense of exploration; his forehead is slightly wrinkled, his eyes wide as he surveys the sky, his lips barely parted, as if on the verge of a question. In stark and breathtaking contrast to his nearly frozen state, the entire Milky Way galaxy rotates rapidly, time-lapse, across the celestial dome. A dazzling, majestic sweep of vibrant purples, deep blues, and stellar dust paints the sky as a scene of movement. This juxtaposition creates a powerful, awe-inspiring atmosphere of fleeting moments of human introspection contrasted with the vast, indifferent passage of cosmic time.", + "seed": 2288, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2288_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": -0.07903948286548257, + "vlm_logprob": -0.07903948286548257, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.07903948286548257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0452631451189518 + ], + "opposite_logprob_tokens": [ + -3.1702630519866943 + ], + "logprob_reward": -0.0452631451189518, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0452631451189518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523209512233734, + 0.11881358176469803, + 0.08179924637079239, + 0.11736023426055908, + 0.06736088544130325, + 0.10073694586753845, + 0.13596245646476746, + 0.10777647793292999, + 0.0594443753361702, + 0.05842486023902893 + ], + "credit_map": [ + 0.02107425406575203, + 0.027761556208133698, + 0.023749155923724174, + 0.02095303311944008, + 0.01693708822131157, + 0.01936650462448597, + 0.01013995986431837, + 0.012339395470917225, + 0.014294932596385479, + 0.017346318811178207, + 0.019110839813947678, + 0.011664040386676788, + 0.010386261157691479, + 0.02808615379035473, + 0.008900880813598633, + 0.009024150669574738, + 0.012714377604424953, + 0.018117545172572136, + 0.008703219704329967, + 0.01102282851934433, + 0.008675933815538883, + 0.007360708434134722, + 0.004992131609469652, + 0.010212498717010021, + 0.019893459975719452, + 0.02325964719057083, + 0.012182570993900299, + 0.009584449231624603, + 0.011972134001553059, + 0.015192544087767601, + 0.00917926337569952, + 0.016096167266368866, + 0.009119587950408459, + 0.015355012379586697, + 0.008555932901799679, + 0.005209044087678194, + 0.006741292309015989, + 0.00790443830192089, + 0.008737938478589058, + 0.005737637635320425, + 0.0074088298715651035, + 0.035184744745492935, + 0.00451611028984189, + 0.0030450145713984966, + 0.0040430473163723946, + 0.038660574704408646, + 0.0046132695861160755, + 0.003265357343479991, + 0.015973011031746864, + 0.01643109880387783, + 0.014883973635733128, + 0.016355330124497414, + 0.017041530460119247, + 0.023119699209928513, + 0.012773539870977402, + 0.019384270533919334, + 0.012852783314883709, + 0.019168861210346222, + 0.015093468129634857, + 0.010487096384167671, + 0.0068613276816904545, + 0.028782010078430176, + 0.006035398691892624, + 0.008495534770190716, + 0.011701264418661594, + 0.014029706828296185, + 0.0044183554127812386, + 0.009137099608778954, + 0.00640683202072978, + 0.005424813833087683, + 0.003931700251996517, + 0.0043946076184511185, + 0.008435795083642006, + 0.009089360013604164, + 0.0059252576902508736, + 0.007651086430996656, + 0.006033048499375582, + 0.0073949191719293594, + 0.006645619869232178, + 0.007249775342643261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0452631451189518 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372610732913017 + ], + "opposite_logprob_tokens": [ + -3.408726215362549 + ], + "logprob_reward": -0.03372610732913017, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372610732913017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5118886828422546, + 0.11989523470401764, + 0.025150595232844353, + 0.04283195734024048, + 0.03439722955226898, + 0.04136863723397255, + 0.058617010712623596, + 0.11042960733175278, + 0.04122544080018997, + 0.014195610769093037 + ], + "credit_map": [ + 0.07629653066396713, + 0.11315944045782089, + 0.07374674826860428, + 0.05240541324019432, + 0.04099512845277786, + 0.08937875926494598, + 0.03303193300962448, + 0.0328747034072876, + 0.015973085537552834, + 0.017227990552783012, + 0.023359136655926704, + 0.01866872049868107, + 0.010532346554100513, + 0.015781966969370842, + 0.00866139866411686, + 0.009690581820905209, + 0.003692745929583907, + 0.006123076658695936, + 0.0029922674875706434, + 0.002972991205751896, + 0.0021063333842903376, + 0.002988044871017337, + 0.0014956563245505095, + 0.0027794784400612116, + 0.007828477770090103, + 0.009082436561584473, + 0.0037784164305776358, + 0.003898969851434231, + 0.004186500329524279, + 0.006706049665808678, + 0.003365685697644949, + 0.0039854212664067745, + 0.005977035034447908, + 0.006733335088938475, + 0.004237023182213306, + 0.004618040751665831, + 0.0023549089673906565, + 0.00393070699647069, + 0.0028972444124519825, + 0.003648934653028846, + 0.0047757672145962715, + 0.011671405285596848, + 0.00350354821421206, + 0.0020802305079996586, + 0.0019084789091721177, + 0.013010254129767418, + 0.00208574952557683, + 0.0023332014679908752, + 0.009084981866180897, + 0.008165928535163403, + 0.006607739720493555, + 0.0066171432845294476, + 0.0054999664425849915, + 0.010697116144001484, + 0.005716744344681501, + 0.006227388978004456, + 0.013724913820624352, + 0.018062010407447815, + 0.014074721373617649, + 0.01122201420366764, + 0.007365360390394926, + 0.030200032517313957, + 0.007137345615774393, + 0.008643196895718575, + 0.00901726447045803, + 0.011035722680389881, + 0.0025969776324927807, + 0.003831559792160988, + 0.003059103386476636, + 0.0053962175734341145, + 0.0031753021758049726, + 0.003113292856141925, + 0.003091365797445178, + 0.001951457466930151, + 0.0015651575522497296, + 0.0013360435841605067, + 0.001785204280167818, + 0.0018240280915051699, + 0.0015084454789757729, + 0.0011339078191667795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03372610732913017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28134429454803467 + ], + "opposite_logprob_tokens": [ + -1.4063442945480347 + ], + "logprob_reward": -0.28134429454803467, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28134429454803467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07485271990299225, + 0.07621447741985321, + 0.1251845806837082, + 0.07319410890340805, + 0.05496509373188019, + 0.3684512674808502, + 0.06408632546663284, + 0.05824854224920273, + 0.06309137493371964, + 0.04171152785420418 + ], + "credit_map": [ + 0.009901589713990688, + 0.01597839593887329, + 0.01108025386929512, + 0.005713091231882572, + 0.007920394651591778, + 0.010663175955414772, + 0.0050511485897004604, + 0.008544668555259705, + 0.008608531206846237, + 0.013142517767846584, + 0.007220917381346226, + 0.004709735978394747, + 0.0063926056027412415, + 0.027548324316740036, + 0.004244950599968433, + 0.004346897359937429, + 0.022457793354988098, + 0.031489040702581406, + 0.013677503913640976, + 0.011923113837838173, + 0.01195515040308237, + 0.014130057767033577, + 0.010013575665652752, + 0.00953834131360054, + 0.01555589959025383, + 0.01668773591518402, + 0.007138175889849663, + 0.006498511880636215, + 0.00610388396307826, + 0.01083656307309866, + 0.005672701634466648, + 0.0047006383538246155, + 0.009369846433401108, + 0.011978481896221638, + 0.006220036186277866, + 0.006756673566997051, + 0.00448273029178381, + 0.00564681738615036, + 0.0056444741785526276, + 0.0048660337924957275, + 0.017884576693177223, + 0.14541509747505188, + 0.011232360266149044, + 0.006055782083421946, + 0.009958378970623016, + 0.16317731142044067, + 0.005939006339758635, + 0.008788736537098885, + 0.007470875512808561, + 0.010044330731034279, + 0.0064104581251740456, + 0.008333387784659863, + 0.006857282482087612, + 0.010354015044867992, + 0.006003726273775101, + 0.008612249046564102, + 0.008857578039169312, + 0.008654219098389149, + 0.006079155020415783, + 0.006695942487567663, + 0.005059263203293085, + 0.008916246704757214, + 0.005665244534611702, + 0.008320894092321396, + 0.017295006662607193, + 0.019771499559283257, + 0.003488218178972602, + 0.003019682364538312, + 0.004551398102194071, + 0.007897360250353813, + 0.003908851183950901, + 0.0031593546736985445, + 0.00789785198867321, + 0.005692521575838327, + 0.0038685339968651533, + 0.0042211441323161125, + 0.0038705102633684874, + 0.004976330790668726, + 0.006786248181015253, + 0.00439838832244277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.28134429454803467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011797142215073109 + ], + "opposite_logprob_tokens": [ + -6.7511796951293945 + ], + "logprob_reward": -0.0011797142215073109, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011797142215073109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12512432038784027, + 0.06763619929552078, + 0.09854728728532791, + 0.08960384875535965, + 0.3103727698326111, + 0.09649502485990524, + 0.044065725058317184, + 0.05957766994833946, + 0.058908335864543915, + 0.0496688075363636 + ], + "credit_map": [ + 0.01672150008380413, + 0.023934978991746902, + 0.017273293808102608, + 0.011600090190768242, + 0.014090641401708126, + 0.020639384165406227, + 0.00890625175088644, + 0.011958187445998192, + 0.007421991787850857, + 0.010367357172071934, + 0.008712919428944588, + 0.005120893474668264, + 0.007138520944863558, + 0.020724374800920486, + 0.003899207804352045, + 0.004250933416187763, + 0.01483340747654438, + 0.031299084424972534, + 0.010896530002355576, + 0.01032115239650011, + 0.007778414059430361, + 0.011660458520054817, + 0.006518278270959854, + 0.005239963997155428, + 0.01609591208398342, + 0.02556534856557846, + 0.010306858457624912, + 0.008497551083564758, + 0.004691127687692642, + 0.010858132503926754, + 0.005417189095169306, + 0.00817172508686781, + 0.06845297664403915, + 0.09110672026872635, + 0.02327347919344902, + 0.029080145061016083, + 0.017814237624406815, + 0.04201902821660042, + 0.01970098167657852, + 0.018925197422504425, + 0.014734867960214615, + 0.024306733161211014, + 0.00789121724665165, + 0.005599035881459713, + 0.006449287757277489, + 0.024440104141831398, + 0.0075568160973489285, + 0.005516963545233011, + 0.006824907381087542, + 0.007279638666659594, + 0.0036602672189474106, + 0.005234413780272007, + 0.003510355716571212, + 0.009662964381277561, + 0.00388945359736681, + 0.004003725480288267, + 0.008091283962130547, + 0.010206382721662521, + 0.004884721711277962, + 0.005261973477900028, + 0.00412306934595108, + 0.017908675596117973, + 0.003890810301527381, + 0.0052107530646026134, + 0.014677425846457481, + 0.018629832193255424, + 0.003661013673990965, + 0.003602108918130398, + 0.0034275043290108442, + 0.007054230198264122, + 0.00522890267893672, + 0.0026273143012076616, + 0.012541748583316803, + 0.009149941615760326, + 0.0043422444723546505, + 0.003996507730334997, + 0.003785722656175494, + 0.006749249063432217, + 0.005174099467694759, + 0.003929296508431435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0011797142215073109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033684153109788895 + ], + "opposite_logprob_tokens": [ + -3.408684253692627 + ], + "logprob_reward": -0.033684153109788895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033684153109788895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07649049907922745, + 0.04544027894735336, + 0.06354545801877975, + 0.028592413291335106, + 0.0856945663690567, + 0.4424062669277191, + 0.10339461266994476, + 0.02721613273024559, + 0.07601529359817505, + 0.05120450630784035 + ], + "credit_map": [ + 0.009734885767102242, + 0.013676958158612251, + 0.01092222798615694, + 0.009910393506288528, + 0.006873350124806166, + 0.010429935529828072, + 0.007766291499137878, + 0.007176456972956657, + 0.005234780255705118, + 0.005401953589171171, + 0.0049658967182040215, + 0.0032044812105596066, + 0.0054846699349582195, + 0.013152120634913445, + 0.004260934889316559, + 0.003735440317541361, + 0.010759339667856693, + 0.015265114605426788, + 0.0067624920047819614, + 0.005974188446998596, + 0.006597424857318401, + 0.0074091823771595955, + 0.003679158166050911, + 0.0070985653437674046, + 0.0036852064076811075, + 0.0036612353287637234, + 0.005077911540865898, + 0.0038948452565819025, + 0.003060327610000968, + 0.0031724076252430677, + 0.002732520457357168, + 0.003307958599179983, + 0.017522016540169716, + 0.020110417157411575, + 0.01017692033201456, + 0.009609176777303219, + 0.005266903899610043, + 0.009413562715053558, + 0.00634863693267107, + 0.007246921304613352, + 0.021257244050502777, + 0.17736788094043732, + 0.013558309525251389, + 0.006813816260546446, + 0.01207659300416708, + 0.1978054791688919, + 0.004596390295773745, + 0.008930543437600136, + 0.012099022977054119, + 0.015019658952951431, + 0.009367648512125015, + 0.014542668126523495, + 0.008992644026875496, + 0.01650049164891243, + 0.013669502921402454, + 0.013202978298068047, + 0.004042451269924641, + 0.005165457725524902, + 0.0017845163820311427, + 0.0020558382384479046, + 0.0017569612246006727, + 0.00840866006910801, + 0.0018351306207478046, + 0.0021671203430742025, + 0.015305914916098118, + 0.019150573760271072, + 0.004938908852636814, + 0.008768679574131966, + 0.005420189816504717, + 0.01011307630687952, + 0.006628526374697685, + 0.0056894258596003056, + 0.0075979833491146564, + 0.006878294982016087, + 0.004566896706819534, + 0.006628029048442841, + 0.007994563318789005, + 0.006358543410897255, + 0.00677128043025732, + 0.004408915061503649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.033684153109788895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a breathtakingly high-speed phantom camera style, a medium shot gracefully circles a high school boy with short, unkempt hair, standing on the rough surface of a city rooftop. The camera's movement is smooth and deliberate, circling him while his own movements unfold in extreme slow motion. He wears a standard white button-down school shirt, the thin fabric fluttering gently in an invisible breeze, and dark trousers. His pose is contemplative, his hands resting lightly on the cold metal safety railing. Illuminated by the cool, ethereal light above and the warm, ambient light of the sprawling city below, his face is etched with a profound sense of exploration; his forehead is slightly wrinkled, his eyes wide as he surveys the sky, his lips barely parted, as if on the verge of a question. In stark and breathtaking contrast to his nearly frozen state, the entire Milky Way galaxy rotates rapidly, time-lapse, across the celestial dome. A dazzling, majestic sweep of vibrant purples, deep blues, and stellar dust paints the sky as a scene of movement. This juxtaposition creates a powerful, awe-inspiring atmosphere of fleeting moments of human introspection contrasted with the vast, indifferent passage of cosmic time.", + "seed": 2289, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2289_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": -0.15428202911280095, + "vlm_logprob": -0.15428202911280095, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.15428202911280095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07458975911140442 + ], + "opposite_logprob_tokens": [ + -2.699589729309082 + ], + "logprob_reward": -0.07458975911140442, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07458975911140442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14752890169620514, + 0.08702594041824341, + 0.27855348587036133, + 0.14673097431659698, + 0.03757501766085625, + 0.02905474230647087, + 0.06244773417711258, + 0.08527308702468872, + 0.04060919210314751, + 0.08520094305276871 + ], + "credit_map": [ + 0.01843121647834778, + 0.027213558554649353, + 0.011771067045629025, + 0.011124960146844387, + 0.026105109602212906, + 0.018071455880999565, + 0.015172001905739307, + 0.0196395181119442, + 0.016441691666841507, + 0.01713019423186779, + 0.01031305082142353, + 0.005265752784907818, + 0.010606016963720322, + 0.008740859106183052, + 0.005952398292720318, + 0.012575970962643623, + 0.028743833303451538, + 0.04976895451545715, + 0.05110546946525574, + 0.016400570049881935, + 0.015770401805639267, + 0.08187791705131531, + 0.012242771685123444, + 0.022643540054559708, + 0.03202173486351967, + 0.020348304882645607, + 0.024544158950448036, + 0.009445265866816044, + 0.011252298951148987, + 0.03359492868185043, + 0.006719596683979034, + 0.008804678916931152, + 0.006788488477468491, + 0.0075088548474013805, + 0.0033743337262421846, + 0.002566529205068946, + 0.003914548084139824, + 0.006862597074359655, + 0.0024955940898507833, + 0.004064070992171764, + 0.0037891834508627653, + 0.005453788675367832, + 0.0034257296938449144, + 0.0028216049540787935, + 0.0038729440420866013, + 0.004258465953171253, + 0.002592712640762329, + 0.0028403098694980145, + 0.013120356015861034, + 0.012004307471215725, + 0.0034412576351314783, + 0.0032776680309325457, + 0.010257783345878124, + 0.011693954467773438, + 0.004315546248108149, + 0.004336860030889511, + 0.009637843817472458, + 0.0098798219114542, + 0.0067362054251134396, + 0.005630881525576115, + 0.02031131088733673, + 0.009700646623969078, + 0.00728602847084403, + 0.01609034650027752, + 0.00528248306363821, + 0.0058576506562530994, + 0.005713404156267643, + 0.0044362107291817665, + 0.006777629256248474, + 0.004654912743717432, + 0.004511800594627857, + 0.0033750971779227257, + 0.01895010471343994, + 0.014163569547235966, + 0.010675106197595596, + 0.008445920422673225, + 0.006737577263265848, + 0.018228398635983467, + 0.002847240772098303, + 0.005153017118573189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.07458975911140442 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005227357149124 + ], + "opposite_logprob_tokens": [ + -2.6950523853302 + ], + "logprob_reward": -0.07005227357149124, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005227357149124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13002996146678925, + 0.06144016608595848, + 0.16800227761268616, + 0.12376640737056732, + 0.02327573485672474, + 0.01850111223757267, + 0.028252439573407173, + 0.05097997188568115, + 0.053304851055145264, + 0.34244704246520996 + ], + "credit_map": [ + 0.018263474106788635, + 0.02315358631312847, + 0.025923866778612137, + 0.007845279760658741, + 0.009259982965886593, + 0.016687657684087753, + 0.008616352453827858, + 0.020279766991734505, + 0.01346711814403534, + 0.014484791085124016, + 0.007592214737087488, + 0.004075288772583008, + 0.004584168549627066, + 0.007708597928285599, + 0.0038176788948476315, + 0.005710311233997345, + 0.017375081777572632, + 0.02557813748717308, + 0.033489055931568146, + 0.009040753357112408, + 0.009228913113474846, + 0.052173905074596405, + 0.006591105833649635, + 0.014525331556797028, + 0.045770321041345596, + 0.013169070705771446, + 0.04204760491847992, + 0.005521073006093502, + 0.0035378241445869207, + 0.005064920987933874, + 0.0023549417965114117, + 0.006300659850239754, + 0.00567229138687253, + 0.0032643633894622326, + 0.0018693184247240424, + 0.0012946007773280144, + 0.002840605331584811, + 0.004590127617120743, + 0.0017587098991498351, + 0.001985719660297036, + 0.004174880683422089, + 0.0028076954185962677, + 0.0019002814078703523, + 0.0010672194184735417, + 0.0016286502359434962, + 0.003967091906815767, + 0.0011107237078249454, + 0.0018445701571181417, + 0.0052961683832108974, + 0.007839430123567581, + 0.0023380937054753304, + 0.0017129640327766538, + 0.0027549252845346928, + 0.003216184675693512, + 0.002833457663655281, + 0.002261215588077903, + 0.009580541402101517, + 0.009702111594378948, + 0.006217846646904945, + 0.004463794641196728, + 0.00428833207115531, + 0.007635277230292559, + 0.0029869398567825556, + 0.006105135660618544, + 0.007530677132308483, + 0.010617947205901146, + 0.007969722151756287, + 0.0043251425959169865, + 0.005952528212219477, + 0.009220176376402378, + 0.0037068892270326614, + 0.003981770481914282, + 0.07760300487279892, + 0.058631766587495804, + 0.046197209507226944, + 0.03790789470076561, + 0.023872671648859978, + 0.06814861297607422, + 0.010090320371091366, + 0.019995583221316338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.07005227357149124 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761833786964417 + ], + "opposite_logprob_tokens": [ + -0.8261833786964417 + ], + "logprob_reward": -0.5761833786964417, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761833786964417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08774145692586899, + 0.08490036427974701, + 0.5484009981155396, + 0.07422658056020737, + 0.0891629159450531, + 0.029319079592823982, + 0.017991196364164352, + 0.0212032999843359, + 0.023840216919779778, + 0.023213816806674004 + ], + "credit_map": [ + 0.018726397305727005, + 0.014594141393899918, + 0.010666902177035809, + 0.005105434916913509, + 0.008243110962212086, + 0.01066567376255989, + 0.005502164829522371, + 0.014237627387046814, + 0.019331777468323708, + 0.020437750965356827, + 0.009257566183805466, + 0.005148822907358408, + 0.006137794815003872, + 0.01057485118508339, + 0.005000187549740076, + 0.009011608548462391, + 0.05507347732782364, + 0.10181674361228943, + 0.0944102481007576, + 0.0340542234480381, + 0.034691981971263885, + 0.15540257096290588, + 0.02540726214647293, + 0.047544512897729874, + 0.02116996794939041, + 0.0071981302462518215, + 0.019044553861021996, + 0.004680086392909288, + 0.004071379080414772, + 0.010930625721812248, + 0.0033648470416665077, + 0.003766992362216115, + 0.015560152940452099, + 0.018131671473383904, + 0.008148071356117725, + 0.006863870657980442, + 0.007918572053313255, + 0.02017357014119625, + 0.004763472825288773, + 0.007603537291288376, + 0.008572906255722046, + 0.003708983538672328, + 0.002278231782838702, + 0.0013755593681707978, + 0.0024504284374415874, + 0.006670889910310507, + 0.0017523826099932194, + 0.0025096971075981855, + 0.0039143948815763, + 0.0035598177928477526, + 0.0015824325382709503, + 0.0014940929831936955, + 0.0013294181553646922, + 0.00251742429099977, + 0.0016811738023534417, + 0.0019124412210658193, + 0.005148222204297781, + 0.002147819846868515, + 0.0019331378862261772, + 0.0021592886187136173, + 0.0014969166368246078, + 0.005583538673818111, + 0.0013349844375625253, + 0.0013993914471939206, + 0.0031649416778236628, + 0.003300303127616644, + 0.0031595712061971426, + 0.002438560128211975, + 0.004079680889844894, + 0.003035291563719511, + 0.0023206258192658424, + 0.002341242739930749, + 0.004131497349590063, + 0.004562130197882652, + 0.0029993169009685516, + 0.0031136528123170137, + 0.001956431893631816, + 0.0037749349139630795, + 0.0013503057416528463, + 0.0013255486264824867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.5761833786964417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10679787397384644, + 0.09091237932443619, + 0.4673689901828766, + 0.1925927847623825, + 0.06159497797489166, + 0.010257123969495296, + 0.014925962314009666, + 0.01199042983353138, + 0.014428133144974709, + 0.029131362214684486 + ], + "credit_map": [ + 0.018897579982876778, + 0.018164588138461113, + 0.015306597575545311, + 0.0070206765085458755, + 0.010236294008791447, + 0.012376825325191021, + 0.010790596716105938, + 0.014004720374941826, + 0.022185008972883224, + 0.022919615730643272, + 0.008259954862296581, + 0.005195708479732275, + 0.0067452238872647285, + 0.009436196647584438, + 0.005212904419749975, + 0.01095777191221714, + 0.04869629442691803, + 0.07672280073165894, + 0.0939481183886528, + 0.025838186964392662, + 0.024537544697523117, + 0.14402471482753754, + 0.01886037364602089, + 0.03474096581339836, + 0.07427680492401123, + 0.02094535529613495, + 0.06758632510900497, + 0.007529869209975004, + 0.005042391363531351, + 0.004742820747196674, + 0.00352913374081254, + 0.008940090425312519, + 0.008726407773792744, + 0.012655659578740597, + 0.005978565663099289, + 0.004678106866776943, + 0.005659208167344332, + 0.01233693491667509, + 0.004871852695941925, + 0.006688239984214306, + 0.0014712178381159902, + 0.0012012969236820936, + 0.0017384333768859506, + 0.0008438557852059603, + 0.001454874873161316, + 0.001406132709234953, + 0.0008360244682990015, + 0.0013052872382104397, + 0.0029397080652415752, + 0.0029956104699522257, + 0.0007805934874340892, + 0.0008695912547409534, + 0.0015895982505753636, + 0.002502783201634884, + 0.0013799710432067513, + 0.0018681060755625367, + 0.001987738534808159, + 0.0013569016009569168, + 0.002114255214110017, + 0.0011086498852819204, + 0.001397527288645506, + 0.0018896529218181968, + 0.0009920927695930004, + 0.0011436111526563764, + 0.001850125496275723, + 0.001570993335917592, + 0.0022872176486998796, + 0.001620613969862461, + 0.002830681623890996, + 0.0016561964293941855, + 0.0015192473074421287, + 0.0010930569842457771, + 0.004085589200258255, + 0.004037762060761452, + 0.004516540095210075, + 0.005675035994499922, + 0.0021591600961983204, + 0.004916340112686157, + 0.0013446746161207557, + 0.0023962603881955147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "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\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864410310983658 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.04864410310983658, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864410310983658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2399112433195114, + 0.14034520089626312, + 0.15865445137023926, + 0.15783809125423431, + 0.05013349652290344, + 0.030338071286678314, + 0.039083097130060196, + 0.05137043073773384, + 0.04622211307287216, + 0.08610384166240692 + ], + "credit_map": [ + 0.027787543833255768, + 0.03858489543199539, + 0.03213682770729065, + 0.016772210597991943, + 0.03205181658267975, + 0.02928791381418705, + 0.030533485114574432, + 0.03275652974843979, + 0.022982370108366013, + 0.020318789407610893, + 0.0224022064357996, + 0.008562585338950157, + 0.008996430784463882, + 0.01669614389538765, + 0.01140555553138256, + 0.02898109331727028, + 0.020141897723078728, + 0.021348658949136734, + 0.024572737514972687, + 0.011049590073525906, + 0.014314082451164722, + 0.03996025398373604, + 0.010895408689975739, + 0.016371803358197212, + 0.05123830586671829, + 0.01514220517128706, + 0.0458376482129097, + 0.008438623510301113, + 0.008899669162929058, + 0.013435657136142254, + 0.007159063126891851, + 0.00768691161647439, + 0.009607890620827675, + 0.007783233188092709, + 0.00425462331622839, + 0.00433792220428586, + 0.0050732893869280815, + 0.010130608454346657, + 0.005690552294254303, + 0.003255373565480113, + 0.006107687950134277, + 0.004189549945294857, + 0.002859456231817603, + 0.0035615782253444195, + 0.003400968387722969, + 0.005826111417263746, + 0.002188390353694558, + 0.002204327145591378, + 0.00745749706402421, + 0.008398974314332008, + 0.0033574707340449095, + 0.0023559066466987133, + 0.004147143103182316, + 0.006074989214539528, + 0.0027509634383022785, + 0.0045401486568152905, + 0.011999939568340778, + 0.0037019469309598207, + 0.005139469169080257, + 0.00466394517570734, + 0.0035670555662363768, + 0.014677309431135654, + 0.003716530976817012, + 0.0039042283315211535, + 0.007376234047114849, + 0.006272138096392155, + 0.0062886797823011875, + 0.004877397790551186, + 0.006893558893352747, + 0.006940343417227268, + 0.003942265175282955, + 0.0036314930766820908, + 0.01366347074508667, + 0.01092542801052332, + 0.008923624642193317, + 0.010277282446622849, + 0.013649784959852695, + 0.01686215028166771, + 0.005869161803275347, + 0.005932922475039959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.04864410310983658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a breathtakingly high-speed phantom camera style, a medium shot gracefully circles a high school boy with short, unkempt hair, standing on the rough surface of a city rooftop. The camera's movement is smooth and deliberate, circling him while his own movements unfold in extreme slow motion. He wears a standard white button-down school shirt, the thin fabric fluttering gently in an invisible breeze, and dark trousers. His pose is contemplative, his hands resting lightly on the cold metal safety railing. Illuminated by the cool, ethereal light above and the warm, ambient light of the sprawling city below, his face is etched with a profound sense of exploration; his forehead is slightly wrinkled, his eyes wide as he surveys the sky, his lips barely parted, as if on the verge of a question. In stark and breathtaking contrast to his nearly frozen state, the entire Milky Way galaxy rotates rapidly, time-lapse, across the celestial dome. A dazzling, majestic sweep of vibrant purples, deep blues, and stellar dust paints the sky as a scene of movement. This juxtaposition creates a powerful, awe-inspiring atmosphere of fleeting moments of human introspection contrasted with the vast, indifferent passage of cosmic time.", + "seed": 2290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2290_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": -0.10378965535201132, + "vlm_logprob": -0.10378965535201132, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.10378965535201132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05272377282381058 + ], + "opposite_logprob_tokens": [ + -3.0527238845825195 + ], + "logprob_reward": -0.05272377282381058, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05272377282381058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27277445793151855, + 0.09165361523628235, + 0.04443506523966789, + 0.055597130209207535, + 0.08646828681230545, + 0.08365307748317719, + 0.049107592552900314, + 0.08144249022006989, + 0.13553649187088013, + 0.09933178126811981 + ], + "credit_map": [ + 0.036590445786714554, + 0.03689068555831909, + 0.0296852458268404, + 0.04132997989654541, + 0.0349811352789402, + 0.02757377177476883, + 0.02872304804623127, + 0.03700016438961029, + 0.010824451223015785, + 0.01200716383755207, + 0.015041222795844078, + 0.008197843097150326, + 0.01686142571270466, + 0.008660349063575268, + 0.0058322567492723465, + 0.014228899963200092, + 0.0069771516136825085, + 0.007665316108614206, + 0.004285651259124279, + 0.005719964392483234, + 0.006119794677942991, + 0.004363254643976688, + 0.00325173232704401, + 0.006052200682461262, + 0.009262810461223125, + 0.006328208837658167, + 0.007666009943932295, + 0.006228485144674778, + 0.006971274968236685, + 0.006593958009034395, + 0.005342983175069094, + 0.007203401066362858, + 0.007239160127937794, + 0.010060684755444527, + 0.007299767341464758, + 0.00763589097186923, + 0.023902690038084984, + 0.014610584825277328, + 0.008080966770648956, + 0.007638545706868172, + 0.012686613015830517, + 0.008029856719076633, + 0.013076463714241982, + 0.008094336837530136, + 0.010856537148356438, + 0.014570188708603382, + 0.00688135763630271, + 0.009457721374928951, + 0.008375448174774647, + 0.00763788353651762, + 0.006090841256082058, + 0.005656423978507519, + 0.004888433963060379, + 0.005998366512358189, + 0.004472018219530582, + 0.005988180171698332, + 0.00971142016351223, + 0.008825328201055527, + 0.01598244532942772, + 0.007530405651777983, + 0.008961455896496773, + 0.010413788259029388, + 0.00616763811558485, + 0.013850010000169277, + 0.028757918626070023, + 0.008522478863596916, + 0.019860420376062393, + 0.009016371332108974, + 0.012209093198180199, + 0.03599657118320465, + 0.009448226541280746, + 0.011725413613021374, + 0.010517358779907227, + 0.014449261128902435, + 0.012057951651513577, + 0.010481186211109161, + 0.013135389424860477, + 0.011117642745375633, + 0.01309101190418005, + 0.014481982216238976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.05272377282381058 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084320843219757 + ], + "opposite_logprob_tokens": [ + -4.14108419418335 + ], + "logprob_reward": -0.016084320843219757, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084320843219757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2635897397994995, + 0.27417370676994324, + 0.06751678138971329, + 0.025555336847901344, + 0.032862868160009384, + 0.06939734518527985, + 0.03927662596106529, + 0.0916338637471199, + 0.08860113471746445, + 0.04739262908697128 + ], + "credit_map": [ + 0.027547262609004974, + 0.0251588374376297, + 0.037451278418302536, + 0.048789218068122864, + 0.023577701300382614, + 0.020550351589918137, + 0.027921143919229507, + 0.05259394273161888, + 0.031630609184503555, + 0.0191840510815382, + 0.04859372228384018, + 0.02146022394299507, + 0.03281625732779503, + 0.022502476349473, + 0.025399841368198395, + 0.07258650660514832, + 0.007834076881408691, + 0.006709052249789238, + 0.007588405627757311, + 0.008158770389854908, + 0.009965387172996998, + 0.01162030827254057, + 0.0057564652524888515, + 0.009884312748908997, + 0.003324018558487296, + 0.004464815836399794, + 0.002356395125389099, + 0.003833712311461568, + 0.003651442239060998, + 0.003162445267662406, + 0.0026361942291259766, + 0.002126311883330345, + 0.004585812333971262, + 0.003446799237281084, + 0.0040897950530052185, + 0.0030608063098043203, + 0.0054139019921422005, + 0.005296411924064159, + 0.002479317132383585, + 0.004490022547543049, + 0.010011360049247742, + 0.005600635893642902, + 0.01515461690723896, + 0.004302662331610918, + 0.0053421612828969955, + 0.011707746423780918, + 0.00439973222091794, + 0.012878427281975746, + 0.006873050704598427, + 0.004885570146143436, + 0.00412124814465642, + 0.0038428448606282473, + 0.0036661301273852587, + 0.006819277070462704, + 0.003502164501696825, + 0.005566337611526251, + 0.009228586219251156, + 0.009853064082562923, + 0.017329130321741104, + 0.008241469971835613, + 0.008418053388595581, + 0.009732726961374283, + 0.007524946238845587, + 0.021305885165929794, + 0.012082611210644245, + 0.00768930371850729, + 0.011253906413912773, + 0.006251321639865637, + 0.007040262687951326, + 0.014826005324721336, + 0.008734496310353279, + 0.020723221823573112, + 0.0052176290191709995, + 0.008779394440352917, + 0.006066549103707075, + 0.005485357716679573, + 0.005892319604754448, + 0.005736212246119976, + 0.0040105534717440605, + 0.006204613018780947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.016084320843219757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707461953163147 + ], + "opposite_logprob_tokens": [ + -1.137074589729309 + ], + "logprob_reward": -0.38707461953163147, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38707461953163147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1352025419473648, + 0.0959649458527565, + 0.09827017784118652, + 0.08304545283317566, + 0.05531331151723862, + 0.07853073626756668, + 0.05202322080731392, + 0.08654603362083435, + 0.20497503876686096, + 0.1101284772157669 + ], + "credit_map": [ + 0.017993753775954247, + 0.015191005542874336, + 0.01618514023721218, + 0.02187076397240162, + 0.01436326839029789, + 0.010839756578207016, + 0.015894101932644844, + 0.02286476083099842, + 0.014961364679038525, + 0.01097080111503601, + 0.014785810373723507, + 0.00861032772809267, + 0.008164944127202034, + 0.007068592589348555, + 0.009305327199399471, + 0.02209778130054474, + 0.013786018826067448, + 0.014541877433657646, + 0.007974077947437763, + 0.010771178640425205, + 0.017437050119042397, + 0.013601241633296013, + 0.006977943237870932, + 0.01318079512566328, + 0.00918820220977068, + 0.012199869379401207, + 0.011516399681568146, + 0.00897742249071598, + 0.01335056684911251, + 0.010848893783986568, + 0.006853471044450998, + 0.010110631585121155, + 0.00562185188755393, + 0.006948312744498253, + 0.006928084883838892, + 0.005818321835249662, + 0.00962982326745987, + 0.006687888875603676, + 0.005320125259459019, + 0.008358909748494625, + 0.007831832394003868, + 0.005284901242703199, + 0.019629565998911858, + 0.005060077644884586, + 0.00539175420999527, + 0.01143675483763218, + 0.005561253521591425, + 0.01833460107445717, + 0.0053881132043898106, + 0.003512577386572957, + 0.013135644607245922, + 0.0061036888509988785, + 0.004991555120795965, + 0.0072209397330880165, + 0.0047293431125581264, + 0.006941362749785185, + 0.013186867348849773, + 0.015512113459408283, + 0.015101052820682526, + 0.006952705327421427, + 0.008284134790301323, + 0.01076907105743885, + 0.0062640076503157616, + 0.010476090013980865, + 0.05340910702943802, + 0.017248177900910378, + 0.019941966980695724, + 0.009472377598285675, + 0.013238665647804737, + 0.06488701701164246, + 0.011629241518676281, + 0.015148522332310677, + 0.013303002342581749, + 0.014644908718764782, + 0.012461746111512184, + 0.009480148553848267, + 0.01148782018572092, + 0.013880414888262749, + 0.01273387111723423, + 0.022136570885777473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.38707461953163147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010417040903121233 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010417040903121233, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010417040903121233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15827372670173645, + 0.10976356267929077, + 0.11774640530347824, + 0.0438690185546875, + 0.08654395490884781, + 0.0284583680331707, + 0.06473031640052795, + 0.059177231043577194, + 0.14552772045135498, + 0.18590961396694183 + ], + "credit_map": [ + 0.017922773957252502, + 0.015553574077785015, + 0.018722886219620705, + 0.021581703796982765, + 0.019674818962812424, + 0.016113806515932083, + 0.016416573897004128, + 0.03228762000799179, + 0.014203082770109177, + 0.009491519071161747, + 0.019032221287488937, + 0.009253503754734993, + 0.013893377035856247, + 0.008463112637400627, + 0.009038424119353294, + 0.02638833597302437, + 0.019491873681545258, + 0.011704222299158573, + 0.01929345913231373, + 0.012039032764732838, + 0.013501673005521297, + 0.010128969326615334, + 0.009327356703579426, + 0.02225983887910843, + 0.006347308866679668, + 0.00593867851421237, + 0.004814279731363058, + 0.00412439089268446, + 0.005423476919531822, + 0.007186883594840765, + 0.0040183174423873425, + 0.006015688646584749, + 0.009522082284092903, + 0.00954399537295103, + 0.010104542598128319, + 0.006433488335460424, + 0.022040128707885742, + 0.013611171394586563, + 0.005342527292668819, + 0.00994602870196104, + 0.004103740677237511, + 0.0025913340505212545, + 0.0034774027299135923, + 0.0031398346181958914, + 0.0028543199878185987, + 0.0059415302239358425, + 0.003005948616191745, + 0.003344262484461069, + 0.009163794107735157, + 0.007773593068122864, + 0.011451606638729572, + 0.008374829776585102, + 0.005582037381827831, + 0.007193605415523052, + 0.008104195818305016, + 0.007086664438247681, + 0.012207381427288055, + 0.006761178839951754, + 0.007509255316108465, + 0.004383246414363384, + 0.005973353516310453, + 0.01029537059366703, + 0.005122622009366751, + 0.00692483177408576, + 0.04117240384221077, + 0.009963658638298512, + 0.013070452958345413, + 0.0055979471653699875, + 0.008841309696435928, + 0.0503666028380394, + 0.006783395539969206, + 0.009731979109346867, + 0.025557449087500572, + 0.018627604469656944, + 0.01929124817252159, + 0.01571398228406906, + 0.020175186917185783, + 0.03286764398217201, + 0.01739780232310295, + 0.03627872094511986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0010417040903121233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202385947108269 + ], + "opposite_logprob_tokens": [ + -2.812023878097534 + ], + "logprob_reward": -0.06202385947108269, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202385947108269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21638454496860504, + 0.0574374683201313, + 0.06159352511167526, + 0.061920467764139175, + 0.060307107865810394, + 0.07317225635051727, + 0.03380218893289566, + 0.026809675619006157, + 0.1967201828956604, + 0.211852565407753 + ], + "credit_map": [ + 0.018041087314486504, + 0.023646071553230286, + 0.02207200787961483, + 0.0400562658905983, + 0.026973038911819458, + 0.02063228189945221, + 0.025461677461862564, + 0.0395021066069603, + 0.0069098141975700855, + 0.0061823283322155476, + 0.008455147966742516, + 0.006223277188837528, + 0.005669985432177782, + 0.005121806636452675, + 0.005145516246557236, + 0.01372959278523922, + 0.006249666679650545, + 0.011641371995210648, + 0.0063080391846597195, + 0.008357478305697441, + 0.006610080599784851, + 0.0057915267534554005, + 0.006442595273256302, + 0.010192766785621643, + 0.008542705327272415, + 0.008270928636193275, + 0.005649097263813019, + 0.004962926730513573, + 0.009357865899801254, + 0.009626150131225586, + 0.005408587399870157, + 0.010102206841111183, + 0.009041284210979939, + 0.006773295812308788, + 0.007241915445774794, + 0.007029911037534475, + 0.009475696831941605, + 0.007477113977074623, + 0.006491350941359997, + 0.006776542402803898, + 0.010058811865746975, + 0.008755507878959179, + 0.007101852912455797, + 0.011234395205974579, + 0.011063283309340477, + 0.008971056900918484, + 0.009034491144120693, + 0.0069528622552752495, + 0.003907368052750826, + 0.0034342820290476084, + 0.0035954120103269815, + 0.00450516864657402, + 0.005019030999392271, + 0.005312645807862282, + 0.003440557047724724, + 0.00458772387355566, + 0.004054018761962652, + 0.00360271823592484, + 0.0036287333350628614, + 0.0023790206760168076, + 0.003236791817471385, + 0.003414080711081624, + 0.0028973203152418137, + 0.0035969926975667477, + 0.04901198670268059, + 0.01667812280356884, + 0.020494656637310982, + 0.009660519659519196, + 0.015374715439975262, + 0.05881958454847336, + 0.012639728374779224, + 0.01404088269919157, + 0.023008376359939575, + 0.018778573721647263, + 0.021969053894281387, + 0.01816718280315399, + 0.02193407341837883, + 0.027023257687687874, + 0.03492201119661331, + 0.046050045639276505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.06202385947108269 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a breathtakingly high-speed phantom camera style, a medium shot gracefully circles a high school boy with short, unkempt hair, standing on the rough surface of a city rooftop. The camera's movement is smooth and deliberate, circling him while his own movements unfold in extreme slow motion. He wears a standard white button-down school shirt, the thin fabric fluttering gently in an invisible breeze, and dark trousers. His pose is contemplative, his hands resting lightly on the cold metal safety railing. Illuminated by the cool, ethereal light above and the warm, ambient light of the sprawling city below, his face is etched with a profound sense of exploration; his forehead is slightly wrinkled, his eyes wide as he surveys the sky, his lips barely parted, as if on the verge of a question. In stark and breathtaking contrast to his nearly frozen state, the entire Milky Way galaxy rotates rapidly, time-lapse, across the celestial dome. A dazzling, majestic sweep of vibrant purples, deep blues, and stellar dust paints the sky as a scene of movement. This juxtaposition creates a powerful, awe-inspiring atmosphere of fleeting moments of human introspection contrasted with the vast, indifferent passage of cosmic time.", + "seed": 2291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_1/index_848_seed_2291_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": -0.09627053689910099, + "vlm_logprob": -0.09627053689910099, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.09627053689910099, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08114692568778992 + ], + "opposite_logprob_tokens": [ + -2.5811469554901123 + ], + "logprob_reward": -0.08114692568778992, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08114692568778992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2873985469341278, + 0.07580938935279846, + 0.04985243082046509, + 0.07874961197376251, + 0.0724356397986412, + 0.038301702588796616, + 0.02478046715259552, + 0.15013308823108673, + 0.0595422200858593, + 0.16299696266651154 + ], + "credit_map": [ + 0.028972480446100235, + 0.05351753532886505, + 0.03384329006075859, + 0.027622664347290993, + 0.058694932609796524, + 0.030018392950296402, + 0.03146219998598099, + 0.023267054930329323, + 0.011061507277190685, + 0.012978209182620049, + 0.005826494190841913, + 0.008022891357541084, + 0.01341088768094778, + 0.008195352740585804, + 0.008413678035140038, + 0.007900364696979523, + 0.00521863391622901, + 0.008495866321027279, + 0.007680944632738829, + 0.004724116530269384, + 0.006390465423464775, + 0.004983451683074236, + 0.006115064490586519, + 0.006243890151381493, + 0.006740953307598829, + 0.009923039004206657, + 0.007867311127483845, + 0.018556978553533554, + 0.007775011006742716, + 0.006400106009095907, + 0.016138466075062752, + 0.00534774549305439, + 0.006825478281825781, + 0.009314317256212234, + 0.008211801759898663, + 0.008048390038311481, + 0.010778002440929413, + 0.007819602265954018, + 0.012973298318684101, + 0.008464743383228779, + 0.004611543845385313, + 0.004229940474033356, + 0.004759913310408592, + 0.004372659604996443, + 0.005187822040170431, + 0.003943145740777254, + 0.0064522335305809975, + 0.004744444973766804, + 0.0037911273539066315, + 0.0033222706988453865, + 0.002655015792697668, + 0.0033549705985933542, + 0.002964768325909972, + 0.002739297691732645, + 0.0029774578288197517, + 0.0029755588620901108, + 0.01914636418223381, + 0.01960580237209797, + 0.021284494549036026, + 0.0234321691095829, + 0.02085593342781067, + 0.017378462478518486, + 0.01585180126130581, + 0.01257806271314621, + 0.014327775686979294, + 0.004406666848808527, + 0.007072437088936567, + 0.006727232132107019, + 0.003576727584004402, + 0.015136338770389557, + 0.0038402138743549585, + 0.00445482786744833, + 0.01960376277565956, + 0.0198562853038311, + 0.016784070059657097, + 0.0315479151904583, + 0.01586226001381874, + 0.0211134385317564, + 0.018326401710510254, + 0.019902821630239487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.08114692568778992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012556660920381546 + ], + "opposite_logprob_tokens": [ + -4.387556552886963 + ], + "logprob_reward": -0.012556660920381546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012556660920381546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2817062735557556, + 0.07618149369955063, + 0.05152056738734245, + 0.08614739030599594, + 0.05094638466835022, + 0.05274453014135361, + 0.06079429015517235, + 0.07951308786869049, + 0.11932754516601562, + 0.1411183774471283 + ], + "credit_map": [ + 0.04187647998332977, + 0.05021892115473747, + 0.04726279526948929, + 0.027023427188396454, + 0.03672870993614197, + 0.026113392785191536, + 0.021251903846859932, + 0.031230656430125237, + 0.008031735196709633, + 0.017674850299954414, + 0.005343106109648943, + 0.009243486449122429, + 0.010830070823431015, + 0.008626469410955906, + 0.008802664466202259, + 0.0076291062869131565, + 0.005288511514663696, + 0.008892695419490337, + 0.004528098739683628, + 0.005011480767279863, + 0.006580370478332043, + 0.005251028575003147, + 0.011223585344851017, + 0.004744793754070997, + 0.010367205366492271, + 0.02064605802297592, + 0.006416722200810909, + 0.014108485542237759, + 0.008660254068672657, + 0.0093750711530447, + 0.011508973315358162, + 0.00506462249904871, + 0.005137200001627207, + 0.0082094119861722, + 0.006620642729103565, + 0.0055542937479913235, + 0.0078056575730443, + 0.006610755808651447, + 0.004720618017017841, + 0.006287803873419762, + 0.006175908260047436, + 0.007570888847112656, + 0.004132410977035761, + 0.0062269652262330055, + 0.005993460305035114, + 0.005674403160810471, + 0.011875971220433712, + 0.005094524472951889, + 0.009904718026518822, + 0.007670390885323286, + 0.005600243341177702, + 0.006130346097052097, + 0.008562020026147366, + 0.00758809270337224, + 0.006635912694036961, + 0.008702565915882587, + 0.01199878379702568, + 0.009480173699557781, + 0.01098219957202673, + 0.010073709301650524, + 0.008180254139006138, + 0.009516200050711632, + 0.007911947555840015, + 0.011369817890226841, + 0.036011721938848495, + 0.006071287672966719, + 0.008004971779882908, + 0.004546534735709429, + 0.0075324964709579945, + 0.04613529145717621, + 0.00464546587318182, + 0.006379777565598488, + 0.019545434042811394, + 0.012751434929668903, + 0.01695123128592968, + 0.01634524017572403, + 0.01649930700659752, + 0.0317629799246788, + 0.012417806312441826, + 0.014844940975308418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.012556660920381546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34864580631256104 + ], + "opposite_logprob_tokens": [ + -1.223645806312561 + ], + "logprob_reward": -0.34864580631256104, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34864580631256104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24748602509498596, + 0.13720352947711945, + 0.08541267365217209, + 0.06359407305717468, + 0.05310720577836037, + 0.02754197269678116, + 0.06905637681484222, + 0.03272629529237747, + 0.21246442198753357, + 0.07140747457742691 + ], + "credit_map": [ + 0.029817523434758186, + 0.050454795360565186, + 0.03635243698954582, + 0.023405928164720535, + 0.03704259172081947, + 0.02530735544860363, + 0.021879862993955612, + 0.023225516080856323, + 0.01643027365207672, + 0.04167723283171654, + 0.009559313766658306, + 0.012432985007762909, + 0.01451078150421381, + 0.01904616504907608, + 0.011988716199994087, + 0.011558066122233868, + 0.008565566502511501, + 0.014268727041780949, + 0.01139909215271473, + 0.009903712198138237, + 0.014655932784080505, + 0.00947205163538456, + 0.010487415827810764, + 0.006660186219960451, + 0.006101588718593121, + 0.00881807878613472, + 0.0061541409231722355, + 0.011804645881056786, + 0.007102270610630512, + 0.005686422344297171, + 0.013619477860629559, + 0.0043074446730315685, + 0.006827293895184994, + 0.00886840932071209, + 0.006224785465747118, + 0.00392126152291894, + 0.009226460009813309, + 0.006542734336107969, + 0.004191017709672451, + 0.007305242121219635, + 0.004060887731611729, + 0.0039047140162438154, + 0.0028795055113732815, + 0.0030108573846518993, + 0.0032342243939638138, + 0.0034152758307754993, + 0.004341731779277325, + 0.0026947755832225084, + 0.010221918113529682, + 0.007615023758262396, + 0.006756435614079237, + 0.006241367198526859, + 0.011181998997926712, + 0.009382064454257488, + 0.0067303976975381374, + 0.010927168652415276, + 0.0048341304063797, + 0.004464366007596254, + 0.005089216399937868, + 0.004470258951187134, + 0.0040839496068656445, + 0.003941149450838566, + 0.003089905483648181, + 0.002753320150077343, + 0.06840252876281738, + 0.00927420798689127, + 0.012656569480895996, + 0.005496196448802948, + 0.008671918883919716, + 0.08922599256038666, + 0.007245182059705257, + 0.011491810902953148, + 0.008639788255095482, + 0.005856962408870459, + 0.010786931030452251, + 0.006256697699427605, + 0.008983852341771126, + 0.014364592730998993, + 0.006044528912752867, + 0.010474123060703278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.34864580631256104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009199915803037584 + ], + "opposite_logprob_tokens": [ + -7.000919818878174 + ], + "logprob_reward": -0.0009199915803037584, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009199915803037584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1522502452135086, + 0.19741953909397125, + 0.0627312958240509, + 0.062517449259758, + 0.027983473613858223, + 0.030925894156098366, + 0.05951488018035889, + 0.03198470175266266, + 0.05798272043466568, + 0.3166898488998413 + ], + "credit_map": [ + 0.01771351508796215, + 0.02990051917731762, + 0.021029934287071228, + 0.014231703244149685, + 0.025329714640975, + 0.01528654433786869, + 0.01514397282153368, + 0.013614349067211151, + 0.029975485056638718, + 0.06208367273211479, + 0.015162521973252296, + 0.012238211929798126, + 0.021145835518836975, + 0.025548893958330154, + 0.011998171918094158, + 0.019266733899712563, + 0.00588873028755188, + 0.007556798867881298, + 0.009470964781939983, + 0.010795745067298412, + 0.006385415326803923, + 0.00481907045468688, + 0.011581694707274437, + 0.006232877261936665, + 0.006801903247833252, + 0.013668393716216087, + 0.004712030291557312, + 0.011275120079517365, + 0.006534255109727383, + 0.006432284601032734, + 0.009618754498660564, + 0.003474703524261713, + 0.004333801567554474, + 0.004123086109757423, + 0.0036934937816113234, + 0.0026928801089525223, + 0.0032911028247326612, + 0.0037799966521561146, + 0.0025506021920591593, + 0.003518511075526476, + 0.00575254624709487, + 0.004542285110801458, + 0.0034132080618292093, + 0.0029080563690513372, + 0.0038274675607681274, + 0.004026924259960651, + 0.0036075632087886333, + 0.0028478442691266537, + 0.010769560001790524, + 0.00837700441479683, + 0.005409650970250368, + 0.0058535304851830006, + 0.008664073422551155, + 0.007245906628668308, + 0.005949873942881823, + 0.007245279848575592, + 0.004409102723002434, + 0.004365698900073767, + 0.0053617628291249275, + 0.00427671056240797, + 0.004265135154128075, + 0.003510786220431328, + 0.0028044278733432293, + 0.0029910814482718706, + 0.01558554545044899, + 0.003907474223524332, + 0.005152215715497732, + 0.0035685927141457796, + 0.0038911388255655766, + 0.01975622959434986, + 0.0031460225582122803, + 0.0029755039140582085, + 0.036517709493637085, + 0.025260314345359802, + 0.05295148491859436, + 0.023795200511813164, + 0.04067932814359665, + 0.06753885000944138, + 0.028863588348031044, + 0.04108336195349693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.0009199915803037584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high school boy with short, unkempt hair wearing a white button-down shirt and dark trousers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boy standing on a rough city rooftop surface next to a metal safety railing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boy's movement captured in extreme slow motion while the Milky Way galaxy rotates rapidly in a time-lapse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky filled with a time-lapse of the Milky Way galaxy showing vibrant purples, deep blues, and stellar dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a medium shot that circles the boy smoothly and deliberately?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808329999446869 + ], + "opposite_logprob_tokens": [ + -3.28808331489563 + ], + "logprob_reward": -0.03808329999446869, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808329999446869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14691339433193207, + 0.11566973477602005, + 0.11562201380729675, + 0.059392932802438736, + 0.06109980121254921, + 0.09022241085767746, + 0.043494515120983124, + 0.059716686606407166, + 0.22085118293762207, + 0.08701729029417038 + ], + "credit_map": [ + 0.021401336416602135, + 0.022710813209414482, + 0.02187885157763958, + 0.014844704419374466, + 0.02381870523095131, + 0.012935785576701164, + 0.013423706404864788, + 0.015899505466222763, + 0.015439731068909168, + 0.01763973757624626, + 0.008160081692039967, + 0.013921425677835941, + 0.024280350655317307, + 0.012285187840461731, + 0.013684455305337906, + 0.010258766822516918, + 0.00986466184258461, + 0.014684738591313362, + 0.008625585585832596, + 0.010695974342525005, + 0.025190627202391624, + 0.014144519343972206, + 0.024016086012125015, + 0.008399823680520058, + 0.005624633748084307, + 0.00906919315457344, + 0.0050948369316756725, + 0.00863728765398264, + 0.013001454062759876, + 0.006838418543338776, + 0.006715300492942333, + 0.004411810543388128, + 0.009249960072338581, + 0.010078124701976776, + 0.006892289500683546, + 0.005686304997652769, + 0.007600642740726471, + 0.0074563161469995975, + 0.00825080368667841, + 0.005885360296815634, + 0.00951083842664957, + 0.011043621227145195, + 0.007991493679583073, + 0.011444156058132648, + 0.009752240031957626, + 0.008726716041564941, + 0.02311677299439907, + 0.008636574260890484, + 0.005692009348422289, + 0.006047938484698534, + 0.004346388392150402, + 0.006824154406785965, + 0.005297368858009577, + 0.004589271731674671, + 0.004531910642981529, + 0.006165476515889168, + 0.008678288199007511, + 0.008829300291836262, + 0.00875312089920044, + 0.007283745799213648, + 0.007269698195159435, + 0.0071783484891057014, + 0.005960023961961269, + 0.005764163564890623, + 0.06551604717969894, + 0.009618000127375126, + 0.014837335795164108, + 0.007759832311421633, + 0.011906764470040798, + 0.09390735626220703, + 0.008594853803515434, + 0.008710985071957111, + 0.010671386495232582, + 0.007389386650174856, + 0.012424381449818611, + 0.007314668968319893, + 0.010623503476381302, + 0.017392240464687347, + 0.008504696190357208, + 0.012697027064859867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.03808329999446869 + } + ], + "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/0000152/group_10/index_tensor([4879, 4879])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_tensor([4879, 4879])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3e7ad9773b68ad8be8947dd29a13bdd1c86dc3ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_tensor([4879, 4879])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, low-angle fisheye lens positions the viewer on the deck of a massive wooden sailing ship, ploughing through the surging waves. At the center of the frame, several heroes, clad in ornate, vintage armor, stand firm, their expressions resolute and their gazes piercing, as they gaze out upon the turbulent distance. The leading male hero, his long hair dancing wildly in the wind, draped in dark armor intricately carved with metal patterns and a thick, crimson cape. His companions, some wielding longbows, others clutching battle axes, their clothing and weapons bearing the marks of age. The fisheye lens's distortion exaggerates the curvature of the ship's hull and the towering mainsail, its canvas billowing and rustling in the fierce wind. The damp, dark wooden deck reflects the indirect, cold light of the gloomy sky, while the splashing whitewater waves all around add a sense of dynamism to the image. The color saturation of the entire picture is deliberately reduced, presenting a faded retro nostalgic film texture with obvious grain and slight dark corners, creating an epic, tragic yet hopeful expedition atmosphere.", + "seed": 7539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7539_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": -0.014947421476244926, + "vlm_logprob": -0.014947421476244926, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.014947421476244926, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02525939978659153 + ], + "opposite_logprob_tokens": [ + -4.275259494781494 + ], + "logprob_reward": -0.02525939978659153, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02525939978659153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16824281215667725, + 0.07072963565587997, + 0.041462793946266174, + 0.08686485886573792, + 0.23161053657531738, + 0.0679091215133667, + 0.030223559588193893, + 0.22303539514541626, + 0.03121541626751423, + 0.04870584234595299 + ], + "credit_map": [ + 0.031444840133190155, + 0.011739213950932026, + 0.013520372100174427, + 0.016768164932727814, + 0.009080972522497177, + 0.011957020498812199, + 0.026325704529881477, + 0.04740651324391365, + 0.005952274426817894, + 0.010430630296468735, + 0.011608632281422615, + 0.014804841950535774, + 0.008813955821096897, + 0.0054997545666992664, + 0.006651689298450947, + 0.0069678593426942825, + 0.0065189870074391365, + 0.0034848435316234827, + 0.0032257067505270243, + 0.003996673971414566, + 0.00570659851655364, + 0.004510032944381237, + 0.010158544406294823, + 0.0038614056538790464, + 0.007196882739663124, + 0.011694565415382385, + 0.017448069527745247, + 0.01170775480568409, + 0.013796512968838215, + 0.00864408165216446, + 0.00901042390614748, + 0.007366567850112915, + 0.02978566475212574, + 0.012476465664803982, + 0.019377408549189568, + 0.03473596274852753, + 0.022169917821884155, + 0.01367226429283619, + 0.08091040700674057, + 0.018482431769371033, + 0.0062186988070607185, + 0.0072241174057126045, + 0.0056547923013567924, + 0.011142022907733917, + 0.005797181278467178, + 0.004019846674054861, + 0.02304496243596077, + 0.0048075043596327305, + 0.0028059810865670443, + 0.002857392420992255, + 0.0038724064361304045, + 0.0034579895436763763, + 0.00646940479055047, + 0.0037156452890485525, + 0.0041435351595282555, + 0.0029012036975473166, + 0.012898549437522888, + 0.062198493629693985, + 0.029572082683444023, + 0.07821548730134964, + 0.010095560923218727, + 0.010463664308190346, + 0.014208114705979824, + 0.005383448675274849, + 0.0039968634955585, + 0.004602605476975441, + 0.004675926174968481, + 0.004777258727699518, + 0.00303510925732553, + 0.0030630514957010746, + 0.0038374224677681923, + 0.0032271789386868477, + 0.00693198898807168, + 0.009348609484732151, + 0.0069037992507219315, + 0.004977100994437933, + 0.006041587796062231, + 0.0055089108645915985, + 0.005344128701835871, + 0.0036497139371931553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.02525939978659153 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005279882811009884 + ], + "opposite_logprob_tokens": [ + -5.255280017852783 + ], + "logprob_reward": -0.005279882811009884, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005279882811009884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1678156554698944, + 0.045921750366687775, + 0.020060904324054718, + 0.056152358651161194, + 0.06901442259550095, + 0.022992528975009918, + 0.03473576158285141, + 0.49318361282348633, + 0.0250617116689682, + 0.06506124138832092 + ], + "credit_map": [ + 0.03184403479099274, + 0.010216468013823032, + 0.010613790713250637, + 0.013024207204580307, + 0.008377498015761375, + 0.01593136042356491, + 0.027480758726596832, + 0.05032753944396973, + 0.003488435409963131, + 0.004731032531708479, + 0.003982344642281532, + 0.006113994400948286, + 0.0046478621661663055, + 0.0053222705610096455, + 0.012504401616752148, + 0.005131410900503397, + 0.0017108051106333733, + 0.0020863611716777086, + 0.003246862906962633, + 0.0027299667708575726, + 0.0023594594094902277, + 0.002915623364970088, + 0.002007654868066311, + 0.0030041709542274475, + 0.005687408149242401, + 0.00848417915403843, + 0.008120360784232616, + 0.007161399815231562, + 0.009727631695568562, + 0.005402472801506519, + 0.005944320932030678, + 0.0056245881132781506, + 0.008120384998619556, + 0.00838033389300108, + 0.008266600780189037, + 0.008600600995123386, + 0.009073571301996708, + 0.007533506955951452, + 0.014708402566611767, + 0.00433102622628212, + 0.0014588820049539208, + 0.0017521115951240063, + 0.0032240706495940685, + 0.003925369121134281, + 0.0023808982223272324, + 0.003167488379403949, + 0.0035647286567837, + 0.00351897906512022, + 0.003474748693406582, + 0.004139347467571497, + 0.0036207898519933224, + 0.004300237633287907, + 0.0067663369700312614, + 0.003911804407835007, + 0.005004796665161848, + 0.0035176980309188366, + 0.03025020658969879, + 0.14397279918193817, + 0.06262219697237015, + 0.167860746383667, + 0.023317039012908936, + 0.023049408569931984, + 0.030504336580634117, + 0.011606873944401741, + 0.0042162067256867886, + 0.0038587860763072968, + 0.003720490727573633, + 0.0028829125221818686, + 0.003127416828647256, + 0.0023148476611822844, + 0.0027119265869259834, + 0.0022291243076324463, + 0.00898708775639534, + 0.014550945721566677, + 0.007523234002292156, + 0.007112203165888786, + 0.008531209081411362, + 0.0058365813456475735, + 0.0074387239292263985, + 0.005081261973828077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.005279882811009884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545477598905563 + ], + "opposite_logprob_tokens": [ + -4.387545585632324 + ], + "logprob_reward": -0.012545477598905563, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545477598905563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1333661675453186, + 0.1389302909374237, + 0.03439944609999657, + 0.06461288034915924, + 0.08000914007425308, + 0.06260542571544647, + 0.05576324090361595, + 0.30939042568206787, + 0.05055128410458565, + 0.07037164270877838 + ], + "credit_map": [ + 0.0254606194794178, + 0.00820712000131607, + 0.007555027026683092, + 0.012290733866393566, + 0.00611913064494729, + 0.012561752460896969, + 0.02194582298398018, + 0.03922596573829651, + 0.011152515187859535, + 0.021168025210499763, + 0.017448972910642624, + 0.023461947217583656, + 0.022214453667402267, + 0.014532906003296375, + 0.01345161721110344, + 0.01549984049052, + 0.0051374915055930614, + 0.0033008374739438295, + 0.002488731639459729, + 0.0033911122009158134, + 0.003985133953392506, + 0.0041776434518396854, + 0.008536730892956257, + 0.0033817640505731106, + 0.007644809782505035, + 0.007383766118437052, + 0.0044869668781757355, + 0.006809903774410486, + 0.014478281140327454, + 0.009906649589538574, + 0.007587810512632132, + 0.006314696744084358, + 0.007701240945607424, + 0.014272364787757397, + 0.011094500310719013, + 0.008783756755292416, + 0.013091124594211578, + 0.009073958732187748, + 0.008847223594784737, + 0.007144967094063759, + 0.0038696760311722755, + 0.005257111042737961, + 0.007345524616539478, + 0.007942096330225468, + 0.007376824505627155, + 0.007457299157977104, + 0.016149356961250305, + 0.007207539398223162, + 0.005827709566801786, + 0.0049066743813455105, + 0.007639463525265455, + 0.008292600512504578, + 0.00908189732581377, + 0.006621083244681358, + 0.007811777759343386, + 0.005582035984843969, + 0.01882059872150421, + 0.08840827643871307, + 0.0403759740293026, + 0.102122463285923, + 0.01539691537618637, + 0.01441400870680809, + 0.02185007743537426, + 0.008002136833965778, + 0.0073938677087426186, + 0.0056519778445363045, + 0.006813114043325186, + 0.009242077358067036, + 0.006855332758277655, + 0.003226611064746976, + 0.007036182098090649, + 0.004332124255597591, + 0.011675816029310226, + 0.006824738346040249, + 0.007999134249985218, + 0.009566737338900566, + 0.007790518458932638, + 0.008454683236777782, + 0.009728712029755116, + 0.008331301622092724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.012545477598905563 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058078907430172 + ], + "opposite_logprob_tokens": [ + -3.895580768585205 + ], + "logprob_reward": -0.02058078907430172, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058078907430172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15132173895835876, + 0.12833240628242493, + 0.02208414301276207, + 0.06302456557750702, + 0.050845079123973846, + 0.036135315895080566, + 0.02358204498887062, + 0.4239782989025116, + 0.07501876354217529, + 0.025677690282464027 + ], + "credit_map": [ + 0.027138089761137962, + 0.011333663016557693, + 0.010114083997905254, + 0.013311518356204033, + 0.009653197601437569, + 0.015762094408273697, + 0.02309337630867958, + 0.04091571643948555, + 0.010948858223855495, + 0.022512709721922874, + 0.013873325660824776, + 0.02047775126993656, + 0.02276620827615261, + 0.013159205205738544, + 0.012172811664640903, + 0.012421525083482265, + 0.002533901482820511, + 0.0022372056264430285, + 0.001674727420322597, + 0.0022846756037324667, + 0.003273541573435068, + 0.002882503205910325, + 0.00426435424014926, + 0.002933232579380274, + 0.005367594305425882, + 0.009003588929772377, + 0.01131370011717081, + 0.00842224806547165, + 0.010596680454909801, + 0.005717888008803129, + 0.006246594712138176, + 0.0063562714494764805, + 0.006307578645646572, + 0.004084847867488861, + 0.004860284738242626, + 0.006692732218652964, + 0.005845602136105299, + 0.0038126539438962936, + 0.01495671272277832, + 0.004284669179469347, + 0.002396394731476903, + 0.0030242870561778545, + 0.0059528653509914875, + 0.005918061826378107, + 0.005307379644364119, + 0.002648084657266736, + 0.007562202401459217, + 0.0033260404597967863, + 0.0020923519041389227, + 0.0016944264061748981, + 0.0025873035192489624, + 0.0037648456636816263, + 0.005683312192559242, + 0.002378273056820035, + 0.003361629555001855, + 0.002019901992753148, + 0.026371615007519722, + 0.12545424699783325, + 0.053162846714258194, + 0.14363814890384674, + 0.019687296822667122, + 0.01967911422252655, + 0.0264937411993742, + 0.009491272270679474, + 0.014464864507317543, + 0.010920447297394276, + 0.014575985260307789, + 0.008824142627418041, + 0.006552061531692743, + 0.006572376005351543, + 0.007282445207238197, + 0.005826442968100309, + 0.00387356779538095, + 0.0030203652568161488, + 0.00396345229819417, + 0.0027348638977855444, + 0.002915371209383011, + 0.0025663580745458603, + 0.0035482218954712152, + 0.003055489156395197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.02058078907430172 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071558110415936 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071558110415936, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071558110415936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06400419026613235, + 0.10902859270572662, + 0.04832906275987625, + 0.0612790584564209, + 0.376899391412735, + 0.053189463913440704, + 0.03559381887316704, + 0.16288378834724426, + 0.051981180906295776, + 0.036811381578445435 + ], + "credit_map": [ + 0.004871814511716366, + 0.00947664026170969, + 0.007616143673658371, + 0.008240309543907642, + 0.012747662141919136, + 0.006458722986280918, + 0.007916921749711037, + 0.006675989367067814, + 0.007547708693891764, + 0.017784351482987404, + 0.011884917505085468, + 0.019015317782759666, + 0.01739201508462429, + 0.01024656742811203, + 0.015253892168402672, + 0.009903826750814915, + 0.007177205290645361, + 0.003928518854081631, + 0.006085237488150597, + 0.005269454792141914, + 0.004526402801275253, + 0.0031777755357325077, + 0.013324059545993805, + 0.004840410780161619, + 0.005291169509291649, + 0.00844629667699337, + 0.010818790644407272, + 0.007683677598834038, + 0.010768687352538109, + 0.006083540618419647, + 0.005934187676757574, + 0.006252713967114687, + 0.034259941428899765, + 0.045784223824739456, + 0.043989259749650955, + 0.04810480400919914, + 0.04333164542913437, + 0.04239518567919731, + 0.07891850918531418, + 0.04011586308479309, + 0.0051823267713189125, + 0.0034896081779152155, + 0.0063227093778550625, + 0.008625737391412258, + 0.0039024436846375465, + 0.0031964266672730446, + 0.018142210319638252, + 0.004328002687543631, + 0.0036319997161626816, + 0.0038519196677953005, + 0.0045356121845543385, + 0.005250478629022837, + 0.005308023653924465, + 0.003990722820162773, + 0.005432512145489454, + 0.0035925537813454866, + 0.00928548350930214, + 0.04314872622489929, + 0.020379455760121346, + 0.051905855536460876, + 0.0103121567517519, + 0.010011669248342514, + 0.011438323184847832, + 0.006402133964002132, + 0.009886202402412891, + 0.006703912280499935, + 0.007761731743812561, + 0.0069444808177649975, + 0.005014313384890556, + 0.004865340888500214, + 0.0057823131792247295, + 0.005022888537496328, + 0.00753745948895812, + 0.006557738408446312, + 0.0037929280661046505, + 0.0031010557431727648, + 0.005299287382513285, + 0.0034331839997321367, + 0.004224695265293121, + 0.0028650350868701935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.011071558110415936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle fisheye lens positions the viewer on the deck of a massive wooden sailing ship, ploughing through the surging waves. At the center of the frame, several heroes, clad in ornate, vintage armor, stand firm, their expressions resolute and their gazes piercing, as they gaze out upon the turbulent distance. The leading male hero, his long hair dancing wildly in the wind, draped in dark armor intricately carved with metal patterns and a thick, crimson cape. His companions, some wielding longbows, others clutching battle axes, their clothing and weapons bearing the marks of age. The fisheye lens's distortion exaggerates the curvature of the ship's hull and the towering mainsail, its canvas billowing and rustling in the fierce wind. The damp, dark wooden deck reflects the indirect, cold light of the gloomy sky, while the splashing whitewater waves all around add a sense of dynamism to the image. The color saturation of the entire picture is deliberately reduced, presenting a faded retro nostalgic film texture with obvious grain and slight dark corners, creating an epic, tragic yet hopeful expedition atmosphere.", + "seed": 7540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7540_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": -0.021064431872218847, + "vlm_logprob": -0.021064431872218847, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.021064431872218847, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02070503681898117 + ], + "opposite_logprob_tokens": [ + -5.395705223083496 + ], + "logprob_reward": -0.02070503681898117, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02070503681898117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11770500987768173, + 0.04838273674249649, + 0.06914740800857544, + 0.15890809893608093, + 0.23918657004833221, + 0.021285193040966988, + 0.1645495742559433, + 0.028969045728445053, + 0.12522193789482117, + 0.02664441615343094 + ], + "credit_map": [ + 0.013251153752207756, + 0.009019934572279453, + 0.024599822238087654, + 0.011883660219609737, + 0.009782454930245876, + 0.014747640118002892, + 0.01515575684607029, + 0.01926458068192005, + 0.00515196705237031, + 0.008039869368076324, + 0.0047719525173306465, + 0.006315719336271286, + 0.0037995800375938416, + 0.007600490469485521, + 0.005893950816243887, + 0.006809206213802099, + 0.006014879327267408, + 0.007785232271999121, + 0.007426009513437748, + 0.012303704395890236, + 0.006181606091558933, + 0.004944177344441414, + 0.01051686704158783, + 0.013974931091070175, + 0.020022809505462646, + 0.006293827667832375, + 0.013930370099842548, + 0.01727757602930069, + 0.008169225417077541, + 0.008128085173666477, + 0.028106672689318657, + 0.056979529559612274, + 0.03152373805642128, + 0.015697380527853966, + 0.026087310165166855, + 0.04480824992060661, + 0.014826486818492413, + 0.02508079819381237, + 0.032339390367269516, + 0.04882321134209633, + 0.0029012414161115885, + 0.001854137284681201, + 0.0027978704310953617, + 0.0033415884245187044, + 0.002228677272796631, + 0.0014781986828893423, + 0.003199254162609577, + 0.0034842246677726507, + 0.023024627938866615, + 0.019301529973745346, + 0.017422962933778763, + 0.0325237512588501, + 0.016390221193432808, + 0.016616541892290115, + 0.02277439646422863, + 0.01649554632604122, + 0.003172505646944046, + 0.003068172838538885, + 0.0029664868488907814, + 0.006301370915025473, + 0.0016808158252388239, + 0.0024057121481746435, + 0.0047669862397015095, + 0.004606996197253466, + 0.01920575089752674, + 0.011599872261285782, + 0.01474797073751688, + 0.01521957479417324, + 0.006963686551898718, + 0.016151150688529015, + 0.016952449455857277, + 0.024381494149565697, + 0.002570725278928876, + 0.002502179006114602, + 0.003826288040727377, + 0.004221092909574509, + 0.002299416810274124, + 0.0030560248997062445, + 0.004561318550258875, + 0.003607371123507619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.02070503681898117 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005978795234113932 + ], + "opposite_logprob_tokens": [ + -5.130978584289551 + ], + "logprob_reward": -0.005978795234113932, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005978795234113932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15818928182125092, + 0.12263133376836777, + 0.07515435665845871, + 0.030630089342594147, + 0.07843799144029617, + 0.037817489355802536, + 0.12013789266347885, + 0.022643422707915306, + 0.3338979482650757, + 0.020460212603211403 + ], + "credit_map": [ + 0.020487820729613304, + 0.014563323929905891, + 0.016835294663906097, + 0.02004987746477127, + 0.012623729184269905, + 0.025187473744153976, + 0.021819163113832474, + 0.026622585952281952, + 0.01887384057044983, + 0.013930391520261765, + 0.01516553945839405, + 0.016228636726737022, + 0.006168829742819071, + 0.011136680841445923, + 0.01763218827545643, + 0.02349521778523922, + 0.010163141414523125, + 0.002896026475355029, + 0.004376628901809454, + 0.0073119825683534145, + 0.0035286699421703815, + 0.004082777537405491, + 0.015698572620749474, + 0.027096552774310112, + 0.0032350842375308275, + 0.0021708060521632433, + 0.0032795416191220284, + 0.004478587303310633, + 0.0035892692394554615, + 0.001994851278141141, + 0.00484948605298996, + 0.007032460533082485, + 0.008582470938563347, + 0.006588795688003302, + 0.009991620667278767, + 0.011960259638726711, + 0.007364588789641857, + 0.009660185314714909, + 0.01195477694272995, + 0.01233528833836317, + 0.004528870806097984, + 0.003640122711658478, + 0.00468652555719018, + 0.007052835077047348, + 0.0029178073164075613, + 0.0025549321435391903, + 0.00552001129835844, + 0.006916381418704987, + 0.01352173462510109, + 0.012439410202205181, + 0.013397214002907276, + 0.026734499260783195, + 0.01085917092859745, + 0.01176366489380598, + 0.017498521134257317, + 0.013923672959208488, + 0.0026161314453929663, + 0.0024762989487499, + 0.0030332093592733145, + 0.0033630630932748318, + 0.0033499584533274174, + 0.0016319885617122054, + 0.003196875099092722, + 0.002975893672555685, + 0.05187027528882027, + 0.0315101258456707, + 0.042084868997335434, + 0.044650256633758545, + 0.018174730241298676, + 0.04069855436682701, + 0.0440056212246418, + 0.06090348958969116, + 0.0022487561218440533, + 0.0024737310595810413, + 0.002497641835361719, + 0.0034914088901132345, + 0.0016362458700314164, + 0.0020416059996932745, + 0.0036225083749741316, + 0.0024483141023665667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005978795234113932 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13210654258728027, + 0.1230744868516922, + 0.045766912400722504, + 0.03022720105946064, + 0.1531020551919937, + 0.0383152961730957, + 0.07949123531579971, + 0.05530541017651558, + 0.3071862757205963, + 0.0354246087372303 + ], + "credit_map": [ + 0.01779741421341896, + 0.009308073669672012, + 0.014112802222371101, + 0.018654413521289825, + 0.00887614767998457, + 0.022114872932434082, + 0.01819130778312683, + 0.023051505908370018, + 0.019475864246487617, + 0.008291874080896378, + 0.012937644496560097, + 0.019155971705913544, + 0.006241762079298496, + 0.010587114840745926, + 0.018381161615252495, + 0.028003092855215073, + 0.004202600568532944, + 0.00421117153018713, + 0.004528924357146025, + 0.007226203568279743, + 0.008215419948101044, + 0.003011498600244522, + 0.008832323364913464, + 0.005538772325962782, + 0.002984102815389633, + 0.0031680187676101923, + 0.0042470707558095455, + 0.005916731897741556, + 0.0032171853817999363, + 0.0019140626536682248, + 0.004308117087930441, + 0.004471910186111927, + 0.014553852379322052, + 0.008914683014154434, + 0.014169410802423954, + 0.023347385227680206, + 0.02284912019968033, + 0.012892148457467556, + 0.027156904339790344, + 0.02921854332089424, + 0.003533148905262351, + 0.0036914534866809845, + 0.004263793583959341, + 0.006745590828359127, + 0.003745998488739133, + 0.0032910790760070086, + 0.008486405946314335, + 0.00455782376229763, + 0.008390708826482296, + 0.007381673436611891, + 0.008592608384788036, + 0.016532951965928078, + 0.006791225168853998, + 0.0072578671388328075, + 0.014395147562026978, + 0.010149052366614342, + 0.004731651395559311, + 0.005279467906802893, + 0.006004468072205782, + 0.00825553573668003, + 0.008644724264740944, + 0.003508823923766613, + 0.013142958283424377, + 0.005737782455980778, + 0.047615427523851395, + 0.028836851939558983, + 0.03845272213220596, + 0.04076436161994934, + 0.01647789590060711, + 0.03807700425386429, + 0.04041101783514023, + 0.05655099079012871, + 0.0037321397103369236, + 0.004169178660959005, + 0.005337750073522329, + 0.006404956337064505, + 0.0029656407423317432, + 0.004462115000933409, + 0.004666577093303204, + 0.0036862525157630444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "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 camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809535130858421 + ], + "opposite_logprob_tokens": [ + -3.288095235824585 + ], + "logprob_reward": -0.03809535130858421, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809535130858421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11078766733407974, + 0.1478264033794403, + 0.14423011243343353, + 0.053048692643642426, + 0.0898967981338501, + 0.032657966017723083, + 0.07439505308866501, + 0.029385874047875404, + 0.2849385142326355, + 0.03283295780420303 + ], + "credit_map": [ + 0.010827446356415749, + 0.01260377187281847, + 0.0197235569357872, + 0.012250596657395363, + 0.007804900407791138, + 0.011897035874426365, + 0.01755768433213234, + 0.01812266744673252, + 0.02051689475774765, + 0.01990460231900215, + 0.021856503561139107, + 0.014460056088864803, + 0.008872640319168568, + 0.011622568592429161, + 0.023604976013302803, + 0.02698815055191517, + 0.017260385677218437, + 0.00875048153102398, + 0.0076451608911156654, + 0.016172941774129868, + 0.007501969579607248, + 0.00754381762817502, + 0.030340956524014473, + 0.0490143820643425, + 0.005734384059906006, + 0.003933065105229616, + 0.005461408291012049, + 0.01158054731786251, + 0.006091550923883915, + 0.004366410430520773, + 0.008198603987693787, + 0.007682718802243471, + 0.008192897774279118, + 0.0056528872810304165, + 0.010358307510614395, + 0.015674328431487083, + 0.0056963288225233555, + 0.007259604521095753, + 0.015077636577188969, + 0.021984798833727837, + 0.0031577125191688538, + 0.003822962986305356, + 0.0030873005744069815, + 0.004520411603152752, + 0.002901717321947217, + 0.0029595938976854086, + 0.005007029511034489, + 0.007201232481747866, + 0.011688572354614735, + 0.007746376562863588, + 0.006474216002970934, + 0.012457475997507572, + 0.006014138925820589, + 0.008962883614003658, + 0.010594016872346401, + 0.010457364842295647, + 0.003052681917324662, + 0.002537667751312256, + 0.0034810376819223166, + 0.004487752914428711, + 0.002418396295979619, + 0.0022219261154532433, + 0.007351385895162821, + 0.0038350229151546955, + 0.043920211493968964, + 0.026419077068567276, + 0.03611450269818306, + 0.0389878936111927, + 0.015188178047537804, + 0.03363313525915146, + 0.03842146694660187, + 0.05225403979420662, + 0.004073356278240681, + 0.004106961190700531, + 0.0038908261340111494, + 0.003898788010701537, + 0.002806312870234251, + 0.0055693406611680984, + 0.0045182472094893456, + 0.003969124983996153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.03809535130858421 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340389624238014 + ], + "opposite_logprob_tokens": [ + -3.6513404846191406 + ], + "logprob_reward": -0.026340389624238014, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340389624238014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09330257028341293, + 0.042711514979600906, + 0.04626790061593056, + 0.014889667741954327, + 0.060839731246232986, + 0.0277791116386652, + 0.01701771467924118, + 0.021215174347162247, + 0.6612681150436401, + 0.014708437956869602 + ], + "credit_map": [ + 0.01281782891601324, + 0.006123262923210859, + 0.011416982859373093, + 0.011818323284387589, + 0.006465286482125521, + 0.014569967053830624, + 0.012851295061409473, + 0.01723961904644966, + 0.007974383421242237, + 0.002937710378319025, + 0.0035718646831810474, + 0.00467350147664547, + 0.0020764877554029226, + 0.0036656190641224384, + 0.005992468912154436, + 0.011819479055702686, + 0.007325292564928532, + 0.0037803486920893192, + 0.003952377010136843, + 0.004882499109953642, + 0.002338706748560071, + 0.0036752282176166773, + 0.0075774057768285275, + 0.01273604016751051, + 0.0016493380535393953, + 0.0014967445749789476, + 0.0014256250578910112, + 0.002206993056461215, + 0.001585741643793881, + 0.0011708037927746773, + 0.00219928496517241, + 0.0031551362480968237, + 0.007206913083791733, + 0.002986065112054348, + 0.004863442853093147, + 0.006650064140558243, + 0.0037740818224847317, + 0.005125015042722225, + 0.01108772587031126, + 0.019146420061588287, + 0.005275375209748745, + 0.00165399385150522, + 0.0028609945438802242, + 0.004441014491021633, + 0.0014854633482173085, + 0.003932875115424395, + 0.0037361851427704096, + 0.004393208771944046, + 0.0025304220616817474, + 0.001412817626260221, + 0.0017434571636840701, + 0.0023499177768826485, + 0.001987871713936329, + 0.0013860462931916118, + 0.002887383569031954, + 0.002719798358157277, + 0.002124705584719777, + 0.002088654087856412, + 0.002399747259914875, + 0.00275977561250329, + 0.001598290866240859, + 0.0019680140540003777, + 0.004247281700372696, + 0.004028705880045891, + 0.10285386443138123, + 0.05993103235960007, + 0.08175656199455261, + 0.09034161269664764, + 0.034141141921281815, + 0.08084040135145187, + 0.08872682601213455, + 0.12267675250768661, + 0.0015653347363695502, + 0.0016876927111297846, + 0.002460239455103874, + 0.001876351423561573, + 0.001113024540245533, + 0.002667085500434041, + 0.0020938399247825146, + 0.0012448690831661224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.026340389624238014 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle fisheye lens positions the viewer on the deck of a massive wooden sailing ship, ploughing through the surging waves. At the center of the frame, several heroes, clad in ornate, vintage armor, stand firm, their expressions resolute and their gazes piercing, as they gaze out upon the turbulent distance. The leading male hero, his long hair dancing wildly in the wind, draped in dark armor intricately carved with metal patterns and a thick, crimson cape. His companions, some wielding longbows, others clutching battle axes, their clothing and weapons bearing the marks of age. The fisheye lens's distortion exaggerates the curvature of the ship's hull and the towering mainsail, its canvas billowing and rustling in the fierce wind. The damp, dark wooden deck reflects the indirect, cold light of the gloomy sky, while the splashing whitewater waves all around add a sense of dynamism to the image. The color saturation of the entire picture is deliberately reduced, presenting a faded retro nostalgic film texture with obvious grain and slight dark corners, creating an epic, tragic yet hopeful expedition atmosphere.", + "seed": 7541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7541_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": -0.02640651650726795, + "vlm_logprob": -0.02640651650726795, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02640651650726795, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017423424869775772 + ], + "opposite_logprob_tokens": [ + -4.892423629760742 + ], + "logprob_reward": -0.017423424869775772, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.017423424869775772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13304024934768677, + 0.22375787794589996, + 0.10694849491119385, + 0.037885166704654694, + 0.08283321559429169, + 0.03430124744772911, + 0.06323863565921783, + 0.031384170055389404, + 0.2371750771999359, + 0.04943586513400078 + ], + "credit_map": [ + 0.010977386496961117, + 0.020723115652799606, + 0.012358124367892742, + 0.023736003786325455, + 0.013352223671972752, + 0.018039366230368614, + 0.015937328338623047, + 0.01791670359671116, + 0.014324313960969448, + 0.023604881018400192, + 0.017423219978809357, + 0.025982746854424477, + 0.014932382851839066, + 0.020047150552272797, + 0.04610193893313408, + 0.06134125590324402, + 0.013216326013207436, + 0.018127895891666412, + 0.015882808715105057, + 0.015588208101689816, + 0.010668615810573101, + 0.009177839383482933, + 0.010133038274943829, + 0.01415376365184784, + 0.0035507408902049065, + 0.005085375159978867, + 0.003733055666089058, + 0.00716569134965539, + 0.006488041952252388, + 0.00429899338632822, + 0.004477673675864935, + 0.003085593692958355, + 0.015091107226908207, + 0.011773519217967987, + 0.006748990621417761, + 0.010009855031967163, + 0.0052835228852927685, + 0.023332301527261734, + 0.005532403010874987, + 0.005061513744294643, + 0.0030984284821897745, + 0.004180614836513996, + 0.00393768260255456, + 0.005012227687984705, + 0.0071616689674556255, + 0.004483099095523357, + 0.004012142773717642, + 0.0024153837002813816, + 0.01413638424128294, + 0.004881700500845909, + 0.006206962279975414, + 0.007053873967379332, + 0.004222158808261156, + 0.005383464973419905, + 0.004623935557901859, + 0.01673015020787716, + 0.0025255123618990183, + 0.00402455311268568, + 0.004491097759455442, + 0.0054968735203146935, + 0.0030291499570012093, + 0.003473683726042509, + 0.004322309046983719, + 0.004020989406853914, + 0.044842109084129333, + 0.012267407029867172, + 0.026633527129888535, + 0.03818652778863907, + 0.006759848445653915, + 0.011963537894189358, + 0.02667011320590973, + 0.06985200941562653, + 0.004241790622472763, + 0.004244767129421234, + 0.0060460628010332584, + 0.008057010360062122, + 0.003118847729638219, + 0.0041148788295686245, + 0.009828402660787106, + 0.009784108959138393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.017423424869775772 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810447692871094 + ], + "opposite_logprob_tokens": [ + -4.634810447692871 + ], + "logprob_reward": -0.009810447692871094, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810447692871094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18637791275978088, + 0.3370739221572876, + 0.1273483783006668, + 0.052851300686597824, + 0.041220828890800476, + 0.02833491936326027, + 0.042118433862924576, + 0.051338762044906616, + 0.08851926028728485, + 0.044816289097070694 + ], + "credit_map": [ + 0.012564347125589848, + 0.028676224872469902, + 0.0179972555488348, + 0.03680044785141945, + 0.01932309940457344, + 0.01995737850666046, + 0.027104847133159637, + 0.023954302072525024, + 0.024330971762537956, + 0.04022109881043434, + 0.03226792439818382, + 0.04014976695179939, + 0.026975447311997414, + 0.03466835245490074, + 0.0628218948841095, + 0.07563848793506622, + 0.022700194269418716, + 0.018260134384036064, + 0.027952736243605614, + 0.01701849140226841, + 0.008974798955023289, + 0.009405076503753662, + 0.009598152711987495, + 0.013438797555863857, + 0.006914510857313871, + 0.005963868461549282, + 0.004889668431133032, + 0.008504245430231094, + 0.007909542880952358, + 0.007299365010112524, + 0.0066698347218334675, + 0.004700264427810907, + 0.0035900180228054523, + 0.006670419126749039, + 0.005633705761283636, + 0.010074748657643795, + 0.004334436729550362, + 0.0031894012354314327, + 0.004488399717956781, + 0.0032396994065493345, + 0.002224918454885483, + 0.004762359894812107, + 0.0036896441597491503, + 0.004097986500710249, + 0.00508300494402647, + 0.002762451069429517, + 0.0031569330021739006, + 0.0025576232001185417, + 0.00542895682156086, + 0.00605027237907052, + 0.004799094516783953, + 0.005950600374490023, + 0.00438424339517951, + 0.004133747890591621, + 0.005261503625661135, + 0.006110014859586954, + 0.003918214701116085, + 0.006581142544746399, + 0.007092487998306751, + 0.007704957854002714, + 0.0047074262984097, + 0.006277552805840969, + 0.008860441856086254, + 0.006196537986397743, + 0.019014334306120872, + 0.0054492587223649025, + 0.014987836591899395, + 0.011536984704434872, + 0.003981614951044321, + 0.004739389289170504, + 0.010262707248330116, + 0.01854713074862957, + 0.0038316380232572556, + 0.005940950009971857, + 0.005862061399966478, + 0.006126858294010162, + 0.003532019443809986, + 0.004435191862285137, + 0.006445973180234432, + 0.00864159595221281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009810447692871094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201998710632324 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014201998710632324, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201998710632324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19128453731536865, + 0.14785486459732056, + 0.08586079627275467, + 0.050972871482372284, + 0.028895528987050056, + 0.045160748064517975, + 0.11932703107595444, + 0.1484057903289795, + 0.10774993896484375, + 0.07448785752058029 + ], + "credit_map": [ + 0.01716442033648491, + 0.02901751548051834, + 0.016915777698159218, + 0.030952980741858482, + 0.023269595578312874, + 0.02392110414803028, + 0.024515271186828613, + 0.02552788332104683, + 0.010420084930956364, + 0.018657641485333443, + 0.014613674953579903, + 0.020716825500130653, + 0.018557382747530937, + 0.025215398520231247, + 0.019268903881311417, + 0.020404968410730362, + 0.009707306511700153, + 0.009233428165316582, + 0.013662219978868961, + 0.009251442737877369, + 0.007116895169019699, + 0.006669994443655014, + 0.013352912850677967, + 0.016866592690348625, + 0.004459039773792028, + 0.007104957010596991, + 0.005433999001979828, + 0.007545783184468746, + 0.007902942597866058, + 0.008964242413640022, + 0.005171883385628462, + 0.004390026442706585, + 0.006653608288615942, + 0.002420729259029031, + 0.00331593444570899, + 0.004113330040127039, + 0.0015563078923150897, + 0.007341840770095587, + 0.0017410210566595197, + 0.0017527563031762838, + 0.005593772511929274, + 0.006817449815571308, + 0.00578096741810441, + 0.005804029759019613, + 0.005140920169651508, + 0.005742174573242664, + 0.004986597690731287, + 0.005294834263622761, + 0.022248724475502968, + 0.011362268589437008, + 0.013469289056956768, + 0.014771545305848122, + 0.009779402054846287, + 0.009716354310512543, + 0.010646474547684193, + 0.02733296900987625, + 0.014170181006193161, + 0.019497478380799294, + 0.016807449981570244, + 0.02150077186524868, + 0.015661019831895828, + 0.015998180955648422, + 0.024739323183894157, + 0.020031379535794258, + 0.006563133094459772, + 0.008177803829312325, + 0.02769874408841133, + 0.01344218198210001, + 0.010092031210660934, + 0.008491050451993942, + 0.02032020129263401, + 0.012964786030352116, + 0.0071145035326480865, + 0.008568981662392616, + 0.011846560053527355, + 0.011784889735281467, + 0.005418379325419664, + 0.006969850976020098, + 0.012582107447087765, + 0.010202579200267792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.014201998710632324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002059370279312 + ], + "opposite_logprob_tokens": [ + -2.6950206756591797 + ], + "logprob_reward": -0.07002059370279312, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002059370279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16863252222537994, + 0.23226645588874817, + 0.10990311205387115, + 0.04755815863609314, + 0.061298172920942307, + 0.039940062910318375, + 0.08020998537540436, + 0.0814596638083458, + 0.1162543073296547, + 0.06247757747769356 + ], + "credit_map": [ + 0.009289138950407505, + 0.029575327411293983, + 0.01627367176115513, + 0.030233018100261688, + 0.02058902569115162, + 0.017481984570622444, + 0.02356008067727089, + 0.021630266681313515, + 0.018653474748134613, + 0.02791665494441986, + 0.023738041520118713, + 0.03031383641064167, + 0.020803682506084442, + 0.029819875955581665, + 0.0364827960729599, + 0.04453808441758156, + 0.019693834707140923, + 0.011919371783733368, + 0.02451840043067932, + 0.011924687772989273, + 0.009005307219922543, + 0.007088326849043369, + 0.01107889972627163, + 0.014674283564090729, + 0.008220121264457703, + 0.0046361228451132774, + 0.003210834925994277, + 0.0061860643327236176, + 0.0072764381766319275, + 0.008770179934799671, + 0.004402042366564274, + 0.0048563554883003235, + 0.012546133249998093, + 0.005341896787285805, + 0.00475163571536541, + 0.007764187175780535, + 0.0035984544083476067, + 0.019592605531215668, + 0.0036334730684757233, + 0.004069787450134754, + 0.005043063312768936, + 0.0050025824457407, + 0.005183321423828602, + 0.005009727086871862, + 0.00496958801522851, + 0.005629921332001686, + 0.004965971224009991, + 0.004135889001190662, + 0.009397733025252819, + 0.009996434673666954, + 0.00910341739654541, + 0.012500402517616749, + 0.00722078001126647, + 0.0060000792145729065, + 0.009878642857074738, + 0.016112497076392174, + 0.008211991749703884, + 0.011155785992741585, + 0.012608159333467484, + 0.011826906353235245, + 0.008072750642895699, + 0.007767478469759226, + 0.011974301189184189, + 0.009842289611697197, + 0.007139995694160461, + 0.005318449344485998, + 0.03190893679857254, + 0.015829598531126976, + 0.00417705811560154, + 0.007175697013735771, + 0.011421100236475468, + 0.033283475786447525, + 0.004454689100384712, + 0.009758331812918186, + 0.008216865360736847, + 0.007395194377750158, + 0.00737241143360734, + 0.007292632944881916, + 0.0068692793138325214, + 0.01111817266792059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.07002059370279312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057611756026745 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.02057611756026745, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057611756026745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08045122027397156, + 0.06798609346151352, + 0.04580722376704216, + 0.02507614530622959, + 0.0077266632579267025, + 0.03146764636039734, + 0.5318259000778198, + 0.06981022655963898, + 0.09609594941139221, + 0.04375297948718071 + ], + "credit_map": [ + 0.007613123394548893, + 0.00914264190942049, + 0.010704047046601772, + 0.013759219087660313, + 0.0060981460846960545, + 0.012564192526042461, + 0.009482597932219505, + 0.011087249033153057, + 0.006061496213078499, + 0.011743277311325073, + 0.007215497083961964, + 0.009205303154885769, + 0.006401618476957083, + 0.009028801694512367, + 0.010034183971583843, + 0.008295913226902485, + 0.0111817866563797, + 0.003212430514395237, + 0.009301448240876198, + 0.005508502013981342, + 0.0026356077287346125, + 0.0032976213842630386, + 0.0036988842766731977, + 0.006970940623432398, + 0.002661773469299078, + 0.0030597138684242964, + 0.0023391060531139374, + 0.003986642230302095, + 0.0031449843663722277, + 0.004155876114964485, + 0.0026973483618348837, + 0.003030699910596013, + 0.001193773583509028, + 0.0007736746920272708, + 0.0009223117376677692, + 0.0011968179605901241, + 0.0012534023262560368, + 0.0010055513121187687, + 0.0007400779868476093, + 0.0006410536589100957, + 0.004022636450827122, + 0.003203078405931592, + 0.004373106174170971, + 0.003969060257077217, + 0.004076302517205477, + 0.005596679635345936, + 0.004177604336291552, + 0.002049180446192622, + 0.08583889901638031, + 0.05509814992547035, + 0.07060292363166809, + 0.06907391548156738, + 0.04884140193462372, + 0.06376904249191284, + 0.04901285469532013, + 0.08958867192268372, + 0.007125504780560732, + 0.0062840781174600124, + 0.011890987865626812, + 0.010053522884845734, + 0.005366128869354725, + 0.008525856770575047, + 0.011029260233044624, + 0.009534889832139015, + 0.020940998569130898, + 0.004648680333048105, + 0.02940620668232441, + 0.015226173214614391, + 0.00304100732319057, + 0.0038301367312669754, + 0.0075574941001832485, + 0.011445253156125546, + 0.004409815184772015, + 0.003558291122317314, + 0.007076702080667019, + 0.0064581227488815784, + 0.0031228559091687202, + 0.00355761987157166, + 0.008262988179922104, + 0.007306584622710943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.02057611756026745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle fisheye lens positions the viewer on the deck of a massive wooden sailing ship, ploughing through the surging waves. At the center of the frame, several heroes, clad in ornate, vintage armor, stand firm, their expressions resolute and their gazes piercing, as they gaze out upon the turbulent distance. The leading male hero, his long hair dancing wildly in the wind, draped in dark armor intricately carved with metal patterns and a thick, crimson cape. His companions, some wielding longbows, others clutching battle axes, their clothing and weapons bearing the marks of age. The fisheye lens's distortion exaggerates the curvature of the ship's hull and the towering mainsail, its canvas billowing and rustling in the fierce wind. The damp, dark wooden deck reflects the indirect, cold light of the gloomy sky, while the splashing whitewater waves all around add a sense of dynamism to the image. The color saturation of the entire picture is deliberately reduced, presenting a faded retro nostalgic film texture with obvious grain and slight dark corners, creating an epic, tragic yet hopeful expedition atmosphere.", + "seed": 7542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_10/index_4879_seed_7542_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": -0.020081108249723912, + "vlm_logprob": -0.020081108249723912, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.020081108249723912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01719343289732933 + ], + "opposite_logprob_tokens": [ + -5.39219331741333 + ], + "logprob_reward": -0.01719343289732933, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.01719343289732933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19713926315307617, + 0.13246363401412964, + 0.038186997175216675, + 0.043828703463077545, + 0.214479461312294, + 0.09546210616827011, + 0.11357013881206512, + 0.02886497974395752, + 0.06183181330561638, + 0.07417289167642593 + ], + "credit_map": [ + 0.018994955345988274, + 0.03339548036456108, + 0.01857941411435604, + 0.02066461183130741, + 0.024804964661598206, + 0.029762711375951767, + 0.030279474332928658, + 0.020657667890191078, + 0.012232352048158646, + 0.011106915771961212, + 0.014391886070370674, + 0.020897045731544495, + 0.018369603902101517, + 0.012324259616434574, + 0.026359306648373604, + 0.01678227074444294, + 0.004302519839257002, + 0.003156934166327119, + 0.003084286116063595, + 0.005347105674445629, + 0.00431566359475255, + 0.004546414595097303, + 0.009938149712979794, + 0.003495924873277545, + 0.004953657276928425, + 0.003854762762784958, + 0.0035339179448783398, + 0.0062935566529631615, + 0.0034618498757481575, + 0.004466779064387083, + 0.010545599274337292, + 0.006718585733324289, + 0.01392365712672472, + 0.012555726803839207, + 0.006249540951102972, + 0.01596630923449993, + 0.02020132541656494, + 0.012994087301194668, + 0.07416925579309464, + 0.05841956287622452, + 0.008373288437724113, + 0.014528574422001839, + 0.009537107311189175, + 0.012088224291801453, + 0.01575457863509655, + 0.011201150715351105, + 0.01238055620342493, + 0.011598634533584118, + 0.011058354750275612, + 0.013765208423137665, + 0.012819409370422363, + 0.013453871943056583, + 0.018028777092695236, + 0.011520500294864178, + 0.017581455409526825, + 0.015342572703957558, + 0.003377426415681839, + 0.0037923939526081085, + 0.0019704715814441442, + 0.004019832704216242, + 0.0032615221571177244, + 0.003283800557255745, + 0.005373174790292978, + 0.0037863608449697495, + 0.008516174741089344, + 0.010049103759229183, + 0.004073774442076683, + 0.007273979485034943, + 0.008972426876425743, + 0.006727710831910372, + 0.009045789949595928, + 0.007172855082899332, + 0.00684093264862895, + 0.005147592630237341, + 0.003566103521734476, + 0.01433251891285181, + 0.003957912791520357, + 0.007698178291320801, + 0.023025039583444595, + 0.009604621678590775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01719343289732933 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007651190273463726 + ], + "opposite_logprob_tokens": [ + -4.882651329040527 + ], + "logprob_reward": -0.007651190273463726, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007651190273463726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08091160655021667, + 0.11071311682462692, + 0.025227822363376617, + 0.03220602869987488, + 0.354118674993515, + 0.09489105641841888, + 0.11258674412965775, + 0.03395848721265793, + 0.041557230055332184, + 0.11382920295000076 + ], + "credit_map": [ + 0.0069412654265761375, + 0.013479892164468765, + 0.008322365581989288, + 0.009808260947465897, + 0.013669501058757305, + 0.007434242870658636, + 0.01111675426363945, + 0.01013932004570961, + 0.00703496253117919, + 0.0106272017583251, + 0.010374171659350395, + 0.017448771744966507, + 0.011565702967345715, + 0.008178731426596642, + 0.03187594935297966, + 0.013607630506157875, + 0.0025049648247659206, + 0.002469217637553811, + 0.0025092249270528555, + 0.0034996788017451763, + 0.002886776113882661, + 0.002825826406478882, + 0.005438977852463722, + 0.0030931555666029453, + 0.0023889814037829638, + 0.0013694895897060633, + 0.001943595358170569, + 0.0027421601116657257, + 0.0017775362357497215, + 0.002348071662709117, + 0.01019288133829832, + 0.009443311020731926, + 0.024643126875162125, + 0.012672566808760166, + 0.006767164915800095, + 0.031038394197821617, + 0.022565966472029686, + 0.019230350852012634, + 0.1364079713821411, + 0.10079315304756165, + 0.010219057090580463, + 0.009714777581393719, + 0.009585616178810596, + 0.010620866902172565, + 0.019919585436582565, + 0.010613709688186646, + 0.012076137587428093, + 0.012141309678554535, + 0.010323832742869854, + 0.011901757679879665, + 0.013666415587067604, + 0.018294991925358772, + 0.011301945894956589, + 0.009521263651549816, + 0.020756307989358902, + 0.01682022586464882, + 0.002660611644387245, + 0.0037122145295143127, + 0.003317546797916293, + 0.00487467972561717, + 0.003500988008454442, + 0.004723397083580494, + 0.004721648525446653, + 0.0064474004320800304, + 0.003668681252747774, + 0.004680837504565716, + 0.0020549495238810778, + 0.005394488573074341, + 0.007039565593004227, + 0.0048816148191690445, + 0.008534877561032772, + 0.005302217788994312, + 0.01324048638343811, + 0.010999570600688457, + 0.005882502067834139, + 0.017661411315202713, + 0.007977666333317757, + 0.013264414854347706, + 0.024537254124879837, + 0.02026590146124363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.007651190273463726 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07844033092260361, + 0.07978895306587219, + 0.017304740846157074, + 0.03189907968044281, + 0.5815423130989075, + 0.02553998865187168, + 0.06649378687143326, + 0.029100436717271805, + 0.05277691408991814, + 0.03711343929171562 + ], + "credit_map": [ + 0.007075241766870022, + 0.013300812803208828, + 0.008295919746160507, + 0.010023674927651882, + 0.010316265746951103, + 0.0073845284059643745, + 0.012069341726601124, + 0.009974556043744087, + 0.006102008745074272, + 0.006065269000828266, + 0.006884933449327946, + 0.012582361698150635, + 0.007155842613428831, + 0.007440435234457254, + 0.020913537591695786, + 0.012644569389522076, + 0.0015712266322225332, + 0.0017418242059648037, + 0.00184821174480021, + 0.0023597783874720335, + 0.0022778327111154795, + 0.0028208093717694283, + 0.002718193456530571, + 0.0019668666645884514, + 0.0022948619443923235, + 0.00168328988365829, + 0.002061370061710477, + 0.002630129922181368, + 0.00227521569468081, + 0.0021739101503044367, + 0.009797937236726284, + 0.008982369676232338, + 0.03914092853665352, + 0.021715756505727768, + 0.012584752403199673, + 0.05063707008957863, + 0.03977898508310318, + 0.029994552955031395, + 0.22307127714157104, + 0.16461901366710663, + 0.0029405110981315374, + 0.0022237428929656744, + 0.0027740050572901964, + 0.003590683452785015, + 0.0046487427316606045, + 0.002822439419105649, + 0.0032334697898477316, + 0.0033063990995287895, + 0.006040158681571484, + 0.00725132878869772, + 0.006301463581621647, + 0.01153282355517149, + 0.006540082860738039, + 0.005486977286636829, + 0.012711971998214722, + 0.010628994554281235, + 0.002546648494899273, + 0.0034170427825301886, + 0.0031159142963588238, + 0.00445755897089839, + 0.0031422749161720276, + 0.0033371886238455772, + 0.004669410642236471, + 0.0044144014827907085, + 0.006337186321616173, + 0.0052241976372897625, + 0.0036669382825493813, + 0.006915833335369825, + 0.004771713633090258, + 0.004071282688528299, + 0.01398628018796444, + 0.00780349038541317, + 0.003268097760155797, + 0.005429134238511324, + 0.0025698505342006683, + 0.0050397044979035854, + 0.0028434956911951303, + 0.003535973373800516, + 0.007986260578036308, + 0.006440926808863878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.014194829389452934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808639943599701 + ], + "opposite_logprob_tokens": [ + -3.288086414337158 + ], + "logprob_reward": -0.03808639943599701, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808639943599701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11056029796600342, + 0.10160853713750839, + 0.01683248206973076, + 0.02044207602739334, + 0.6270919442176819, + 0.04189136624336243, + 0.030209720134735107, + 0.014010450802743435, + 0.01632573828101158, + 0.02102743089199066 + ], + "credit_map": [ + 0.011065829545259476, + 0.018000472337007523, + 0.008775036782026291, + 0.01374982763081789, + 0.01782129518687725, + 0.01590476930141449, + 0.013095532543957233, + 0.012147531844675541, + 0.007123331073671579, + 0.007074343506246805, + 0.00936751440167427, + 0.01481657661497593, + 0.009529621340334415, + 0.007277802564203739, + 0.031093522906303406, + 0.015325816348195076, + 0.002433981280773878, + 0.0015419645933434367, + 0.0016188430599868298, + 0.0020790814887732267, + 0.0014602186856791377, + 0.0023338287137448788, + 0.004105375148355961, + 0.0012591899139806628, + 0.0019477609312161803, + 0.001289163250476122, + 0.0013493887381628156, + 0.001829861430451274, + 0.0015300811501219869, + 0.0021567095536738634, + 0.004741423763334751, + 0.005597688723355532, + 0.04273626208305359, + 0.020238807424902916, + 0.011694733053445816, + 0.05512901395559311, + 0.03542637079954147, + 0.03250883147120476, + 0.24626041948795319, + 0.1830974519252777, + 0.004046425223350525, + 0.004086869768798351, + 0.0042501175776124, + 0.006332830060273409, + 0.007187036331743002, + 0.004745986312627792, + 0.005848392378538847, + 0.005393709987401962, + 0.002939059864729643, + 0.0035933784674853086, + 0.0030011660419404507, + 0.004542078822851181, + 0.0035497008357197046, + 0.0028581845108419657, + 0.005193671677261591, + 0.00453247781842947, + 0.0011636283015832305, + 0.0015573793789371848, + 0.0010406112996861339, + 0.0022253706119954586, + 0.0013156906934455037, + 0.001380827510729432, + 0.00277682370506227, + 0.002550119534134865, + 0.002070285379886627, + 0.0015351349720731378, + 0.001027285703457892, + 0.0018851406639441848, + 0.0024660215713083744, + 0.00199426943436265, + 0.0033728801645338535, + 0.0019747207406908274, + 0.0030591061804443598, + 0.0015331009635701776, + 0.0009834262309595942, + 0.0029187893960624933, + 0.0016827277140691876, + 0.0025692670606076717, + 0.00508180120959878, + 0.003199211321771145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.03808639943599701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle with a fisheye lens distortion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are several heroes wearing ornate vintage armor visible on the deck of a massive wooden sailing ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading male hero depicted with long hair, dark armor with metal patterns, and a thick crimson cape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are companions visible wielding longbows or battle axes on the ship?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the image exhibit a faded retro nostalgic film texture with obvious grain and slight dark corners?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279689252376556 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279689252376556, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279689252376556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08112625777721405, + 0.19744762778282166, + 0.030817890539765358, + 0.030812252312898636, + 0.38247939944267273, + 0.04740093648433685, + 0.03790172189474106, + 0.03984599560499191, + 0.018928907811641693, + 0.13323906064033508 + ], + "credit_map": [ + 0.006772354245185852, + 0.013354604132473469, + 0.00790509581565857, + 0.013542823493480682, + 0.011231992393732071, + 0.007727136369794607, + 0.009873482398688793, + 0.01071876846253872, + 0.02017500437796116, + 0.013866007328033447, + 0.014105873182415962, + 0.02216845564544201, + 0.01644601859152317, + 0.02075439877808094, + 0.050291601568460464, + 0.03964027762413025, + 0.004151578526943922, + 0.0023585751187056303, + 0.0022526823449879885, + 0.003640874056145549, + 0.002459821756929159, + 0.0029461290687322617, + 0.010292120277881622, + 0.0027161103207618, + 0.0031848365906625986, + 0.0026960873510688543, + 0.004050114192068577, + 0.004826141055673361, + 0.0025101916398853064, + 0.0022953085135668516, + 0.006549161393195391, + 0.004700410179793835, + 0.02584548108279705, + 0.011752975173294544, + 0.007674547377973795, + 0.033616118133068085, + 0.021247414872050285, + 0.01927875354886055, + 0.15049999952316284, + 0.1125640869140625, + 0.005069300532341003, + 0.008360474370419979, + 0.004905331414192915, + 0.008025681599974632, + 0.0056501831859350204, + 0.005759598221629858, + 0.005495268851518631, + 0.004135096445679665, + 0.003759159240871668, + 0.004614738281816244, + 0.004872462712228298, + 0.004679148551076651, + 0.004405209328979254, + 0.0034003285691142082, + 0.0062032658606767654, + 0.00596740934997797, + 0.004122117534279823, + 0.005229093600064516, + 0.0034541741479188204, + 0.005097022280097008, + 0.005099577363580465, + 0.0034263161942362785, + 0.0072111268527805805, + 0.0062065706588327885, + 0.0022478047758340836, + 0.0018761834362521768, + 0.0011133707594126463, + 0.0025117516051977873, + 0.002217991277575493, + 0.0021130815148353577, + 0.004099090117961168, + 0.0027496344409883022, + 0.01259972807019949, + 0.008236399851739407, + 0.004373342264443636, + 0.02181093767285347, + 0.008008135482668877, + 0.012850718572735786, + 0.047783609479665756, + 0.0175761915743351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.023279689252376556 + } + ], + "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/0000152/group_11/index_tensor([1756, 1756])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_tensor([1756, 1756])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7c15321a0630cfd3a9a8e0ce03a0c55c602990f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_tensor([1756, 1756])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static medium shot captures a rain-slicked, narrow alleyway in a sprawling cyberpunk cityscape, rendered in stunningly photorealistic CGI. The scene is drenched in a pulsating glow of magenta and cyan neon emanating from an unseen source above, casting long, deep shadows across the grimy, graffiti-covered concrete walls. At the center of the frame, a four-legged robot dog crouches on the ground, its chassis constructed from interlocking, scuffed panels of gunmetal gray and matte black. Its optical sensors emit a harsh electric blue, casting a faint, partial glow on its snout as it steadily and deliberately gnaws at a piece of abandoned high-tech machinery—a cracked server casing, revealing gleaming circuit boards. Its jagged metal jaws grind with mechanical force, and with each powerful bite, its head twitches, sending a trail of tiny sparks into the humid air. Suddenly, the massive, fast-moving shadow of a hovercraft passing overhead sweeps across the scene from left to right, momentarily plunging the alley into near-total darkness before vibrant neon floods back in, its dynamic shifts highlighting the intricate details of the robot’s articulated limbs and the textured sheen of its metal frame against the wet, reflective ground.", + "seed": 15316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_1756_seed_15316_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": -0.3001602296717465, + "vlm_logprob": -0.3001602296717465, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.3001602296717465, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.526314377784729 + ], + "opposite_logprob_tokens": [ + -0.901314377784729 + ], + "logprob_reward": -0.526314377784729, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.526314377784729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2497255504131317, + 0.09225249290466309, + 0.14100229740142822, + 0.04484495148062706, + 0.09073353558778763, + 0.053755197674036026, + 0.03163022920489311, + 0.12113294005393982, + 0.13784678280353546, + 0.03707604482769966 + ], + "credit_map": [ + 0.03520721569657326, + 0.02257663942873478, + 0.033315591514110565, + 0.025653565302491188, + 0.022144602611660957, + 0.025118937715888023, + 0.02939661592245102, + 0.056312382221221924, + 0.01513061486184597, + 0.01161291915923357, + 0.013643805868923664, + 0.010904665105044842, + 0.007433885242789984, + 0.010928490199148655, + 0.0066701434552669525, + 0.015927966684103012, + 0.021113121882081032, + 0.01804996281862259, + 0.019699091091752052, + 0.015173189342021942, + 0.009699826128780842, + 0.017564769834280014, + 0.010762770660221577, + 0.028939561918377876, + 0.004676055628806353, + 0.00412798672914505, + 0.005416989326477051, + 0.006667593494057655, + 0.004763042088598013, + 0.005146655719727278, + 0.006044276524335146, + 0.008002350106835365, + 0.009668229147791862, + 0.008400863967835903, + 0.011800117790699005, + 0.014902173541486263, + 0.01129860244691372, + 0.011695243418216705, + 0.008590707555413246, + 0.014377604238688946, + 0.006063756532967091, + 0.004431765526533127, + 0.008421086706221104, + 0.008583943359553814, + 0.009012499824166298, + 0.0060435407795012, + 0.004637989215552807, + 0.006560617126524448, + 0.0033735879696905613, + 0.004092426039278507, + 0.004516297020018101, + 0.005999842658638954, + 0.0031529965344816446, + 0.003327332902699709, + 0.002527435775846243, + 0.004640310537070036, + 0.01581927202641964, + 0.009116735309362411, + 0.018588624894618988, + 0.015757035464048386, + 0.015889093279838562, + 0.02035466581583023, + 0.010480852797627449, + 0.015126665122807026, + 0.029689105227589607, + 0.019897745922207832, + 0.018539350479841232, + 0.016050608828663826, + 0.015504853799939156, + 0.012706510722637177, + 0.011005719192326069, + 0.014452889561653137, + 0.005247735884040594, + 0.004005898255854845, + 0.006805200595408678, + 0.004496261477470398, + 0.0044698151759803295, + 0.0024438316468149424, + 0.003679745364934206, + 0.0059275575913488865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.526314377784729 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306991025805473 + ], + "opposite_logprob_tokens": [ + -3.168069839477539 + ], + "logprob_reward": -0.04306991025805473, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306991025805473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17723658680915833, + 0.08973310142755508, + 0.11999412626028061, + 0.2585645914077759, + 0.03527829796075821, + 0.02193198725581169, + 0.037122368812561035, + 0.032412271946668625, + 0.16432850062847137, + 0.06339819729328156 + ], + "credit_map": [ + 0.01980002596974373, + 0.01514190062880516, + 0.021827183663845062, + 0.024028586223721504, + 0.01990322954952717, + 0.01770414598286152, + 0.02028394117951393, + 0.0385475791990757, + 0.009472891688346863, + 0.007280837744474411, + 0.013177460059523582, + 0.011268840171396732, + 0.010591395199298859, + 0.009952642023563385, + 0.010612856596708298, + 0.01737617887556553, + 0.011532648466527462, + 0.0086034145206213, + 0.017907746136188507, + 0.020969612523913383, + 0.011320821940898895, + 0.008900916203856468, + 0.013058135285973549, + 0.02770082652568817, + 0.021469319239258766, + 0.017114294692873955, + 0.0318170003592968, + 0.043287765234708786, + 0.03678550198674202, + 0.020066499710083008, + 0.039584510028362274, + 0.048439688980579376, + 0.005176664795726538, + 0.0028356306720525026, + 0.005070874467492104, + 0.005693531129509211, + 0.0021672819275408983, + 0.003988244570791721, + 0.0043088700622320175, + 0.006037198938429356, + 0.0025254699867218733, + 0.00210966425947845, + 0.00323257758282125, + 0.0031383216846734285, + 0.0028110716957598925, + 0.0032118158414959908, + 0.002319921972230077, + 0.0025831449311226606, + 0.004878176376223564, + 0.0030075747054070234, + 0.005411350633949041, + 0.0049753328785300255, + 0.005988402292132378, + 0.004581762477755547, + 0.0027948517818003893, + 0.005484916269779205, + 0.0039954897947609425, + 0.004201657138764858, + 0.005117163527756929, + 0.004476103466004133, + 0.005595873575657606, + 0.002730009378865361, + 0.003155733225867152, + 0.003140241140499711, + 0.017711294814944267, + 0.017684316262602806, + 0.02338465489447117, + 0.023847227916121483, + 0.033942144364118576, + 0.014649491757154465, + 0.018813690170645714, + 0.014295668341219425, + 0.007974375039339066, + 0.010045838542282581, + 0.010874640196561813, + 0.009355519898235798, + 0.010478967800736427, + 0.003131855046376586, + 0.005794825963675976, + 0.005742175038903952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.04306991025805473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25421324372291565, + 0.17431308329105377, + 0.09547353535890579, + 0.06704545766115189, + 0.06146097555756569, + 0.05763420835137367, + 0.028710780665278435, + 0.06520411372184753, + 0.07489980757236481, + 0.1210448294878006 + ], + "credit_map": [ + 0.03796432167291641, + 0.033393070101737976, + 0.03422393277287483, + 0.03290674090385437, + 0.033101119101047516, + 0.026801809668540955, + 0.02492302842438221, + 0.030899187549948692, + 0.021001433953642845, + 0.016773324459791183, + 0.02483813464641571, + 0.02609911747276783, + 0.012295895256102085, + 0.01853499561548233, + 0.02291726879775524, + 0.031852904707193375, + 0.011943821795284748, + 0.006760014686733484, + 0.014423160813748837, + 0.014374852180480957, + 0.009831767529249191, + 0.007681036833673716, + 0.015640608966350555, + 0.014818262308835983, + 0.008962527848780155, + 0.005777291487902403, + 0.009126165881752968, + 0.00940773170441389, + 0.005078669171780348, + 0.0077035631984472275, + 0.005417535547167063, + 0.015571963973343372, + 0.006463939324021339, + 0.005494957789778709, + 0.008748319931328297, + 0.009386848658323288, + 0.006970988120883703, + 0.006560987792909145, + 0.006882769521325827, + 0.010952156037092209, + 0.006175589747726917, + 0.007761063985526562, + 0.008565129712224007, + 0.008618813008069992, + 0.0071558053605258465, + 0.0063627054914832115, + 0.004385252948850393, + 0.008609842509031296, + 0.003907578997313976, + 0.0024611991830170155, + 0.003945636563003063, + 0.004138079006224871, + 0.003143882378935814, + 0.0025360824074596167, + 0.0024012417998164892, + 0.006177076138556004, + 0.01128710713237524, + 0.009691394865512848, + 0.00966937281191349, + 0.008669748902320862, + 0.006819895002990961, + 0.006125194951891899, + 0.004866134840995073, + 0.008075252175331116, + 0.012901827692985535, + 0.00994174275547266, + 0.009279950521886349, + 0.009417031891644001, + 0.012658513151109219, + 0.005366369616240263, + 0.0062973108142614365, + 0.009037045761942863, + 0.019400013610720634, + 0.02020827680826187, + 0.019253378733992577, + 0.013427108526229858, + 0.020190779119729996, + 0.007388472557067871, + 0.009508240036666393, + 0.011668547987937927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738083902746439 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738083902746439, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738083902746439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23086534440517426, + 0.0913965031504631, + 0.07630719244480133, + 0.08085409551858902, + 0.06686418503522873, + 0.06960608810186386, + 0.053285084664821625, + 0.13175562024116516, + 0.13807353377342224, + 0.060992393642663956 + ], + "credit_map": [ + 0.027768997475504875, + 0.033680111169815063, + 0.023943066596984863, + 0.02478647790849209, + 0.04133906587958336, + 0.02378722093999386, + 0.026933519169688225, + 0.02862686477601528, + 0.013244345784187317, + 0.010385176166892052, + 0.012150991708040237, + 0.011631843633949757, + 0.007222477346658707, + 0.012627624906599522, + 0.010109199211001396, + 0.014024843461811543, + 0.0118631012737751, + 0.009703473187983036, + 0.010192124173045158, + 0.009182670153677464, + 0.00646368321031332, + 0.010214860551059246, + 0.0056060864590108395, + 0.013081194832921028, + 0.010435721836984158, + 0.010554309003055096, + 0.011204065755009651, + 0.01027682889252901, + 0.005987200420349836, + 0.012063538655638695, + 0.007084255572408438, + 0.013248180039227009, + 0.007890795357525349, + 0.005828308407217264, + 0.009479548782110214, + 0.010227322578430176, + 0.008174889720976353, + 0.00902389083057642, + 0.004733234643936157, + 0.01150619424879551, + 0.009901284240186214, + 0.009333179332315922, + 0.009842763654887676, + 0.009317637421190739, + 0.008371290750801563, + 0.0067687248811125755, + 0.005241143051534891, + 0.010830067098140717, + 0.00627939635887742, + 0.006986528635025024, + 0.006943980697542429, + 0.010990769602358341, + 0.006231615319848061, + 0.0051034362986683846, + 0.004127591382712126, + 0.006621762178838253, + 0.01798623614013195, + 0.012994424439966679, + 0.020881151780486107, + 0.0185084231197834, + 0.018062379211187363, + 0.017452048137784004, + 0.011211410164833069, + 0.014659545384347439, + 0.03367506340146065, + 0.022136140614748, + 0.018296167254447937, + 0.015815481543540955, + 0.01403514388948679, + 0.011874411255121231, + 0.009167453274130821, + 0.013073676265776157, + 0.008429723791778088, + 0.00967604573816061, + 0.009731155820190907, + 0.007308854255825281, + 0.010620749555528164, + 0.004004387184977531, + 0.004873866215348244, + 0.006347615271806717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.006738083902746439 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983297944068909 + ], + "opposite_logprob_tokens": [ + -0.5233297944068909 + ], + "logprob_reward": -0.8983297944068909, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983297944068909, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1429491490125656, + 0.07325224578380585, + 0.10381565243005753, + 0.06114901602268219, + 0.10412281006574631, + 0.04528484120965004, + 0.08857112377882004, + 0.07025022059679031, + 0.1603514403104782, + 0.1502535492181778 + ], + "credit_map": [ + 0.014879856258630753, + 0.012310253456234932, + 0.017857566475868225, + 0.02047230862081051, + 0.01998402737081051, + 0.014165688306093216, + 0.01653260737657547, + 0.02674684301018715, + 0.010065770708024502, + 0.007346647325903177, + 0.00950187910348177, + 0.009537503123283386, + 0.010502081364393234, + 0.006694725714623928, + 0.007619786076247692, + 0.011983852833509445, + 0.011951733380556107, + 0.009358871728181839, + 0.013711852952837944, + 0.013929687440395355, + 0.01339793298393488, + 0.010617499239742756, + 0.00960241537541151, + 0.021245652809739113, + 0.0068155755288898945, + 0.005580838304013014, + 0.009070279076695442, + 0.009561988525092602, + 0.006449705455452204, + 0.006545974873006344, + 0.006978877820074558, + 0.010145776905119419, + 0.011289466172456741, + 0.009215679951012135, + 0.015383448451757431, + 0.016102515161037445, + 0.008883083239197731, + 0.012110650539398193, + 0.011732477694749832, + 0.019405486062169075, + 0.0062117562629282475, + 0.003307806793600321, + 0.007403441239148378, + 0.0057798465713858604, + 0.007852815091609955, + 0.003390805097296834, + 0.004970134235918522, + 0.0063682361505925655, + 0.009844128042459488, + 0.008674719370901585, + 0.011415264569222927, + 0.012399493716657162, + 0.013690845109522343, + 0.01004887092858553, + 0.006786935031414032, + 0.015710871666669846, + 0.010895754210650921, + 0.0057169911451637745, + 0.012765704654157162, + 0.011485126800835133, + 0.007013860624283552, + 0.010089360177516937, + 0.004998812451958656, + 0.007284615188837051, + 0.021791482344269753, + 0.01920256018638611, + 0.022954806685447693, + 0.028590474277734756, + 0.023932863026857376, + 0.01222850289195776, + 0.014685496687889099, + 0.016965243965387344, + 0.018443278968334198, + 0.025099432095885277, + 0.022787636145949364, + 0.02126576378941536, + 0.02365640178322792, + 0.012190314009785652, + 0.012454097159206867, + 0.014356625266373158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.8983297944068909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, static medium shot captures a rain-slicked, narrow alleyway in a sprawling cyberpunk cityscape, rendered in stunningly photorealistic CGI. The scene is drenched in a pulsating glow of magenta and cyan neon emanating from an unseen source above, casting long, deep shadows across the grimy, graffiti-covered concrete walls. At the center of the frame, a four-legged robot dog crouches on the ground, its chassis constructed from interlocking, scuffed panels of gunmetal gray and matte black. Its optical sensors emit a harsh electric blue, casting a faint, partial glow on its snout as it steadily and deliberately gnaws at a piece of abandoned high-tech machinery—a cracked server casing, revealing gleaming circuit boards. Its jagged metal jaws grind with mechanical force, and with each powerful bite, its head twitches, sending a trail of tiny sparks into the humid air. Suddenly, the massive, fast-moving shadow of a hovercraft passing overhead sweeps across the scene from left to right, momentarily plunging the alley into near-total darkness before vibrant neon floods back in, its dynamic shifts highlighting the intricate details of the robot’s articulated limbs and the textured sheen of its metal frame against the wet, reflective ground.", + "seed": 15317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_1756_seed_15317_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": -0.25648828288540243, + "vlm_logprob": -0.25648828288540243, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.25648828288540243, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3178260028362274 + ], + "opposite_logprob_tokens": [ + -1.3178260326385498 + ], + "logprob_reward": -0.3178260028362274, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3178260028362274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42216330766677856, + 0.0989556685090065, + 0.09998129308223724, + 0.04250478744506836, + 0.07463377714157104, + 0.019271887838840485, + 0.037513747811317444, + 0.031119773164391518, + 0.06271055340766907, + 0.11114514619112015 + ], + "credit_map": [ + 0.03270978108048439, + 0.03121613897383213, + 0.06415285170078278, + 0.0371258482336998, + 0.054157160222530365, + 0.07930797338485718, + 0.05151935666799545, + 0.07197419553995132, + 0.006475246511399746, + 0.0061273653991520405, + 0.02364717423915863, + 0.007332213688641787, + 0.007612967398017645, + 0.012863394804298878, + 0.008456004783511162, + 0.026441296562552452, + 0.016177019104361534, + 0.006373388227075338, + 0.014570701867341995, + 0.007544584572315216, + 0.006682677194476128, + 0.022194549441337585, + 0.01029460970312357, + 0.016143765300512314, + 0.003918189089745283, + 0.003267816500738263, + 0.007706673350185156, + 0.0026918193325400352, + 0.005739700049161911, + 0.004304676782339811, + 0.006392048671841621, + 0.00848386250436306, + 0.009055825881659985, + 0.007792883086949587, + 0.013234858401119709, + 0.005611496511846781, + 0.00655378308147192, + 0.015087730251252651, + 0.007033313158899546, + 0.010263888165354729, + 0.0028942201752215624, + 0.0016058547189459205, + 0.0022572081070393324, + 0.001541648874990642, + 0.0024331077001988888, + 0.004441278986632824, + 0.0016513761365786195, + 0.002447193255648017, + 0.0036791344173252583, + 0.002495341934263706, + 0.009243487380445004, + 0.0020193629898130894, + 0.003118969965726137, + 0.004032195080071688, + 0.002563798101618886, + 0.010361455380916595, + 0.005593872629106045, + 0.0030592430848628283, + 0.004962925333529711, + 0.0024137741420418024, + 0.0036403837148100138, + 0.004019290674477816, + 0.0028784554451704025, + 0.0045518288388848305, + 0.0102577805519104, + 0.0048300535418093204, + 0.012603272683918476, + 0.0062120468355715275, + 0.005074128974229097, + 0.010900186374783516, + 0.004519420675933361, + 0.008313662372529507, + 0.010891473852097988, + 0.008587567135691643, + 0.026180652901530266, + 0.008410338312387466, + 0.006116093602031469, + 0.02344399131834507, + 0.00881592370569706, + 0.018699100241065025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.3178260028362274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809695690870285 + ], + "opposite_logprob_tokens": [ + -3.2880969047546387 + ], + "logprob_reward": -0.03809695690870285, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809695690870285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15716996788978577, + 0.24476391077041626, + 0.07023120671510696, + 0.1460304707288742, + 0.05158413574099541, + 0.0932493507862091, + 0.033491894602775574, + 0.053545914590358734, + 0.06139488145709038, + 0.088538259267807 + ], + "credit_map": [ + 0.010888058692216873, + 0.0166543610394001, + 0.018748503178358078, + 0.011486867442727089, + 0.03215969726443291, + 0.026505684480071068, + 0.021184584125876427, + 0.01954222097992897, + 0.02258128672838211, + 0.021369822323322296, + 0.03010091930627823, + 0.028814807534217834, + 0.03017544001340866, + 0.022747758775949478, + 0.05020783096551895, + 0.038766048848629, + 0.008152053691446781, + 0.006813880521804094, + 0.012643098831176758, + 0.009832888841629028, + 0.005780387669801712, + 0.007324106525629759, + 0.009388994425535202, + 0.010295793414115906, + 0.008044463582336903, + 0.007197953760623932, + 0.047131385654211044, + 0.00704951211810112, + 0.00779553409665823, + 0.007326195016503334, + 0.010342790745198727, + 0.05114263668656349, + 0.005910580512136221, + 0.006472917273640633, + 0.007413478102535009, + 0.005370090249925852, + 0.0065706209279596806, + 0.004804987460374832, + 0.008316106162965298, + 0.00672535365447402, + 0.009525747038424015, + 0.007455067243427038, + 0.013906074687838554, + 0.008958769962191582, + 0.009389162063598633, + 0.015415544621646404, + 0.016781646758317947, + 0.011817337945103645, + 0.003692112397402525, + 0.003585695056244731, + 0.005611506290733814, + 0.004348220769315958, + 0.004334222059696913, + 0.002784202340990305, + 0.004439999349415302, + 0.004695936571806669, + 0.005925796460360289, + 0.0056961458176374435, + 0.0074266549199819565, + 0.005603143945336342, + 0.007429249119013548, + 0.0064446269534528255, + 0.008165130391716957, + 0.006855170242488384, + 0.006276790983974934, + 0.008393943309783936, + 0.008858286775648594, + 0.008154105395078659, + 0.008879334665834904, + 0.007307702209800482, + 0.005945101380348206, + 0.007579614408314228, + 0.008623226545751095, + 0.011350599117577076, + 0.015124641358852386, + 0.009464374743402004, + 0.007963989861309528, + 0.014444335363805294, + 0.006746220402419567, + 0.014820869080722332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.03809695690870285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577985793352127 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577985793352127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577985793352127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14070259034633636, + 0.061558641493320465, + 0.026561185717582703, + 0.0593726746737957, + 0.08356077969074249, + 0.04793524742126465, + 0.26786190271377563, + 0.1912752240896225, + 0.043936312198638916, + 0.0772353857755661 + ], + "credit_map": [ + 0.010070856660604477, + 0.016102289780974388, + 0.02163495495915413, + 0.01773756369948387, + 0.014203972183167934, + 0.016415469348430634, + 0.017237434163689613, + 0.02730005420744419, + 0.005931316874921322, + 0.0061629824340343475, + 0.009491209872066975, + 0.006428132764995098, + 0.00787417870014906, + 0.0074602290987968445, + 0.008624645881354809, + 0.009585949592292309, + 0.0026535294018685818, + 0.0032481912057846785, + 0.003403155133128166, + 0.0033460059203207493, + 0.0026063371915370226, + 0.00397828035056591, + 0.003635055385529995, + 0.0036906320601701736, + 0.004651058930903673, + 0.0023975830990821123, + 0.02040022425353527, + 0.002762838965281844, + 0.0027198782190680504, + 0.003262000158429146, + 0.0028199744410812855, + 0.020359119400382042, + 0.008463690057396889, + 0.009951253421604633, + 0.010328922420740128, + 0.008585033938288689, + 0.012501589022576809, + 0.009925800375640392, + 0.012763813138008118, + 0.01104068011045456, + 0.006748838350176811, + 0.004146352875977755, + 0.00590249989181757, + 0.003625574056059122, + 0.006406312808394432, + 0.009512447752058506, + 0.005388667806982994, + 0.006204558536410332, + 0.015940001234412193, + 0.014180405996739864, + 0.09312903881072998, + 0.010623600333929062, + 0.015020223334431648, + 0.01282583549618721, + 0.01015725638717413, + 0.09598559141159058, + 0.013894234783947468, + 0.010355898179113865, + 0.0629805400967598, + 0.01008915901184082, + 0.0147524643689394, + 0.009120061062276363, + 0.009972638450562954, + 0.060110244899988174, + 0.006544893607497215, + 0.003721517976373434, + 0.007961099036037922, + 0.0056712087243795395, + 0.004946476314216852, + 0.005591573193669319, + 0.003566349158063531, + 0.00593319907784462, + 0.008064808323979378, + 0.008161425590515137, + 0.011780030094087124, + 0.006911230739206076, + 0.005889609456062317, + 0.012213019654154778, + 0.00660439720377326, + 0.017610875889658928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.020577985793352127 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2968222200870514, + 0.08090552687644958, + 0.06929904967546463, + 0.15584565699100494, + 0.055849287658929825, + 0.058129679411649704, + 0.03067972883582115, + 0.11554234474897385, + 0.051160603761672974, + 0.08576589077711105 + ], + "credit_map": [ + 0.025720637291669846, + 0.02730252780020237, + 0.03727823495864868, + 0.028191672638058662, + 0.0421687588095665, + 0.05416455864906311, + 0.041285447776317596, + 0.040710385888814926, + 0.006947551388293505, + 0.007231445517390966, + 0.010910801589488983, + 0.007111605256795883, + 0.008962417021393776, + 0.019514571875333786, + 0.009706448763608932, + 0.010520683601498604, + 0.01099997479468584, + 0.005010196007788181, + 0.010792634449899197, + 0.004959879908710718, + 0.005512461997568607, + 0.014248739928007126, + 0.007768430281430483, + 0.010006731376051903, + 0.012405003421008587, + 0.005038297735154629, + 0.05422210320830345, + 0.007114742416888475, + 0.006656231824308634, + 0.008923672139644623, + 0.007568769156932831, + 0.053916823118925095, + 0.006577319465577602, + 0.006445426028221846, + 0.008909059688448906, + 0.0050862617790699005, + 0.005491436924785376, + 0.009960063733160496, + 0.006194382440298796, + 0.007185336202383041, + 0.0066998787224292755, + 0.005318229552358389, + 0.006506324280053377, + 0.004991538356989622, + 0.006360097322613001, + 0.013494783081114292, + 0.007727075833827257, + 0.00703175226226449, + 0.003561431774869561, + 0.003917754162102938, + 0.004017099738121033, + 0.0027435338124632835, + 0.005587775260210037, + 0.0044144256971776485, + 0.0031705303117632866, + 0.0032671792432665825, + 0.00960450153797865, + 0.003000513883307576, + 0.04206220433115959, + 0.0048788171261549, + 0.004276873543858528, + 0.006664099171757698, + 0.003610940184444189, + 0.04144439473748207, + 0.006660119630396366, + 0.0051523661240935326, + 0.009037967771291733, + 0.006639396771788597, + 0.006873574107885361, + 0.006291200872510672, + 0.004888144787400961, + 0.005617831367999315, + 0.007729982491582632, + 0.009335820563137531, + 0.021589186042547226, + 0.007819962687790394, + 0.005806641187518835, + 0.013515830039978027, + 0.00611325865611434, + 0.013855207711458206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.007625873666256666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983145952224731 + ], + "opposite_logprob_tokens": [ + -0.5233145952224731 + ], + "logprob_reward": -0.8983145952224731, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983145952224731, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16200488805770874, + 0.06616176664829254, + 0.01944979466497898, + 0.075562983751297, + 0.027428431436419487, + 0.07474219799041748, + 0.04803328588604927, + 0.14071770012378693, + 0.06877058744430542, + 0.31712833046913147 + ], + "credit_map": [ + 0.012889872305095196, + 0.01833132468163967, + 0.021404800936579704, + 0.019264254719018936, + 0.021890683099627495, + 0.022966060787439346, + 0.019298618659377098, + 0.02595929242670536, + 0.006024625618010759, + 0.006504059303551912, + 0.008009319193661213, + 0.007567073218524456, + 0.009034108370542526, + 0.007960125803947449, + 0.012402467429637909, + 0.008659996092319489, + 0.0023827524855732918, + 0.0017872900934889913, + 0.003310668980702758, + 0.0023491065949201584, + 0.0019190722377970815, + 0.0024049316998571157, + 0.002579374937340617, + 0.002716600429266691, + 0.005979789886623621, + 0.00652021961286664, + 0.01152166910469532, + 0.009113037958741188, + 0.009224601089954376, + 0.0078111267648637295, + 0.017386747524142265, + 0.008005795069038868, + 0.0029310567770153284, + 0.0025775074027478695, + 0.004924648441374302, + 0.002217681612819433, + 0.0025319496635347605, + 0.00413505919277668, + 0.004141815006732941, + 0.003968717064708471, + 0.007996282540261745, + 0.007005166728049517, + 0.008210516534745693, + 0.008826726116240025, + 0.009110048413276672, + 0.010237502865493298, + 0.012702807784080505, + 0.010653155855834484, + 0.0038053561002016068, + 0.002463130746036768, + 0.014929481782019138, + 0.0026298323646187782, + 0.00285010552033782, + 0.003450489602982998, + 0.002426596824079752, + 0.015478298999369144, + 0.01167907565832138, + 0.0037166110705584288, + 0.050615113228559494, + 0.00754137896001339, + 0.004537687636911869, + 0.007673459593206644, + 0.005230724811553955, + 0.04972367361187935, + 0.010606098920106888, + 0.008094422519207, + 0.009401516057550907, + 0.010163692757487297, + 0.008226334117352962, + 0.008454347960650921, + 0.006390685215592384, + 0.007433496881276369, + 0.030622554942965508, + 0.027713248506188393, + 0.058464501053094864, + 0.03080114535987377, + 0.02416428178548813, + 0.07474453002214432, + 0.019716842100024223, + 0.050901249051094055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.8983145952224731 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, static medium shot captures a rain-slicked, narrow alleyway in a sprawling cyberpunk cityscape, rendered in stunningly photorealistic CGI. The scene is drenched in a pulsating glow of magenta and cyan neon emanating from an unseen source above, casting long, deep shadows across the grimy, graffiti-covered concrete walls. At the center of the frame, a four-legged robot dog crouches on the ground, its chassis constructed from interlocking, scuffed panels of gunmetal gray and matte black. Its optical sensors emit a harsh electric blue, casting a faint, partial glow on its snout as it steadily and deliberately gnaws at a piece of abandoned high-tech machinery—a cracked server casing, revealing gleaming circuit boards. Its jagged metal jaws grind with mechanical force, and with each powerful bite, its head twitches, sending a trail of tiny sparks into the humid air. Suddenly, the massive, fast-moving shadow of a hovercraft passing overhead sweeps across the scene from left to right, momentarily plunging the alley into near-total darkness before vibrant neon floods back in, its dynamic shifts highlighting the intricate details of the robot’s articulated limbs and the textured sheen of its metal frame against the wet, reflective ground.", + "seed": 15318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_1756_seed_15318_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": -0.1585242362692952, + "vlm_logprob": -0.1585242362692952, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.1585242362692952, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043290942907333374 + ], + "opposite_logprob_tokens": [ + -3.293290853500366 + ], + "logprob_reward": -0.043290942907333374, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.043290942907333374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2394053339958191, + 0.08336324244737625, + 0.04841775447130203, + 0.08306771516799927, + 0.03550584241747856, + 0.056385599076747894, + 0.07408463209867477, + 0.033920273184776306, + 0.14997614920139313, + 0.1958734542131424 + ], + "credit_map": [ + 0.03825527802109718, + 0.024973096325993538, + 0.028664348646998405, + 0.02132558822631836, + 0.025087056681513786, + 0.03849497810006142, + 0.029542740434408188, + 0.033062249422073364, + 0.01374698057770729, + 0.009176847524940968, + 0.010860194452106953, + 0.010057210922241211, + 0.007296242751181126, + 0.010968324728310108, + 0.01145781110972166, + 0.009799634106457233, + 0.007514071650803089, + 0.0064627667888998985, + 0.005728783085942268, + 0.004480734001845121, + 0.003565836464986205, + 0.006957532837986946, + 0.006133439019322395, + 0.007574588991701603, + 0.01457150001078844, + 0.017767416313290596, + 0.009673538617789745, + 0.0064234184101223946, + 0.005673747975379229, + 0.013770973309874535, + 0.006251542828977108, + 0.008935577236115932, + 0.0056640105322003365, + 0.0038965886924415827, + 0.0048809475265443325, + 0.0037789279595017433, + 0.0033134720288217068, + 0.0048119984567165375, + 0.005274713039398193, + 0.0038851851131767035, + 0.00904632918536663, + 0.0101071298122406, + 0.006619293708354235, + 0.005612732842564583, + 0.003748183837160468, + 0.009833784773945808, + 0.004764434415847063, + 0.0066537135280668736, + 0.010838225483894348, + 0.013025769963860512, + 0.01045816671103239, + 0.007029339671134949, + 0.006327922455966473, + 0.010054203681647778, + 0.006758722476661205, + 0.00959227979183197, + 0.0036821900866925716, + 0.005263908766210079, + 0.004242505878210068, + 0.003989561926573515, + 0.0034719984978437424, + 0.004944575019180775, + 0.004472475498914719, + 0.003853056812658906, + 0.02439822442829609, + 0.013522998429834843, + 0.02655038796365261, + 0.01352554652839899, + 0.013171753846108913, + 0.022878864780068398, + 0.011438657529652119, + 0.02448972314596176, + 0.024383092299103737, + 0.017075849696993828, + 0.031418055295944214, + 0.016733450815081596, + 0.021334383636713028, + 0.024486061185598373, + 0.031139956787228584, + 0.02930261939764023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.043290942907333374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033696141093969345 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.033696141093969345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033696141093969345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20647592842578888, + 0.08777085691690445, + 0.08753260225057602, + 0.11704987287521362, + 0.05722297728061676, + 0.027024410665035248, + 0.029904969036579132, + 0.06052646413445473, + 0.12484346330165863, + 0.20164839923381805 + ], + "credit_map": [ + 0.027353430166840553, + 0.03173931688070297, + 0.023093583062291145, + 0.022697601467370987, + 0.031297001987695694, + 0.023498481139540672, + 0.026410119608044624, + 0.020386414602398872, + 0.008770359680056572, + 0.010709136724472046, + 0.009667103178799152, + 0.011282137595117092, + 0.012821011245250702, + 0.009492190554738045, + 0.01318273413926363, + 0.011846194043755531, + 0.007434484548866749, + 0.008973886258900166, + 0.007633695844560862, + 0.009730109944939613, + 0.01071811281144619, + 0.016678446903824806, + 0.011806760914623737, + 0.014557107351720333, + 0.01249422412365675, + 0.012601015157997608, + 0.012961277738213539, + 0.016095958650112152, + 0.01827465184032917, + 0.012267973273992538, + 0.019127827137708664, + 0.013226953335106373, + 0.006090407259762287, + 0.005517773795872927, + 0.006530000828206539, + 0.008472561836242676, + 0.0069237095303833485, + 0.006417122203856707, + 0.009831284172832966, + 0.007440121378749609, + 0.002603400032967329, + 0.0022782154846936464, + 0.002876709448173642, + 0.004459693096578121, + 0.003250280162319541, + 0.0028225965797901154, + 0.005255593452602625, + 0.003477925667539239, + 0.0025020006578415632, + 0.0040355101227760315, + 0.0025955350138247013, + 0.003058966249227524, + 0.003592011984437704, + 0.005392829887568951, + 0.004265608265995979, + 0.004462507553398609, + 0.006514681037515402, + 0.00655435910448432, + 0.004909094888716936, + 0.007382113486528397, + 0.009998871944844723, + 0.0077988081611692905, + 0.011141321621835232, + 0.006227222271263599, + 0.014914654195308685, + 0.012065475806593895, + 0.015279104933142662, + 0.012820018455386162, + 0.02542627975344658, + 0.012168623507022858, + 0.01747424341738224, + 0.014695072546601295, + 0.017395082861185074, + 0.012803819961845875, + 0.02246600203216076, + 0.018031498417258263, + 0.03748275712132454, + 0.023448655381798744, + 0.04680658504366875, + 0.023214012384414673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.033696141093969345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818636804819107 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.01818636804819107, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818636804819107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2211187183856964, + 0.10229913890361786, + 0.054547689855098724, + 0.050891630351543427, + 0.07020843774080276, + 0.04701942950487137, + 0.04161247983574867, + 0.03173041716217995, + 0.0699363648891449, + 0.31063568592071533 + ], + "credit_map": [ + 0.025999873876571655, + 0.03261658549308777, + 0.02453315630555153, + 0.021025922149419785, + 0.037744734436273575, + 0.02627290040254593, + 0.03201278671622276, + 0.020912742242217064, + 0.017383459955453873, + 0.012378445826470852, + 0.01437890063971281, + 0.011706828139722347, + 0.0104136997833848, + 0.01030177716165781, + 0.014867566525936127, + 0.010868457145988941, + 0.008762937039136887, + 0.008282549679279327, + 0.006467999424785376, + 0.005573700647801161, + 0.005182729102671146, + 0.005605361890047789, + 0.008433862589299679, + 0.0062385485507547855, + 0.007301939185708761, + 0.00544761773198843, + 0.00562990503385663, + 0.0058237467892467976, + 0.005768629722297192, + 0.0056814271956682205, + 0.008992472663521767, + 0.006245891097933054, + 0.007446007337421179, + 0.009094965644180775, + 0.009007046930491924, + 0.008482106029987335, + 0.009202600456774235, + 0.007518717087805271, + 0.01090477779507637, + 0.00855221226811409, + 0.006371976342052221, + 0.008740958757698536, + 0.005152526777237654, + 0.005134926177561283, + 0.0050486051477491856, + 0.0054631768725812435, + 0.006724636070430279, + 0.004382621496915817, + 0.0032088521402329206, + 0.00357779860496521, + 0.0038858652114868164, + 0.0031618517823517323, + 0.004144728183746338, + 0.010373705066740513, + 0.004519764333963394, + 0.008739912882447243, + 0.004730521235615015, + 0.0032373485155403614, + 0.0033258788753300905, + 0.0029087127186357975, + 0.005682629533112049, + 0.003911856561899185, + 0.003885821672156453, + 0.00404764711856842, + 0.009872195310890675, + 0.005485235713422298, + 0.008577220141887665, + 0.006148146465420723, + 0.012560700997710228, + 0.008245635777711868, + 0.01008966937661171, + 0.008957556448876858, + 0.032854631543159485, + 0.017172036692500114, + 0.05084099993109703, + 0.01919035241007805, + 0.047792669385671616, + 0.060386527329683304, + 0.041155632585287094, + 0.0412428192794323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.01818636804819107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094076342880726 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004094076342880726, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094076342880726, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23744893074035645, + 0.05473263934254646, + 0.03652495890855789, + 0.030522970482707024, + 0.030274413526058197, + 0.06626427173614502, + 0.07973739504814148, + 0.05043325573205948, + 0.05928506329655647, + 0.3547760844230652 + ], + "credit_map": [ + 0.02876223251223564, + 0.027332348749041557, + 0.025812026113271713, + 0.026925625279545784, + 0.03860201686620712, + 0.03168648108839989, + 0.03046562895178795, + 0.02786257304251194, + 0.00942353717982769, + 0.006139171309769154, + 0.009541592560708523, + 0.0054256608709692955, + 0.005221094470471144, + 0.006290475372225046, + 0.0060223001055419445, + 0.006668809335678816, + 0.004427413921803236, + 0.0038021744694560766, + 0.005974511615931988, + 0.0036381317768245935, + 0.0038424355443567038, + 0.005769307725131512, + 0.004143252037465572, + 0.004927730653434992, + 0.00407015485689044, + 0.004129898734390736, + 0.0026556900702416897, + 0.003019556635990739, + 0.00374817312695086, + 0.005550250876694918, + 0.004367456771433353, + 0.0029817898757755756, + 0.003799654310569167, + 0.0036109574139118195, + 0.003781961975619197, + 0.003927708137780428, + 0.0030926265753805637, + 0.004319922998547554, + 0.0047121779061853886, + 0.003029405837878585, + 0.012206371873617172, + 0.012499190866947174, + 0.008556986227631569, + 0.0058096665889024734, + 0.004893715027719736, + 0.00901162438094616, + 0.006941123399883509, + 0.0063455975614488125, + 0.011298613622784615, + 0.012729506939649582, + 0.009097963571548462, + 0.00622983742505312, + 0.005735377315431833, + 0.013335020281374454, + 0.006422885227948427, + 0.014888190664350986, + 0.008025516755878925, + 0.007538496516644955, + 0.0062798731960356236, + 0.004884449765086174, + 0.005964555777609348, + 0.005916413851082325, + 0.005692708306014538, + 0.006131240166723728, + 0.007708961144089699, + 0.009224272333085537, + 0.00701384712010622, + 0.005392323713749647, + 0.008702932856976986, + 0.0071704210713505745, + 0.006097845733165741, + 0.007974459789693356, + 0.051398251205682755, + 0.038142066448926926, + 0.04454336315393448, + 0.03345712274312973, + 0.042008791118860245, + 0.056162070482969284, + 0.04187200963497162, + 0.04719242453575134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.004094076342880726 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6933536529541016 + ], + "opposite_logprob_tokens": [ + -0.6933536529541016 + ], + "logprob_reward": -0.6933536529541016, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933536529541016, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09281130880117416, + 0.04396481066942215, + 0.03405012935400009, + 0.06313415616750717, + 0.03575974702835083, + 0.027313070371747017, + 0.08072824776172638, + 0.0811290591955185, + 0.0696197971701622, + 0.47148969769477844 + ], + "credit_map": [ + 0.009814831428229809, + 0.00985098909586668, + 0.013463854789733887, + 0.010412517003715038, + 0.014126947149634361, + 0.010991078801453114, + 0.01415407657623291, + 0.009997011162340641, + 0.004752577282488346, + 0.0056022945791482925, + 0.005025554448366165, + 0.004293723031878471, + 0.006290771067142487, + 0.004465756472200155, + 0.006952218245714903, + 0.006581912748515606, + 0.005132521037012339, + 0.003682877402752638, + 0.003989980556070805, + 0.0038101773243397474, + 0.004068072885274887, + 0.004148018080741167, + 0.00446731224656105, + 0.00475116865709424, + 0.00616550212725997, + 0.007101988419890404, + 0.0070519824512302876, + 0.006711733527481556, + 0.008629283867776394, + 0.007891743443906307, + 0.011988845653831959, + 0.007593072950839996, + 0.005232613533735275, + 0.007664787583053112, + 0.0048148310743272305, + 0.00393719831481576, + 0.0027145096100866795, + 0.004067405592650175, + 0.003931088838726282, + 0.0033973108511418104, + 0.0042182705365121365, + 0.005335445981472731, + 0.00312767899595201, + 0.0018398413667455316, + 0.002097787568345666, + 0.004385410808026791, + 0.0029316917061805725, + 0.0033769395668059587, + 0.007965222001075745, + 0.006902043707668781, + 0.006969014182686806, + 0.0043777902610599995, + 0.006220380309969187, + 0.021993687376379967, + 0.006544855423271656, + 0.019755244255065918, + 0.013751828111708164, + 0.011156974360346794, + 0.00984871480613947, + 0.006863126996904612, + 0.008747892454266548, + 0.011722318828105927, + 0.008294240571558475, + 0.010743957944214344, + 0.010838969610631466, + 0.006846078205853701, + 0.008632909506559372, + 0.007492329925298691, + 0.007347778882831335, + 0.008970784954726696, + 0.010385634377598763, + 0.00910530611872673, + 0.06259649991989136, + 0.04114605113863945, + 0.062475740909576416, + 0.0366082601249218, + 0.05679934471845627, + 0.09253036230802536, + 0.049317024648189545, + 0.07001639157533646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.6933536529541016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "low-angle, static medium shot captures a rain-slicked, narrow alleyway in a sprawling cyberpunk cityscape, rendered in stunningly photorealistic CGI. The scene is drenched in a pulsating glow of magenta and cyan neon emanating from an unseen source above, casting long, deep shadows across the grimy, graffiti-covered concrete walls. At the center of the frame, a four-legged robot dog crouches on the ground, its chassis constructed from interlocking, scuffed panels of gunmetal gray and matte black. Its optical sensors emit a harsh electric blue, casting a faint, partial glow on its snout as it steadily and deliberately gnaws at a piece of abandoned high-tech machinery—a cracked server casing, revealing gleaming circuit boards. Its jagged metal jaws grind with mechanical force, and with each powerful bite, its head twitches, sending a trail of tiny sparks into the humid air. Suddenly, the massive, fast-moving shadow of a hovercraft passing overhead sweeps across the scene from left to right, momentarily plunging the alley into near-total darkness before vibrant neon floods back in, its dynamic shifts highlighting the intricate details of the robot’s articulated limbs and the textured sheen of its metal frame against the wet, reflective ground.", + "seed": 15319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_11/index_1756_seed_15319_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": -0.22463167514652013, + "vlm_logprob": -0.22463167514652013, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.22463167514652013, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1476559191942215 + ], + "opposite_logprob_tokens": [ + -2.022655963897705 + ], + "logprob_reward": -0.1476559191942215, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1476559191942215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23388820886611938, + 0.10396576672792435, + 0.25820621848106384, + 0.03382355347275734, + 0.03468862920999527, + 0.02312634140253067, + 0.03215194121003151, + 0.03847206383943558, + 0.042831022292375565, + 0.19884617626667023 + ], + "credit_map": [ + 0.024528102949261665, + 0.029231125488877296, + 0.03536304086446762, + 0.024778127670288086, + 0.01712922379374504, + 0.053899455815553665, + 0.014730183407664299, + 0.03422898054122925, + 0.009281965903937817, + 0.0026831626892089844, + 0.04032759740948677, + 0.005386647302657366, + 0.003778129583224654, + 0.005139343906193972, + 0.004214614164084196, + 0.03315431997179985, + 0.020148128271102905, + 0.007787670008838177, + 0.10300596803426743, + 0.011779592372477055, + 0.007235757075250149, + 0.010506083257496357, + 0.005834668874740601, + 0.09190835803747177, + 0.0048041264526546, + 0.003564554499462247, + 0.003876647213473916, + 0.004348557908087969, + 0.004432314075529575, + 0.005462840665131807, + 0.0027051023207604885, + 0.00462941313162446, + 0.0034265678841620684, + 0.0020941251423209906, + 0.01014698576182127, + 0.003704313887283206, + 0.0021407529711723328, + 0.002497804118320346, + 0.0019071274437010288, + 0.00877095852047205, + 0.00233822176232934, + 0.0015327917644754052, + 0.004093791823834181, + 0.0028161881491541862, + 0.0013724126620218158, + 0.004313875455409288, + 0.002597733633592725, + 0.0040613277815282345, + 0.002596487058326602, + 0.002716465387493372, + 0.004674233961850405, + 0.0030208874959498644, + 0.004613111261278391, + 0.00520826643332839, + 0.0031910822726786137, + 0.006131408736109734, + 0.004443466663360596, + 0.0031190295703709126, + 0.006585428956896067, + 0.006627049762755632, + 0.004391021560877562, + 0.003903751028701663, + 0.004185102880001068, + 0.00521721551194787, + 0.00511822197586298, + 0.004131137393414974, + 0.0055166250094771385, + 0.00597170228138566, + 0.006472416687756777, + 0.004491597414016724, + 0.004317439626902342, + 0.006811885163187981, + 0.01564723253250122, + 0.012551041319966316, + 0.021396372467279434, + 0.01654074899852276, + 0.026465313509106636, + 0.036095693707466125, + 0.022046910598874092, + 0.04810287803411484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.1476559191942215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865352809429169 + ], + "opposite_logprob_tokens": [ + -3.0486536026000977 + ], + "logprob_reward": -0.04865352809429169, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865352809429169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1835242062807083, + 0.30461713671684265, + 0.0466378889977932, + 0.11509041488170624, + 0.051889825612306595, + 0.029682785272598267, + 0.11662168800830841, + 0.07901523262262344, + 0.031280431896448135, + 0.04164035618305206 + ], + "credit_map": [ + 0.01828041672706604, + 0.02047625370323658, + 0.048866547644138336, + 0.01672535575926304, + 0.013285394757986069, + 0.0139400539919734, + 0.014855178073048592, + 0.03709501400589943, + 0.02633087895810604, + 0.007085665129125118, + 0.11877185851335526, + 0.01855158992111683, + 0.008997165597975254, + 0.012151661328971386, + 0.013822715729475021, + 0.09890560805797577, + 0.003929873928427696, + 0.0029585508164018393, + 0.011285990476608276, + 0.004405337385833263, + 0.0038059991784393787, + 0.003966922871768475, + 0.00469775777310133, + 0.011587459594011307, + 0.007932203821837902, + 0.009821712039411068, + 0.017907030880451202, + 0.013448142446577549, + 0.020013947039842606, + 0.012348424643278122, + 0.013879290781915188, + 0.019739659503102303, + 0.0049188025295734406, + 0.004884921479970217, + 0.009245265275239944, + 0.006027974653989077, + 0.006895647384226322, + 0.007308142725378275, + 0.0054674637503921986, + 0.0071416073478758335, + 0.0026238542050123215, + 0.0027656718157231808, + 0.004039484076201916, + 0.0071174027398228645, + 0.0026959567330777645, + 0.002854833146557212, + 0.0031258035451173782, + 0.0044597783125936985, + 0.009699285961687565, + 0.013447755947709084, + 0.01746758446097374, + 0.011912696994841099, + 0.013175425119698048, + 0.019421318545937538, + 0.010200160555541515, + 0.021297458559274673, + 0.006884085014462471, + 0.010190113447606564, + 0.00797970499843359, + 0.008634688332676888, + 0.010263868607580662, + 0.00845907349139452, + 0.012904859147965908, + 0.013698840513825417, + 0.002669053850695491, + 0.003712268080562353, + 0.004533363040536642, + 0.004026592709124088, + 0.0055570947006344795, + 0.0029201849829405546, + 0.003318407107144594, + 0.004543468356132507, + 0.0034992045257240534, + 0.004217613022774458, + 0.004565668758004904, + 0.0036038022954016924, + 0.006870110984891653, + 0.005151658318936825, + 0.006381936836987734, + 0.00735035864636302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.04865352809429169 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284580558538437 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.023284580558538437, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284580558538437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11617300659418106, + 0.6960355639457703, + 0.03569870814681053, + 0.03182138502597809, + 0.03505363687872887, + 0.011530248448252678, + 0.01335874106734991, + 0.017691249027848244, + 0.013149973936378956, + 0.029487429186701775 + ], + "credit_map": [ + 0.011722879484295845, + 0.008320820517838001, + 0.03336014598608017, + 0.00897494237869978, + 0.009704098105430603, + 0.011267609894275665, + 0.0076598068699240685, + 0.025162702426314354, + 0.06100120395421982, + 0.011158440262079239, + 0.28691476583480835, + 0.036225151270627975, + 0.013313600793480873, + 0.02714592032134533, + 0.018205512315034866, + 0.2420710325241089, + 0.0034401093143969774, + 0.0027542919851839542, + 0.010474528186023235, + 0.0032270776573568583, + 0.0020839679054915905, + 0.0028732812497764826, + 0.0017901554238051176, + 0.009055295959115028, + 0.0019173301989212632, + 0.0033553384710103273, + 0.0048283361829817295, + 0.0033495179377496243, + 0.005613340530544519, + 0.004510185681283474, + 0.0030120285227894783, + 0.005235306452959776, + 0.0038411079440265894, + 0.003213384421542287, + 0.0052627441473305225, + 0.004220530390739441, + 0.005879057105630636, + 0.004187642131000757, + 0.0039008900057524443, + 0.0045482804998755455, + 0.0009690463775768876, + 0.0008751940913498402, + 0.0021753194741904736, + 0.0012489069486036897, + 0.0010740428697317839, + 0.0013273744843900204, + 0.0011214329861104488, + 0.0027389314491301775, + 0.0013085780665278435, + 0.0011172184022143483, + 0.002353751566261053, + 0.0018690134165808558, + 0.001400265726260841, + 0.001745572779327631, + 0.001214321469888091, + 0.0023500204551965, + 0.0020548037718981504, + 0.0015355207724496722, + 0.001949415309354663, + 0.0026790956035256386, + 0.002702793339267373, + 0.0016402414767071605, + 0.0022274930961430073, + 0.0029018858913332224, + 0.001051508355885744, + 0.0013770632212981582, + 0.0021120989695191383, + 0.001363483490422368, + 0.0024764928966760635, + 0.0015862556174397469, + 0.0008742643985897303, + 0.0023088071029633284, + 0.002236283151432872, + 0.0021216808818280697, + 0.002494917018339038, + 0.00333620747551322, + 0.003375991014763713, + 0.003660729620605707, + 0.005174362100660801, + 0.007087258622050285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.023284580558538437 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.15602436661720276, + 0.4155300557613373, + 0.08450599759817123, + 0.04250477999448776, + 0.07417289167642593, + 0.028800057247281075, + 0.07852926105260849, + 0.03324580565094948, + 0.0167042575776577, + 0.06998252123594284 + ], + "credit_map": [ + 0.016804946586489677, + 0.013238705694675446, + 0.052592482417821884, + 0.011243019253015518, + 0.006225569639354944, + 0.009975165128707886, + 0.010095731355249882, + 0.035848747938871384, + 0.03705092519521713, + 0.00700360769405961, + 0.17074957489967346, + 0.022000649943947792, + 0.008394611068069935, + 0.016254408285021782, + 0.011662492528557777, + 0.14241380989551544, + 0.006872781086713076, + 0.0046652634628117085, + 0.029892485588788986, + 0.005194465629756451, + 0.003679456189274788, + 0.004542839247733355, + 0.0026351117994636297, + 0.027023594826459885, + 0.0034452807158231735, + 0.0032884785905480385, + 0.008149120025336742, + 0.00540706655010581, + 0.004891548305749893, + 0.00454316521063447, + 0.003721150103956461, + 0.009058969095349312, + 0.006845687981694937, + 0.004160688724368811, + 0.021553892642259598, + 0.0065169972367584705, + 0.0056665497832000256, + 0.005983459763228893, + 0.004003265406936407, + 0.019442353397607803, + 0.0021947589702904224, + 0.0018634950974956155, + 0.006104294676333666, + 0.003138912608847022, + 0.003066873177886009, + 0.003690337296575308, + 0.0022425162605941296, + 0.006498871836811304, + 0.00823402963578701, + 0.005157016217708588, + 0.013068058528006077, + 0.011683792807161808, + 0.007967189885675907, + 0.012125205248594284, + 0.006044486071914434, + 0.014249488711357117, + 0.0027824428398162127, + 0.0035327032674103975, + 0.003938029054552317, + 0.005364065058529377, + 0.003982476890087128, + 0.004233292769640684, + 0.004231203347444534, + 0.005181597080081701, + 0.0019379418808966875, + 0.0013259421102702618, + 0.0025945811066776514, + 0.002439345931634307, + 0.0021808433812111616, + 0.0017928758170455694, + 0.0019532740116119385, + 0.002479453105479479, + 0.004807257559150457, + 0.005627877544611692, + 0.006656462326645851, + 0.010327032767236233, + 0.008015701547265053, + 0.006750540342181921, + 0.015711138024926186, + 0.012086514383554459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 narrow alleyway in a cyberpunk cityscape with rain-slicked, graffiti-covered concrete walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a four-legged robot dog crouching at the center of the frame, gnawing on a cracked server casing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the robot dog have a gunmetal gray and matte black chassis with optical sensors emitting electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by pulsating magenta and cyan neon light casting long, deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a massive shadow of a hovercraft sweep across the scene from left to right, momentarily plunging the alley into darkness?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8983126878738403 + ], + "opposite_logprob_tokens": [ + -0.5233126878738403 + ], + "logprob_reward": -0.8983126878738403, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983126878738403, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08273840695619583, + 0.43020644783973694, + 0.07230938225984573, + 0.028327476233243942, + 0.080751433968544, + 0.0459611751139164, + 0.048869673162698746, + 0.04015091061592102, + 0.022189397364854813, + 0.14849574863910675 + ], + "credit_map": [ + 0.007909965701401234, + 0.010723100043833256, + 0.014413181692361832, + 0.008240430615842342, + 0.007849651388823986, + 0.016086947172880173, + 0.006433905567973852, + 0.011081226170063019, + 0.03800167888402939, + 0.0068603879772126675, + 0.17686818540096283, + 0.022240981459617615, + 0.00840879324823618, + 0.017092173919081688, + 0.011532975360751152, + 0.14920124411582947, + 0.007019837386906147, + 0.0033661695197224617, + 0.023623192682862282, + 0.005293253343552351, + 0.004896762315183878, + 0.004148195497691631, + 0.004192972555756569, + 0.019769003614783287, + 0.0034102911595255136, + 0.0040734815411269665, + 0.0034811212681233883, + 0.004039278253912926, + 0.0026173319201916456, + 0.004898230079561472, + 0.0022439025342464447, + 0.003563839243724942, + 0.006105345673859119, + 0.0031570640858262777, + 0.028964588418602943, + 0.006121804937720299, + 0.003289892105385661, + 0.004100222606211901, + 0.002757719252258539, + 0.026254795491695404, + 0.00530256936326623, + 0.004895810969173908, + 0.009515433572232723, + 0.005689661949872971, + 0.0026512935291975737, + 0.004450396168977022, + 0.003313161665573716, + 0.010142850689589977, + 0.0051786634139716625, + 0.004413696471601725, + 0.009349944069981575, + 0.008274592459201813, + 0.006471213418990374, + 0.005562186241149902, + 0.004034772049635649, + 0.005584607366472483, + 0.002998827025294304, + 0.0030285792890936136, + 0.004540658555924892, + 0.007025743369013071, + 0.004757216200232506, + 0.0029273941181600094, + 0.008310901001095772, + 0.00656158896163106, + 0.00234802532941103, + 0.0024856729432940483, + 0.0039698039181530476, + 0.0027902910951524973, + 0.004133733455091715, + 0.002169115701690316, + 0.0016580568626523018, + 0.0026346987579017878, + 0.013663874007761478, + 0.014613957144320011, + 0.019950490444898605, + 0.019720789045095444, + 0.013819458894431591, + 0.025509949773550034, + 0.015374223701655865, + 0.025843007490038872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.8983126878738403 + } + ], + "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/0000152/group_12/index_tensor([1605, 1605])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_tensor([1605, 1605])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..343e437af4397d8c1903cdfaa7c25bca9be6eebb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_tensor([1605, 1605])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, first-person handheld camera shot oscillates forward through a vast, abstract digital space constructed in a low-poly style. At the center of the image is a vast information flow representing the Sina portal website. A majestic cuboid, composed of countless tiny, low-polygonal facets gleaming in the signature Sina red and dotted with blue and white dots, it hovers in the pitch-black void, emitting an intense, steady digital glow that illuminates the floating, fragmented geometric structures around it. An invisible force slowly begins to warp this massive object, rippling and cracking its surface. Finally, in a silent yet visually striking slow-motion explosion, it shatters, sending vast red polygonal fragments and countless smaller dots of light scattering violently in all directions. As these fragments tumble and fly in the chaotic chaos, they miraculously transform and reshape, transforming into countless luminous, three-dimensional Chinese news headlines that hurtle toward every corner of the space at varying speeds and trajectories. The handheld camera dangerously shuttles, sways and tilts between these rapidly passing data streams and news headlines, striving to capture the dynamic moment of this information explosion. Light and shadow change rapidly in front of the lens, creating a strong visual atmosphere that is both chaotic and spectacular, about information overload and deconstruction.", + "seed": 9323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_1605_seed_9323_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": -0.026603216282092036, + "vlm_logprob": -0.026603216282092036, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.026603216282092036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09110694378614426 + ], + "opposite_logprob_tokens": [ + -3.09110689163208 + ], + "logprob_reward": -0.09110694378614426, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.09110694378614426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13927726447582245, + 0.3387065529823303, + 0.05837283656001091, + 0.0334891639649868, + 0.024518484249711037, + 0.0603545643389225, + 0.05480543151497841, + 0.06368984282016754, + 0.08535320311784744, + 0.14143261313438416 + ], + "credit_map": [ + 0.021818501874804497, + 0.01276890467852354, + 0.028523029759526253, + 0.012043518014252186, + 0.009193256497383118, + 0.019964324310421944, + 0.010141353122889996, + 0.024824369698762894, + 0.04279705882072449, + 0.06436904519796371, + 0.04145171120762825, + 0.0384591743350029, + 0.04616357013583183, + 0.02427363209426403, + 0.046320267021656036, + 0.03487209975719452, + 0.005653162952512503, + 0.008280422538518906, + 0.005855583120137453, + 0.004871921613812447, + 0.0035712039098143578, + 0.015536738559603691, + 0.0036932320799678564, + 0.010910571552813053, + 0.002810654928907752, + 0.004071836359798908, + 0.003841155907139182, + 0.003754069097340107, + 0.004528389777988195, + 0.004472589120268822, + 0.003576879156753421, + 0.0064335898496210575, + 0.0029544420540332794, + 0.0035656567197293043, + 0.0025489088147878647, + 0.0031069652177393436, + 0.002693519229069352, + 0.00358657562173903, + 0.0034375861287117004, + 0.0026248302310705185, + 0.0062551358714699745, + 0.00436051283031702, + 0.006302140187472105, + 0.00564687792211771, + 0.008498352952301502, + 0.00816689245402813, + 0.014442975632846355, + 0.006681674625724554, + 0.004421810619533062, + 0.002990233013406396, + 0.009916562587022781, + 0.007564605679363012, + 0.005495046265423298, + 0.004302019719034433, + 0.015548432245850563, + 0.004566721152514219, + 0.0072011398151516914, + 0.011724916286766529, + 0.008980131708085537, + 0.010428016074001789, + 0.006101714447140694, + 0.008457332849502563, + 0.005678429733961821, + 0.005118159111589193, + 0.010157499462366104, + 0.015419275499880314, + 0.012597729451954365, + 0.008753160014748573, + 0.005379424896091223, + 0.007010608911514282, + 0.017304755747318268, + 0.008730748668313026, + 0.020384356379508972, + 0.024034572765231133, + 0.023300569504499435, + 0.02316228486597538, + 0.015049011446535587, + 0.01029378455132246, + 0.01613418199121952, + 0.009073835797607899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.09110694378614426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804898872971535 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.009804898872971535, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804898872971535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2617386281490326, + 0.1486460417509079, + 0.09194532036781311, + 0.03327758610248566, + 0.04443097859621048, + 0.04125262051820755, + 0.12896694242954254, + 0.10001019388437271, + 0.09165080636739731, + 0.058080848306417465 + ], + "credit_map": [ + 0.03146227076649666, + 0.02843013033270836, + 0.04217098653316498, + 0.024920368567109108, + 0.023894324898719788, + 0.04016031324863434, + 0.023741193115711212, + 0.04695907235145569, + 0.017769385129213333, + 0.024189341813325882, + 0.01724489964544773, + 0.01830139383673668, + 0.019052503630518913, + 0.015073534101247787, + 0.0191323421895504, + 0.017882663756608963, + 0.01186088752001524, + 0.009829122573137283, + 0.01369036827236414, + 0.007111681159585714, + 0.00690313708037138, + 0.024099504575133324, + 0.007087859325110912, + 0.011362772434949875, + 0.004819875117391348, + 0.004388024564832449, + 0.004209858365356922, + 0.003352464409545064, + 0.0036623093765228987, + 0.004981012549251318, + 0.0032545248977839947, + 0.004609520547091961, + 0.006040336098521948, + 0.01075847540050745, + 0.004857612308114767, + 0.005576490890234709, + 0.005664569325745106, + 0.004423920996487141, + 0.003637318266555667, + 0.0034722599666565657, + 0.004150480963289738, + 0.003659224370494485, + 0.004928924143314362, + 0.003934117499738932, + 0.005775548983365297, + 0.006352144293487072, + 0.00805454608052969, + 0.004397635813802481, + 0.00943575520068407, + 0.0067054773680865765, + 0.018048331141471863, + 0.01969168335199356, + 0.011500694788992405, + 0.00913791824132204, + 0.04382282495498657, + 0.01062428206205368, + 0.009052817709743977, + 0.011270260438323021, + 0.013919439166784286, + 0.016365552321076393, + 0.011082041077315807, + 0.022081000730395317, + 0.008091743104159832, + 0.008147354237735271, + 0.00823926366865635, + 0.014240991324186325, + 0.013873792253434658, + 0.007537114433944225, + 0.007606285624206066, + 0.009886200539767742, + 0.02116401121020317, + 0.009103155694901943, + 0.007335781119763851, + 0.007639814168214798, + 0.010191640816628933, + 0.01133530680090189, + 0.005207838490605354, + 0.005583163816481829, + 0.0057661645114421844, + 0.0050211455672979355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009804898872971535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011834054021164775 + ], + "opposite_logprob_tokens": [ + -6.75118350982666 + ], + "logprob_reward": -0.0011834054021164775, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011834054021164775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19687311351299286, + 0.07568102329969406, + 0.09505777060985565, + 0.06336146593093872, + 0.04751252382993698, + 0.11392027139663696, + 0.07331125438213348, + 0.07875774055719376, + 0.18188932538032532, + 0.07363550364971161 + ], + "credit_map": [ + 0.029165755957365036, + 0.02362244203686714, + 0.027051853016018867, + 0.01998516172170639, + 0.01905287243425846, + 0.03417269513010979, + 0.019896436482667923, + 0.023925891146063805, + 0.011719402857124805, + 0.009259688667953014, + 0.009391136467456818, + 0.008475320413708687, + 0.006941931322216988, + 0.00915060006082058, + 0.006372522562742233, + 0.014370422810316086, + 0.009614795446395874, + 0.01368794310837984, + 0.012664166279137135, + 0.00866120494902134, + 0.010045037604868412, + 0.01733795925974846, + 0.007168799173086882, + 0.015877874568104744, + 0.006807380355894566, + 0.012844052165746689, + 0.010384585708379745, + 0.0039239912293851376, + 0.0069616311229765415, + 0.007560985628515482, + 0.005559816490858793, + 0.009319022297859192, + 0.007989645004272461, + 0.008998420089483261, + 0.005612086039036512, + 0.006064458284527063, + 0.006807238794863224, + 0.0036509742494672537, + 0.00442512147128582, + 0.003964581061154604, + 0.018300149589776993, + 0.011624366976320744, + 0.013597938232123852, + 0.012632310390472412, + 0.015367867425084114, + 0.017150986939668655, + 0.01499843504279852, + 0.010248224250972271, + 0.006763112265616655, + 0.0049283732660114765, + 0.011652367189526558, + 0.009989192709326744, + 0.009130307473242283, + 0.007689598016440868, + 0.01639680191874504, + 0.006761500611901283, + 0.007455774117261171, + 0.008525880984961987, + 0.010488247498869896, + 0.012576955370604992, + 0.00943132396787405, + 0.01783131994307041, + 0.005224849563091993, + 0.00722339004278183, + 0.018204059451818466, + 0.032034922391176224, + 0.02689455635845661, + 0.016602249816060066, + 0.016238680109381676, + 0.016467448323965073, + 0.03920712694525719, + 0.016240287572145462, + 0.009959863498806953, + 0.00800009910017252, + 0.010664897039532661, + 0.013783412985503674, + 0.006279150489717722, + 0.004764689598232508, + 0.010247929021716118, + 0.00993545725941658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0011834054021164775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023278988897800446 + ], + "opposite_logprob_tokens": [ + -3.7732789516448975 + ], + "logprob_reward": -0.023278988897800446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023278988897800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16663861274719238, + 0.10508181154727936, + 0.07228308916091919, + 0.03553497791290283, + 0.04069299250841141, + 0.047628846019506454, + 0.06833633035421371, + 0.21067920327186584, + 0.15502217411994934, + 0.09810202568769455 + ], + "credit_map": [ + 0.02482524886727333, + 0.014488752000033855, + 0.026233240962028503, + 0.016516664996743202, + 0.01035246066749096, + 0.025755492970347404, + 0.010609914548695087, + 0.037856828421354294, + 0.014062747359275818, + 0.01608741097152233, + 0.012790980748832226, + 0.010988770052790642, + 0.011407017707824707, + 0.010377473197877407, + 0.012820571660995483, + 0.01654682122170925, + 0.008255256339907646, + 0.009832113049924374, + 0.008891047909855843, + 0.006487133912742138, + 0.0067214155569672585, + 0.013465603813529015, + 0.005456591956317425, + 0.013173915445804596, + 0.007224455010145903, + 0.004204625263810158, + 0.0036307901609688997, + 0.005289153195917606, + 0.0040520490147173405, + 0.004252339247614145, + 0.003217660589143634, + 0.003663900773972273, + 0.0044626095332205296, + 0.009877112694084644, + 0.004217307548969984, + 0.004865165334194899, + 0.003933433443307877, + 0.004290962591767311, + 0.004848373122513294, + 0.004198024515062571, + 0.003550618188455701, + 0.004023770801723003, + 0.005543932784348726, + 0.005310372915118933, + 0.006590632721781731, + 0.004599414300173521, + 0.012106231413781643, + 0.0059038689360022545, + 0.005082403309643269, + 0.004263097420334816, + 0.00741724343970418, + 0.008121686987578869, + 0.009463343769311905, + 0.005116578657180071, + 0.022738464176654816, + 0.006133506074547768, + 0.021906498819589615, + 0.034007471054792404, + 0.03166018798947334, + 0.03159264475107193, + 0.02148493006825447, + 0.03727192059159279, + 0.015913819894194603, + 0.016841724514961243, + 0.018461791798472404, + 0.032920099794864655, + 0.022255688905715942, + 0.015977639704942703, + 0.015090697444975376, + 0.01601351797580719, + 0.02150888554751873, + 0.01279382687062025, + 0.011697093024849892, + 0.013549775816500187, + 0.01856248453259468, + 0.02071831189095974, + 0.010085750371217728, + 0.005843240302056074, + 0.008693471550941467, + 0.008951888419687748 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.023278988897800446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1556733250617981, + 0.23344780504703522, + 0.03182987496256828, + 0.03460555523633957, + 0.02654329128563404, + 0.035900961607694626, + 0.13213208317756653, + 0.04642293602228165, + 0.05232779681682587, + 0.2511163651943207 + ], + "credit_map": [ + 0.030451379716396332, + 0.011177843436598778, + 0.025338763371109962, + 0.013894247822463512, + 0.00851079449057579, + 0.034427009522914886, + 0.00868725124746561, + 0.023186033591628075, + 0.029239149764180183, + 0.04240864887833595, + 0.02918890304863453, + 0.02776465378701687, + 0.03151753544807434, + 0.01757282204926014, + 0.031290292739868164, + 0.024465784430503845, + 0.003253054805099964, + 0.0057114409282803535, + 0.0038146553561091423, + 0.0036598267033696175, + 0.002212864812463522, + 0.004920319188386202, + 0.0037016752175986767, + 0.0045560370199382305, + 0.0035672567319124937, + 0.006067742593586445, + 0.005836796946823597, + 0.0038024771492928267, + 0.004147351253777742, + 0.004102608188986778, + 0.0030815796926617622, + 0.0039997403509914875, + 0.002952177543193102, + 0.0028155120089650154, + 0.0027717221528291702, + 0.002695497591048479, + 0.0035680115688592196, + 0.005132309626787901, + 0.002678686985746026, + 0.003929374739527702, + 0.0028976888861507177, + 0.0035262182354927063, + 0.003437558887526393, + 0.004928270820528269, + 0.005449571646749973, + 0.0035038217902183533, + 0.008016505278646946, + 0.004141325131058693, + 0.009194541722536087, + 0.007001996971666813, + 0.01889965496957302, + 0.02190914936363697, + 0.011342203244566917, + 0.008314093574881554, + 0.044948745518922806, + 0.010521689429879189, + 0.004755055066198111, + 0.00628392631188035, + 0.0056835864670574665, + 0.007255141157656908, + 0.007634822744876146, + 0.004903318826109171, + 0.005446447990834713, + 0.004460635129362345, + 0.00616022665053606, + 0.011249855160713196, + 0.006690496578812599, + 0.005886341445147991, + 0.004838193766772747, + 0.004807440098375082, + 0.0073705376125872135, + 0.005324703641235828, + 0.022076306864619255, + 0.02438441477715969, + 0.036916401237249374, + 0.050941601395606995, + 0.0283479206264019, + 0.02071480266749859, + 0.03481492027640343, + 0.032919999212026596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.00764184445142746 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, first-person handheld camera shot oscillates forward through a vast, abstract digital space constructed in a low-poly style. At the center of the image is a vast information flow representing the Sina portal website. A majestic cuboid, composed of countless tiny, low-polygonal facets gleaming in the signature Sina red and dotted with blue and white dots, it hovers in the pitch-black void, emitting an intense, steady digital glow that illuminates the floating, fragmented geometric structures around it. An invisible force slowly begins to warp this massive object, rippling and cracking its surface. Finally, in a silent yet visually striking slow-motion explosion, it shatters, sending vast red polygonal fragments and countless smaller dots of light scattering violently in all directions. As these fragments tumble and fly in the chaotic chaos, they miraculously transform and reshape, transforming into countless luminous, three-dimensional Chinese news headlines that hurtle toward every corner of the space at varying speeds and trajectories. The handheld camera dangerously shuttles, sways and tilts between these rapidly passing data streams and news headlines, striving to capture the dynamic moment of this information explosion. Light and shadow change rapidly in front of the lens, creating a strong visual atmosphere that is both chaotic and spectacular, about information overload and deconstruction.", + "seed": 9324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_1605_seed_9324_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": -0.03618229473941028, + "vlm_logprob": -0.03618229473941028, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.03618229473941028, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07251341640949249 + ], + "opposite_logprob_tokens": [ + -3.5725133419036865 + ], + "logprob_reward": -0.07251341640949249, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.07251341640949249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13476045429706573, + 0.17119218409061432, + 0.06652060151100159, + 0.04856225848197937, + 0.04421614482998848, + 0.05145752802491188, + 0.09723718464374542, + 0.06254296749830246, + 0.0927664041519165, + 0.23074421286582947 + ], + "credit_map": [ + 0.015458250418305397, + 0.010538894683122635, + 0.022668946534395218, + 0.01658634841442108, + 0.011900397017598152, + 0.018390432000160217, + 0.019660435616970062, + 0.01955675147473812, + 0.01947326771914959, + 0.02870766445994377, + 0.021573513746261597, + 0.022454092279076576, + 0.013400713913142681, + 0.020498152822256088, + 0.01907772198319435, + 0.026007065549492836, + 0.006322323810309172, + 0.005396803375333548, + 0.00978325679898262, + 0.0049074869602918625, + 0.004714136011898518, + 0.011696367524564266, + 0.009741751477122307, + 0.013958470895886421, + 0.005888520739972591, + 0.004161015618592501, + 0.0052160401828587055, + 0.004828176461160183, + 0.007639897521585226, + 0.00598573312163353, + 0.006834962870925665, + 0.00800790823996067, + 0.003631654428318143, + 0.004008539952337742, + 0.008008046075701714, + 0.008283212780952454, + 0.007479908410459757, + 0.005234051030129194, + 0.00425699632614851, + 0.0033137351274490356, + 0.00425697909668088, + 0.006217717658728361, + 0.0060916468501091, + 0.008615769445896149, + 0.005980749148875475, + 0.006754986941814423, + 0.0054607093334198, + 0.008078970946371555, + 0.025112822651863098, + 0.009317578747868538, + 0.015577212907373905, + 0.01123417541384697, + 0.009277133271098137, + 0.009695653803646564, + 0.009800108149647713, + 0.007222493179142475, + 0.012704920023679733, + 0.005978820845484734, + 0.006118786986917257, + 0.007900465279817581, + 0.007543842773884535, + 0.009623058140277863, + 0.0051757520996034145, + 0.007497323211282492, + 0.009259769693017006, + 0.01176215149462223, + 0.011322668753564358, + 0.011103885248303413, + 0.01191563531756401, + 0.01532648503780365, + 0.011256128549575806, + 0.010819676332175732, + 0.023531552404165268, + 0.02461051009595394, + 0.031065136194229126, + 0.03461319953203201, + 0.03357390686869621, + 0.02963920310139656, + 0.029525049030780792, + 0.02418564446270466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.07251341640949249 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026386020705103874 + ], + "opposite_logprob_tokens": [ + -3.651386022567749 + ], + "logprob_reward": -0.026386020705103874, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026386020705103874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21922948956489563, + 0.11773339658975601, + 0.06214085966348648, + 0.07419873774051666, + 0.07571391761302948, + 0.07092740386724472, + 0.06489437818527222, + 0.07385803014039993, + 0.11553747951984406, + 0.125766322016716 + ], + "credit_map": [ + 0.02470792643725872, + 0.021280569955706596, + 0.03570476546883583, + 0.03500112518668175, + 0.01580633968114853, + 0.028081819415092468, + 0.036749184131622314, + 0.02189774625003338, + 0.010446314699947834, + 0.018819384276866913, + 0.01846223697066307, + 0.015784122049808502, + 0.012927192263305187, + 0.013159691356122494, + 0.012935805134475231, + 0.015198647044599056, + 0.006150260102003813, + 0.0072851660661399364, + 0.006798745598644018, + 0.007565653417259455, + 0.0048626079224050045, + 0.00673905573785305, + 0.011449198238551617, + 0.011290174908936024, + 0.006986691616475582, + 0.007762621156871319, + 0.008513884618878365, + 0.008151644840836525, + 0.010129034519195557, + 0.011787531897425652, + 0.011274589225649834, + 0.009592738933861256, + 0.005105925723910332, + 0.00643176632001996, + 0.010075568221509457, + 0.01021492201834917, + 0.017068829387426376, + 0.010236302390694618, + 0.008876296691596508, + 0.007704317569732666, + 0.006199374794960022, + 0.005495960358530283, + 0.012713273987174034, + 0.01122167706489563, + 0.007026642095297575, + 0.010418526828289032, + 0.009957998059689999, + 0.007893947884440422, + 0.011786337941884995, + 0.006214193534106016, + 0.011212189681828022, + 0.007189675699919462, + 0.006718347314745188, + 0.007736122235655785, + 0.007543710060417652, + 0.006493795663118362, + 0.012807452119886875, + 0.006423110142350197, + 0.0063637662678956985, + 0.00731121888384223, + 0.008024627342820168, + 0.014869465492665768, + 0.00643253605812788, + 0.01162585336714983, + 0.009806946851313114, + 0.010713126510381699, + 0.01585613191127777, + 0.01885768212378025, + 0.01498712133616209, + 0.011125477962195873, + 0.01582077331840992, + 0.01837022788822651, + 0.010117103345692158, + 0.016975896432995796, + 0.017394136637449265, + 0.016419092193245888, + 0.019195185974240303, + 0.012797338888049126, + 0.017724312841892242, + 0.015143242664635181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.026386020705103874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002203776268288493 + ], + "opposite_logprob_tokens": [ + -6.127203941345215 + ], + "logprob_reward": -0.002203776268288493, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002203776268288493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23990458250045776, + 0.10962314903736115, + 0.09348222613334656, + 0.0549582913517952, + 0.025086387991905212, + 0.03426087647676468, + 0.04977846145629883, + 0.0656944140791893, + 0.06577005982398987, + 0.26144155859947205 + ], + "credit_map": [ + 0.0249212384223938, + 0.023875048384070396, + 0.037621039897203445, + 0.029786549508571625, + 0.027675887569785118, + 0.032690174877643585, + 0.03336431458592415, + 0.02997034415602684, + 0.0108074015006423, + 0.012002001516520977, + 0.01144505012780428, + 0.011852655559778214, + 0.007996157743036747, + 0.01857656240463257, + 0.010843045078217983, + 0.026100274175405502, + 0.00937862042337656, + 0.005610302090644836, + 0.01906386762857437, + 0.006242609117180109, + 0.00678182952105999, + 0.017556995153427124, + 0.011251875199377537, + 0.017596133053302765, + 0.007183829788118601, + 0.005267396569252014, + 0.0077985539101064205, + 0.005485979840159416, + 0.0051822359673678875, + 0.011487840674817562, + 0.00650918809697032, + 0.006043263711035252, + 0.0019295683596283197, + 0.0028238247614353895, + 0.0034964189399033785, + 0.003323109820485115, + 0.0038249818608164787, + 0.0028610588051378727, + 0.003257510717958212, + 0.0035699130967259407, + 0.0029460699297487736, + 0.0031972266733646393, + 0.004180027171969414, + 0.00523032620549202, + 0.004517375025898218, + 0.004695196636021137, + 0.0041625467129051685, + 0.005332109052687883, + 0.007207790855318308, + 0.004241690970957279, + 0.007247891277074814, + 0.005431328434497118, + 0.004862835630774498, + 0.006031271070241928, + 0.008625796996057034, + 0.006129855290055275, + 0.009638182818889618, + 0.005070212297141552, + 0.006748105399310589, + 0.006562291644513607, + 0.0065644425339996815, + 0.0128386951982975, + 0.006140487734228373, + 0.012132002040743828, + 0.004932577256113291, + 0.006787189282476902, + 0.005270090419799089, + 0.011484540067613125, + 0.014592248946428299, + 0.005834958516061306, + 0.009516269899904728, + 0.007352183572947979, + 0.02458132803440094, + 0.023506516590714455, + 0.03852657601237297, + 0.03367406874895096, + 0.03120509162545204, + 0.031744301319122314, + 0.04586844518780708, + 0.03233522176742554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.002203776268288493 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001103460788727 + ], + "opposite_logprob_tokens": [ + -2.6950111389160156 + ], + "logprob_reward": -0.07001103460788727, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001103460788727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354093700647354, + 0.05600826069712639, + 0.03734955936670303, + 0.01801575906574726, + 0.032126832753419876, + 0.04479130357503891, + 0.15060925483703613, + 0.04420044273138046, + 0.08668138831853867, + 0.3948077857494354 + ], + "credit_map": [ + 0.016273902729153633, + 0.01401540357619524, + 0.020999301224946976, + 0.013054956682026386, + 0.011792547069489956, + 0.01880842074751854, + 0.02089008502662182, + 0.019574755802750587, + 0.00542921619489789, + 0.00787422526627779, + 0.005656515248119831, + 0.0046380506828427315, + 0.004287087824195623, + 0.00913300458341837, + 0.005381477996706963, + 0.013608686625957489, + 0.003730854019522667, + 0.003880062373355031, + 0.004526856821030378, + 0.004109738860279322, + 0.004166129045188427, + 0.004724016413092613, + 0.005619608797132969, + 0.006592294201254845, + 0.0018079360015690327, + 0.001793581759557128, + 0.002155369147658348, + 0.0019942105282098055, + 0.002150442684069276, + 0.002414040034636855, + 0.003025276120752096, + 0.0026749027892947197, + 0.002621303778141737, + 0.002516398672014475, + 0.005506723653525114, + 0.004079434089362621, + 0.005557778291404247, + 0.00453402241691947, + 0.003657042048871517, + 0.003654130734503269, + 0.0035493725445121527, + 0.0023017588537186384, + 0.005390094127506018, + 0.011093667708337307, + 0.002208329038694501, + 0.003757819999009371, + 0.004464278928935528, + 0.012025982141494751, + 0.03168538212776184, + 0.014472179114818573, + 0.023902947083115578, + 0.017935186624526978, + 0.011820429936051369, + 0.01509646512567997, + 0.023177005350589752, + 0.0125196548178792, + 0.006936705205589533, + 0.004374601412564516, + 0.0040381732396781445, + 0.004366648383438587, + 0.0036838429514318705, + 0.007292376831173897, + 0.004122420679777861, + 0.009385679848492146, + 0.007311687804758549, + 0.012184680439531803, + 0.008013983257114887, + 0.013159622438251972, + 0.013603214174509048, + 0.009671942330896854, + 0.011946544051170349, + 0.010789724066853523, + 0.03745797649025917, + 0.03779185935854912, + 0.04673900455236435, + 0.05818149447441101, + 0.0513082779943943, + 0.044015128165483475, + 0.07272515445947647, + 0.04658893123269081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.07001103460788727 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797225706279278 + ], + "opposite_logprob_tokens": [ + -4.634797096252441 + ], + "logprob_reward": -0.009797225706279278, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797225706279278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1744881272315979, + 0.07432623952627182, + 0.09879133105278015, + 0.04986200109124184, + 0.07385676354169846, + 0.10878787189722061, + 0.07681777328252792, + 0.06341557204723358, + 0.10088999569416046, + 0.17876429855823517 + ], + "credit_map": [ + 0.017286116257309914, + 0.02314036153256893, + 0.02500607632100582, + 0.025693483650684357, + 0.021032650023698807, + 0.022334879264235497, + 0.01754555106163025, + 0.022449029609560966, + 0.00974506139755249, + 0.008800809271633625, + 0.012209899723529816, + 0.010680163279175758, + 0.005203706212341785, + 0.008746445178985596, + 0.009019463323056698, + 0.009920697659254074, + 0.009491391479969025, + 0.008647297509014606, + 0.02268495224416256, + 0.009157630614936352, + 0.009288998320698738, + 0.015356487594544888, + 0.010070722550153732, + 0.01409385446459055, + 0.006705761421471834, + 0.004120053257793188, + 0.006484239362180233, + 0.006324229761958122, + 0.004000071436166763, + 0.01221390813589096, + 0.00423465995118022, + 0.0057790824212133884, + 0.0056406715884804726, + 0.006222925614565611, + 0.007909275591373444, + 0.008989956229925156, + 0.017270708456635475, + 0.01027423795312643, + 0.00924120657145977, + 0.008307788521051407, + 0.01467579510062933, + 0.009927290491759777, + 0.013590890914201736, + 0.01790444366633892, + 0.01455788966268301, + 0.014738243073225021, + 0.012695341370999813, + 0.010697990655899048, + 0.009398965165019035, + 0.007129705976694822, + 0.011902978643774986, + 0.009931308217346668, + 0.007835442200303078, + 0.00887353252619505, + 0.014490796253085136, + 0.007255051750689745, + 0.010375977493822575, + 0.00632597366347909, + 0.005063159391283989, + 0.00730764539912343, + 0.0062683699652552605, + 0.00879696849733591, + 0.006305740214884281, + 0.012971743009984493, + 0.008282314985990524, + 0.009591612964868546, + 0.01206883229315281, + 0.01311475969851017, + 0.017125986516475677, + 0.016178186982870102, + 0.011736043728888035, + 0.012792264111340046, + 0.019191185012459755, + 0.017712349072098732, + 0.04045579582452774, + 0.021032245829701424, + 0.022377822548151016, + 0.016803111881017685, + 0.0264959167689085, + 0.014695881865918636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.009797225706279278 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, first-person handheld camera shot oscillates forward through a vast, abstract digital space constructed in a low-poly style. At the center of the image is a vast information flow representing the Sina portal website. A majestic cuboid, composed of countless tiny, low-polygonal facets gleaming in the signature Sina red and dotted with blue and white dots, it hovers in the pitch-black void, emitting an intense, steady digital glow that illuminates the floating, fragmented geometric structures around it. An invisible force slowly begins to warp this massive object, rippling and cracking its surface. Finally, in a silent yet visually striking slow-motion explosion, it shatters, sending vast red polygonal fragments and countless smaller dots of light scattering violently in all directions. As these fragments tumble and fly in the chaotic chaos, they miraculously transform and reshape, transforming into countless luminous, three-dimensional Chinese news headlines that hurtle toward every corner of the space at varying speeds and trajectories. The handheld camera dangerously shuttles, sways and tilts between these rapidly passing data streams and news headlines, striving to capture the dynamic moment of this information explosion. Light and shadow change rapidly in front of the lens, creating a strong visual atmosphere that is both chaotic and spectacular, about information overload and deconstruction.", + "seed": 9325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_1605_seed_9325_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": -0.04939407501369715, + "vlm_logprob": -0.04939407501369715, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04939407501369715, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06768443435430527 + ], + "opposite_logprob_tokens": [ + -3.3176844120025635 + ], + "logprob_reward": -0.06768443435430527, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06768443435430527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19326989352703094, + 0.06542957574129105, + 0.05647045373916626, + 0.09904661029577255, + 0.24025820195674896, + 0.06880754232406616, + 0.05556182935833931, + 0.10243524610996246, + 0.0645899772644043, + 0.05413064733147621 + ], + "credit_map": [ + 0.015871822834014893, + 0.016544265672564507, + 0.023339472711086273, + 0.05206454545259476, + 0.019355403259396553, + 0.016151851043105125, + 0.03680974990129471, + 0.013132779859006405, + 0.005276842042803764, + 0.010954630561172962, + 0.0058451308868825436, + 0.01321479119360447, + 0.008073622360825539, + 0.005456691607832909, + 0.011294697411358356, + 0.005313175730407238, + 0.008430812507867813, + 0.008131113834679127, + 0.006844434887170792, + 0.006498062051832676, + 0.0057452828623354435, + 0.00479630334302783, + 0.01017365325242281, + 0.005850792862474918, + 0.011153210885822773, + 0.007847856730222702, + 0.016139844432473183, + 0.013884477317333221, + 0.010325422510504723, + 0.012038559652864933, + 0.010752839967608452, + 0.016904402524232864, + 0.02288704738020897, + 0.023295927792787552, + 0.026121601462364197, + 0.03642984479665756, + 0.034256722778081894, + 0.05486416444182396, + 0.025999369099736214, + 0.01640351116657257, + 0.006463005673140287, + 0.00650138920173049, + 0.007757124025374651, + 0.006591327488422394, + 0.02017669938504696, + 0.006548234727233648, + 0.0075486223213374615, + 0.007221135310828686, + 0.006620906759053469, + 0.005765073467046022, + 0.0057615903206169605, + 0.005995533429086208, + 0.0060425447300076485, + 0.01110292598605156, + 0.008090554736554623, + 0.006182701326906681, + 0.01036770548671484, + 0.010857791639864445, + 0.01618064194917679, + 0.01205609180033207, + 0.006357600912451744, + 0.007812353782355785, + 0.016614854335784912, + 0.022188197821378708, + 0.004446488805115223, + 0.007059056311845779, + 0.01575412228703499, + 0.010913250036537647, + 0.0034230914898216724, + 0.009793172590434551, + 0.007848748937249184, + 0.0053520482033491135, + 0.006000456400215626, + 0.005197762046009302, + 0.005383600480854511, + 0.007700424175709486, + 0.006151209585368633, + 0.00592866912484169, + 0.00778032885864377, + 0.00998819898813963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.06768443435430527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381375268101692 + ], + "opposite_logprob_tokens": [ + -3.651381492614746 + ], + "logprob_reward": -0.026381375268101692, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381375268101692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23784884810447693, + 0.0989200621843338, + 0.09071417897939682, + 0.10577908903360367, + 0.11720157414674759, + 0.057157181203365326, + 0.0652894377708435, + 0.07425528019666672, + 0.08759241551160812, + 0.06524188071489334 + ], + "credit_map": [ + 0.07323235273361206, + 0.01310345996171236, + 0.08141916990280151, + 0.015042992308735847, + 0.013952309265732765, + 0.0134187126532197, + 0.013846859335899353, + 0.013832977041602135, + 0.007153763435781002, + 0.011863191612064838, + 0.01131333876401186, + 0.016350550577044487, + 0.010184712707996368, + 0.00806642696261406, + 0.020478565245866776, + 0.013509519398212433, + 0.006791253574192524, + 0.00955161638557911, + 0.01368272490799427, + 0.010697977617383003, + 0.010100086219608784, + 0.009668760001659393, + 0.01783515326678753, + 0.012386604212224483, + 0.012147177942097187, + 0.00819574948400259, + 0.014986478723585606, + 0.01601467654109001, + 0.014700669795274734, + 0.01141950860619545, + 0.013981278985738754, + 0.014333546161651611, + 0.01230650581419468, + 0.01093241572380066, + 0.015614663250744343, + 0.012422796338796616, + 0.01601492613554001, + 0.024660220369696617, + 0.010006253607571125, + 0.015243800356984138, + 0.006703003775328398, + 0.005623298231512308, + 0.005309985484927893, + 0.004772912245243788, + 0.014454493299126625, + 0.006696943659335375, + 0.006986449472606182, + 0.006610094103962183, + 0.006978517863899469, + 0.005566321313381195, + 0.0075221010483801365, + 0.012152439914643764, + 0.006861139088869095, + 0.005332768429070711, + 0.012159177102148533, + 0.008716976270079613, + 0.009198851883411407, + 0.00640446599572897, + 0.011203193105757236, + 0.009622560814023018, + 0.004297673236578703, + 0.004955685697495937, + 0.01232246682047844, + 0.016250377520918846, + 0.006086539011448622, + 0.00883703026920557, + 0.026367420330643654, + 0.016178853809833527, + 0.004769712220877409, + 0.013413392007350922, + 0.00574460020288825, + 0.006194875575602055, + 0.0045015523210167885, + 0.0071195876225829124, + 0.007341137155890465, + 0.010795710608363152, + 0.005027191713452339, + 0.007292621303349733, + 0.012083004228770733, + 0.011081075295805931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.026381375268101692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259131081402302 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259131081402302, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259131081402302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15644782781600952, + 0.05719010531902313, + 0.24689029157161713, + 0.09883558005094528, + 0.11089174449443817, + 0.04205641895532608, + 0.03699203580617905, + 0.11168095469474792, + 0.0756763443350792, + 0.06333870440721512 + ], + "credit_map": [ + 0.035361506044864655, + 0.010285238735377789, + 0.04750281199812889, + 0.013550921343266964, + 0.00852003414183855, + 0.010514352470636368, + 0.014914819039404392, + 0.015798142179846764, + 0.004185978323221207, + 0.008635886013507843, + 0.007341067772358656, + 0.007524814456701279, + 0.005454356782138348, + 0.005298533942550421, + 0.010146895423531532, + 0.008602566085755825, + 0.01827586442232132, + 0.028992874547839165, + 0.022944120690226555, + 0.03474820405244827, + 0.030089735984802246, + 0.0181697029620409, + 0.0711742490530014, + 0.02249552123248577, + 0.012808908708393574, + 0.008254834450781345, + 0.013683476485311985, + 0.017466312274336815, + 0.010776606388390064, + 0.009996440261602402, + 0.014187777414917946, + 0.011661214753985405, + 0.010881501249969006, + 0.013983584940433502, + 0.01418590359389782, + 0.01645566336810589, + 0.017300862818956375, + 0.013938987627625465, + 0.011644049547612667, + 0.012501192279160023, + 0.005120397079735994, + 0.004196216352283955, + 0.004810504615306854, + 0.00427427189424634, + 0.006434094160795212, + 0.004798268433660269, + 0.007260294631123543, + 0.005162369459867477, + 0.004490092862397432, + 0.0046673850156366825, + 0.005351123865693808, + 0.00409806938841939, + 0.004648098722100258, + 0.003918822854757309, + 0.004930575378239155, + 0.004887866321951151, + 0.008889122866094112, + 0.007935111410915852, + 0.019933242350816727, + 0.017222555354237556, + 0.00677459454163909, + 0.00903092697262764, + 0.020755788311362267, + 0.0211396012455225, + 0.0054036034271121025, + 0.007003441918641329, + 0.025231413543224335, + 0.012797989882528782, + 0.0037444208282977343, + 0.009891695342957973, + 0.006018744315952063, + 0.005585023667663336, + 0.004268383141607046, + 0.006369868293404579, + 0.007018616423010826, + 0.011066591367125511, + 0.0053915949538350105, + 0.007079242262989283, + 0.011881323531270027, + 0.01026307512074709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.005259131081402302 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.127041757106781 + ], + "opposite_logprob_tokens": [ + -2.127041816711426 + ], + "logprob_reward": -0.127041757106781, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.127041757106781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16163717210292816, + 0.0514475479722023, + 0.063743457198143, + 0.10424993187189102, + 0.24270950257778168, + 0.17423617839813232, + 0.04701319336891174, + 0.059234291315078735, + 0.062052566558122635, + 0.03367616981267929 + ], + "credit_map": [ + 0.029404321685433388, + 0.010521741583943367, + 0.025563919916749, + 0.04100201278924942, + 0.013497349806129932, + 0.007686709053814411, + 0.02677849866449833, + 0.007182602770626545, + 0.003187253838405013, + 0.00909943412989378, + 0.004400631878525019, + 0.009268105961382389, + 0.007956990040838718, + 0.004044283647090197, + 0.008848272264003754, + 0.004642575513571501, + 0.006836462300270796, + 0.008530338294804096, + 0.008957074955105782, + 0.0072336806915700436, + 0.007519503589719534, + 0.00451886560767889, + 0.01161441020667553, + 0.008533119224011898, + 0.008195247501134872, + 0.007214349694550037, + 0.016328660771250725, + 0.012235802598297596, + 0.008511585183441639, + 0.016313116997480392, + 0.009941832162439823, + 0.025509340688586235, + 0.029150763526558876, + 0.022038336843252182, + 0.03727981820702553, + 0.024769343435764313, + 0.029093950986862183, + 0.04378172382712364, + 0.0191381573677063, + 0.037457410246133804, + 0.02773738093674183, + 0.019388360902667046, + 0.0179903544485569, + 0.01422329805791378, + 0.024307455867528915, + 0.024937890470027924, + 0.025026019662618637, + 0.020625414326786995, + 0.003493479685857892, + 0.0041960556991398335, + 0.003931795246899128, + 0.006342669017612934, + 0.004156764131039381, + 0.006627743132412434, + 0.011391942389309406, + 0.006872743368148804, + 0.005271981470286846, + 0.005051382817327976, + 0.010445604100823402, + 0.008689279668033123, + 0.0032796680461615324, + 0.004527311306446791, + 0.01248138677328825, + 0.00948767364025116, + 0.004207759629935026, + 0.006836692802608013, + 0.010707941837608814, + 0.008541710674762726, + 0.004579547792673111, + 0.01156547386199236, + 0.0069926124997437, + 0.00862082839012146, + 0.00280621275305748, + 0.003930125385522842, + 0.003922495525330305, + 0.004640669096261263, + 0.002932856557890773, + 0.0038403652142733335, + 0.0061713517643511295, + 0.005432092119008303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.127041757106781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060367725789547 + ], + "opposite_logprob_tokens": [ + -3.895603656768799 + ], + "logprob_reward": -0.02060367725789547, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060367725789547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11623646318912506, + 0.048200976103544235, + 0.10852331668138504, + 0.12105849385261536, + 0.20677849650382996, + 0.06267014890909195, + 0.0907839983701706, + 0.1212267056107521, + 0.05706454813480377, + 0.06745684146881104 + ], + "credit_map": [ + 0.01782836765050888, + 0.012210317887365818, + 0.021117188036441803, + 0.020225433632731438, + 0.008361240848898888, + 0.012590154074132442, + 0.013600335456430912, + 0.01030342373996973, + 0.003637206507846713, + 0.006720948498696089, + 0.003952563740313053, + 0.009524766355752945, + 0.004729775711894035, + 0.004601136781275272, + 0.010514148510992527, + 0.0045204306952655315, + 0.009639010764658451, + 0.012176452204585075, + 0.013583759777247906, + 0.013233006931841373, + 0.01233445294201374, + 0.009539397433400154, + 0.02412988990545273, + 0.01388735044747591, + 0.009696116670966148, + 0.00995490700006485, + 0.018926333636045456, + 0.0162329263985157, + 0.01270910631865263, + 0.01614266261458397, + 0.013335668481886387, + 0.024060767143964767, + 0.02146066352725029, + 0.02048565074801445, + 0.020168911665678024, + 0.026340287178754807, + 0.03551265224814415, + 0.04416881501674652, + 0.020696744322776794, + 0.017944760620594025, + 0.008968960493803024, + 0.0066213724203407764, + 0.008029436692595482, + 0.006410479545593262, + 0.008288764394819736, + 0.008781008422374725, + 0.007228196132928133, + 0.008341933600604534, + 0.009609106928110123, + 0.00809989683330059, + 0.01088577602058649, + 0.010577283799648285, + 0.010637897998094559, + 0.01510635670274496, + 0.015535988844931126, + 0.010331694036722183, + 0.012342415750026703, + 0.00901422742754221, + 0.018440933898091316, + 0.017315998673439026, + 0.007916978560388088, + 0.009010148234665394, + 0.020464900881052017, + 0.02672109752893448, + 0.004825239535421133, + 0.008788127452135086, + 0.011308287270367146, + 0.006671037059277296, + 0.004051093943417072, + 0.008656000718474388, + 0.006890196818858385, + 0.00587456626817584, + 0.005407902877777815, + 0.006603359244763851, + 0.008733042515814304, + 0.010504676960408688, + 0.005933154374361038, + 0.008371831849217415, + 0.010642703622579575, + 0.011260171420872211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.02060367725789547 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, first-person handheld camera shot oscillates forward through a vast, abstract digital space constructed in a low-poly style. At the center of the image is a vast information flow representing the Sina portal website. A majestic cuboid, composed of countless tiny, low-polygonal facets gleaming in the signature Sina red and dotted with blue and white dots, it hovers in the pitch-black void, emitting an intense, steady digital glow that illuminates the floating, fragmented geometric structures around it. An invisible force slowly begins to warp this massive object, rippling and cracking its surface. Finally, in a silent yet visually striking slow-motion explosion, it shatters, sending vast red polygonal fragments and countless smaller dots of light scattering violently in all directions. As these fragments tumble and fly in the chaotic chaos, they miraculously transform and reshape, transforming into countless luminous, three-dimensional Chinese news headlines that hurtle toward every corner of the space at varying speeds and trajectories. The handheld camera dangerously shuttles, sways and tilts between these rapidly passing data streams and news headlines, striving to capture the dynamic moment of this information explosion. Light and shadow change rapidly in front of the lens, creating a strong visual atmosphere that is both chaotic and spectacular, about information overload and deconstruction.", + "seed": 9326, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_12/index_1605_seed_9326_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": -0.052467736694961783, + "vlm_logprob": -0.052467736694961783, + "vlm_yes_no_margin": 3.874999976158142, + "vlm_token_credit": -0.052467736694961783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10779090970754623 + ], + "opposite_logprob_tokens": [ + -2.607790946960449 + ], + "logprob_reward": -0.10779090970754623, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.10779090970754623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1319759488105774, + 0.14424221217632294, + 0.15834593772888184, + 0.03586664795875549, + 0.033274095505476, + 0.020035764202475548, + 0.2671736180782318, + 0.09802000224590302, + 0.07185612618923187, + 0.039209648966789246 + ], + "credit_map": [ + 0.018380243331193924, + 0.010783939622342587, + 0.01610121689736843, + 0.014895963482558727, + 0.01274503767490387, + 0.016604356467723846, + 0.015997225418686867, + 0.02646796405315399, + 0.01862078532576561, + 0.010174859315156937, + 0.020187098532915115, + 0.013423156924545765, + 0.007620407734066248, + 0.020614909008145332, + 0.012443086132407188, + 0.041157905012369156, + 0.015303467400372028, + 0.009964851662516594, + 0.016522124409675598, + 0.010585264302790165, + 0.006439919117838144, + 0.0475710965692997, + 0.009270335547626019, + 0.04268887639045715, + 0.004029070492833853, + 0.00600834097713232, + 0.002977091586217284, + 0.0035942080430686474, + 0.0032962847035378218, + 0.003851821878924966, + 0.007363707758486271, + 0.0047461241483688354, + 0.003504622494801879, + 0.004432470537722111, + 0.006577960215508938, + 0.004311091732233763, + 0.00525958277285099, + 0.0026727053336799145, + 0.0037066612858325243, + 0.0028090011328458786, + 0.003014139598235488, + 0.0025218315422534943, + 0.0023611015640199184, + 0.002930903807282448, + 0.0023804442025721073, + 0.0017020776867866516, + 0.0034318838734179735, + 0.0016933823935687542, + 0.021015582606196404, + 0.07973777502775192, + 0.08991335332393646, + 0.014436704106628895, + 0.011280758306384087, + 0.01602873206138611, + 0.0172507893294096, + 0.01750992424786091, + 0.009824574925005436, + 0.014936464838683605, + 0.013012424111366272, + 0.016338184475898743, + 0.011355040594935417, + 0.014686495997011662, + 0.011172868311405182, + 0.006693946197628975, + 0.009122799150645733, + 0.010741646401584148, + 0.006508269812911749, + 0.00576405692845583, + 0.01223029289394617, + 0.01260995864868164, + 0.006173932459205389, + 0.008705166168510914, + 0.0053231120109558105, + 0.005741671659052372, + 0.002562248846516013, + 0.004310016054660082, + 0.0039039799012243748, + 0.007960586808621883, + 0.0060505042783916, + 0.003357528243213892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.10779090970754623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011099381372332573 + ], + "opposite_logprob_tokens": [ + -4.511099338531494 + ], + "logprob_reward": -0.011099381372332573, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011099381372332573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12331777065992355, + 0.07451397180557251, + 0.12019890546798706, + 0.03807588294148445, + 0.1298905909061432, + 0.024112679064273834, + 0.1148107573390007, + 0.22153128683567047, + 0.10802832245826721, + 0.045519888401031494 + ], + "credit_map": [ + 0.013876231387257576, + 0.016548406332731247, + 0.015688758343458176, + 0.011433425359427929, + 0.012066418305039406, + 0.017794419080018997, + 0.01146097481250763, + 0.02444913052022457, + 0.010390006005764008, + 0.005582097917795181, + 0.01136701088398695, + 0.006716317962855101, + 0.005387394223362207, + 0.011065511032938957, + 0.00784054771065712, + 0.01616508886218071, + 0.01279566902667284, + 0.009234207682311535, + 0.0213958490639925, + 0.010382882319390774, + 0.0067102136090397835, + 0.02405240200459957, + 0.009614250622689724, + 0.02601342462003231, + 0.00467164721339941, + 0.004516209941357374, + 0.006640071514993906, + 0.004488492384552956, + 0.002964129438623786, + 0.004761230666190386, + 0.004906584043055773, + 0.005127519369125366, + 0.011304433457553387, + 0.015819653868675232, + 0.01936446875333786, + 0.02045176737010479, + 0.021951885893940926, + 0.013193908147513866, + 0.01563085801899433, + 0.01217362005263567, + 0.004284949507564306, + 0.0030308172572404146, + 0.002988436259329319, + 0.003018125658854842, + 0.0022431733086705208, + 0.002854137448593974, + 0.002591375494375825, + 0.003101664362475276, + 0.007562049198895693, + 0.025866473093628883, + 0.03294967859983444, + 0.006972427945584059, + 0.009122012183070183, + 0.009602474048733711, + 0.01601181924343109, + 0.006723820697516203, + 0.024396205320954323, + 0.03758697584271431, + 0.03251565247774124, + 0.04488570615649223, + 0.024155136197805405, + 0.020109226927161217, + 0.025056390091776848, + 0.012825977988541126, + 0.008999453857541084, + 0.019913937896490097, + 0.010968850925564766, + 0.012697771191596985, + 0.017745021730661392, + 0.013972798362374306, + 0.012021680362522602, + 0.01170880626887083, + 0.0056630768813192844, + 0.00609925389289856, + 0.003291517496109009, + 0.005260334350168705, + 0.006118530873209238, + 0.007724560331553221, + 0.0071005611680448055, + 0.00426205201074481 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.011099381372332573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200921531766653 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002200921531766653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200921531766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08300434052944183, + 0.10550583899021149, + 0.12631601095199585, + 0.031919367611408234, + 0.04805929958820343, + 0.024015624076128006, + 0.21543966233730316, + 0.2155841588973999, + 0.10813254863023758, + 0.042023222893476486 + ], + "credit_map": [ + 0.01147620938718319, + 0.007914437912404537, + 0.012421644292771816, + 0.008327383548021317, + 0.008661960251629353, + 0.012352466583251953, + 0.010204355232417583, + 0.011645876802504063, + 0.014971651136875153, + 0.009418176487088203, + 0.016894107684493065, + 0.014548963867127895, + 0.009211420081555843, + 0.015047488734126091, + 0.010716586373746395, + 0.01469743438065052, + 0.015051360242068768, + 0.010563842952251434, + 0.022380396723747253, + 0.011612476781010628, + 0.006096294149756432, + 0.026157675310969353, + 0.009512515738606453, + 0.024941429495811462, + 0.0036660954356193542, + 0.005010869819670916, + 0.00245979824103415, + 0.002955606672912836, + 0.003128690179437399, + 0.004217324312776327, + 0.006235247943550348, + 0.004245732445269823, + 0.004173703026026487, + 0.005798054859042168, + 0.006994567811489105, + 0.006268068682402372, + 0.008638269267976284, + 0.005358334165066481, + 0.005857526324689388, + 0.004970774054527283, + 0.003272454487159848, + 0.0024948862846940756, + 0.002965911291539669, + 0.0032371648121625185, + 0.0028150719590485096, + 0.002443630015477538, + 0.0043862359598279, + 0.002400267869234085, + 0.0171397402882576, + 0.0631164088845253, + 0.06838660687208176, + 0.013581018894910812, + 0.01222634594887495, + 0.015485010109841824, + 0.014061043038964272, + 0.011443473398685455, + 0.02525637298822403, + 0.043548133224248886, + 0.03306451439857483, + 0.036828070878982544, + 0.02480914443731308, + 0.017597325146198273, + 0.02157372049987316, + 0.012906862422823906, + 0.009971805848181248, + 0.01577381230890751, + 0.010808403603732586, + 0.007453407626599073, + 0.024617716670036316, + 0.014571018517017365, + 0.011812862940132618, + 0.01312351506203413, + 0.004723740741610527, + 0.005525536369532347, + 0.00296621210873127, + 0.005991815589368343, + 0.004083819221705198, + 0.005432272795587778, + 0.008756788447499275, + 0.004543033894151449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.002200921531766653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703430652618408 + ], + "opposite_logprob_tokens": [ + -2.1270341873168945 + ], + "logprob_reward": -0.12703430652618408, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703430652618408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11380726844072342, + 0.03370087221264839, + 0.15532360970973969, + 0.02840036153793335, + 0.017136383801698685, + 0.016976868733763695, + 0.2023327648639679, + 0.2699180841445923, + 0.130754753947258, + 0.031649019569158554 + ], + "credit_map": [ + 0.016286000609397888, + 0.011632386595010757, + 0.010890752077102661, + 0.010640674270689487, + 0.01833266019821167, + 0.014837745577096939, + 0.014550426043570042, + 0.016636615619063377, + 0.005092752631753683, + 0.00287340534850955, + 0.004097917582839727, + 0.002586851129308343, + 0.002340964274480939, + 0.004556660540401936, + 0.004198677372187376, + 0.0079536447301507, + 0.009322536177933216, + 0.004644112195819616, + 0.006630852818489075, + 0.004669074434787035, + 0.004248251672834158, + 0.06203870102763176, + 0.004770177416503429, + 0.05899989604949951, + 0.003111964324489236, + 0.005846185609698296, + 0.003431487362831831, + 0.0021403662394732237, + 0.0028256485238671303, + 0.003615545341745019, + 0.003808412468060851, + 0.003620751202106476, + 0.0023352866992354393, + 0.002272621961310506, + 0.0037312472704797983, + 0.0016135808546096087, + 0.0016818721778690815, + 0.001980840926989913, + 0.001570491585880518, + 0.0019504440715536475, + 0.0022674291394650936, + 0.0019166418351233006, + 0.00191802519839257, + 0.0022047972306609154, + 0.002256185980513692, + 0.0016459198668599129, + 0.003348728409036994, + 0.0014191408408805728, + 0.012602345086634159, + 0.02547839842736721, + 0.024311605840921402, + 0.015352264046669006, + 0.022084707394242287, + 0.02239084802567959, + 0.06219444423913956, + 0.017918134108185768, + 0.029202373698353767, + 0.044554658234119415, + 0.03561164811253548, + 0.05431436747312546, + 0.02925153076648712, + 0.027626793831586838, + 0.03208402916789055, + 0.017272673547267914, + 0.010967956855893135, + 0.025719448924064636, + 0.010043079033493996, + 0.008901288732886314, + 0.029313717037439346, + 0.017699504271149635, + 0.01892879791557789, + 0.009180948138237, + 0.00353871937841177, + 0.003591969609260559, + 0.0020327342208474874, + 0.004984273109585047, + 0.003405573545023799, + 0.004446184728294611, + 0.0063973236829042435, + 0.003252238966524601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.12703430652618408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot first-person and handheld, oscillating forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a vast cuboid composed of low-polygonal facets in Sina red with blue and white dots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cuboid warp, ripple, and crack before shattering into red polygonal fragments and dots of light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered fragments transform into three-dimensional Chinese news headlines that hurtle in all directions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the handheld camera sways, tilt, and shuttle between the data streams and headlines during the explosion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421316433697939 + ], + "opposite_logprob_tokens": [ + -4.2642130851745605 + ], + "logprob_reward": -0.01421316433697939, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421316433697939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13721396028995514, + 0.11802685260772705, + 0.06367971748113632, + 0.07597403973340988, + 0.06666619330644608, + 0.022539803758263588, + 0.05862155184149742, + 0.2661837339401245, + 0.1342090517282486, + 0.05688513070344925 + ], + "credit_map": [ + 0.017582343891263008, + 0.011097349226474762, + 0.018803486600518227, + 0.01435950305312872, + 0.012784077785909176, + 0.018478991463780403, + 0.01634373888373375, + 0.027764471247792244, + 0.014319658279418945, + 0.009804747998714447, + 0.022086581215262413, + 0.016249656677246094, + 0.010934324003756046, + 0.014526914805173874, + 0.011068298481404781, + 0.01903666742146015, + 0.0059271203354001045, + 0.005389814730733633, + 0.010300925932824612, + 0.006147285923361778, + 0.004111775662750006, + 0.014380304142832756, + 0.004352071322500706, + 0.013070415705442429, + 0.006681741215288639, + 0.013179085217416286, + 0.006325928028672934, + 0.007333000190556049, + 0.008819001726806164, + 0.00576668931171298, + 0.018256841227412224, + 0.009611749090254307, + 0.006833544000983238, + 0.008337565697729588, + 0.006614426616579294, + 0.014806886203587055, + 0.009496482089161873, + 0.005939557682722807, + 0.007374915760010481, + 0.0072628166526556015, + 0.00318168755620718, + 0.0025495553854852915, + 0.0026172443758696318, + 0.0028480556793510914, + 0.002759638475254178, + 0.0021536420099437237, + 0.004057273268699646, + 0.0023727091029286385, + 0.005512709263712168, + 0.008851788938045502, + 0.012680023908615112, + 0.004335835110396147, + 0.006664953660219908, + 0.00736842630431056, + 0.008138521574437618, + 0.005069292616099119, + 0.028464823961257935, + 0.03750183433294296, + 0.03862893208861351, + 0.05919680744409561, + 0.028137708082795143, + 0.025449734181165695, + 0.03348065912723541, + 0.015323244035243988, + 0.010263233445584774, + 0.017379939556121826, + 0.013517295010387897, + 0.011614296585321426, + 0.028254453092813492, + 0.01600894331932068, + 0.02167455293238163, + 0.015496339648962021, + 0.006176883354783058, + 0.006349810399115086, + 0.0033955592662096024, + 0.00959648285061121, + 0.005239552818238735, + 0.009378262795507908, + 0.009887691587209702, + 0.006860889494419098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01421316433697939 + } + ], + "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/0000152/group_13/index_tensor([3755, 3755])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_tensor([3755, 3755])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b489f07b5ab5e1f1ad041b84efd73875b873b483 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_tensor([3755, 3755])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up shot from a slightly overhead angle, framed in minimalist black and white, captures a male athlete in excruciating pain. A harsh, slanting light carves out his tense facial features. Against the high-contrast backdrop, his eyes are tightly closed, his brow knitted in agony. Between the deep wrinkles of a rickety crease, large beads of sweat form, trickling down his jawline, accentuated by his clenched teeth. His head tosses and turns slowly and painfully against a deep, almost indecipherable black background. Each twist is a silent struggle, as if attempting to escape the inescapable pain within him. His short, damp hair clings to his forehead and temples. The entire scene is shrouded in a primal and oppressive atmosphere. The intense visual impact focuses the viewer's attention on the sheer physical suffering, creating a suffocating sense of tension.", + "seed": 5369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5369_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": -0.03416112884879112, + "vlm_logprob": -0.03416112884879112, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.03416112884879112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07196445763111115 + ], + "opposite_logprob_tokens": [ + -2.6969645023345947 + ], + "logprob_reward": -0.07196445763111115, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07196445763111115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09461995214223862, + 0.06729380786418915, + 0.07467491179704666, + 0.03532055765390396, + 0.09964194148778915, + 0.12878507375717163, + 0.3194134831428528, + 0.07945988327264786, + 0.0629739910364151, + 0.03781641274690628 + ], + "credit_map": [ + 0.012977797538042068, + 0.01181096863001585, + 0.010839407332241535, + 0.009691283106803894, + 0.00893221702426672, + 0.013094238005578518, + 0.01261257752776146, + 0.014661460183560848, + 0.012357578612864017, + 0.00684781838208437, + 0.00556951342150569, + 0.006580569315701723, + 0.005320629570633173, + 0.007715543266385794, + 0.012683128006756306, + 0.010219032876193523, + 0.009307934902608395, + 0.00929480604827404, + 0.008858623914420605, + 0.011160585097968578, + 0.008059409447014332, + 0.006236139684915543, + 0.012050974182784557, + 0.009706441313028336, + 0.0032092202454805374, + 0.003419822547584772, + 0.0037439714651554823, + 0.003613931592553854, + 0.004610054194927216, + 0.005105002783238888, + 0.00683950912207365, + 0.004779048264026642, + 0.015311493538320065, + 0.008762290701270103, + 0.008530252613127232, + 0.010398863814771175, + 0.011541230604052544, + 0.010618717409670353, + 0.01642589271068573, + 0.018053200095891953, + 0.017237456515431404, + 0.017359711229801178, + 0.011078627780079842, + 0.01754138059914112, + 0.009451929479837418, + 0.016007719561457634, + 0.01648450829088688, + 0.023623740300536156, + 0.028147447854280472, + 0.048311978578567505, + 0.03776739165186882, + 0.039007507264614105, + 0.032087408006191254, + 0.0509839691221714, + 0.045588407665491104, + 0.03751937299966812, + 0.005928835831582546, + 0.009659718722105026, + 0.00922967679798603, + 0.016704915091395378, + 0.008422245271503925, + 0.00831121951341629, + 0.011040966957807541, + 0.010162306018173695, + 0.007215190213173628, + 0.005705951247364283, + 0.0062032099813222885, + 0.010421473532915115, + 0.006722388789057732, + 0.006863743998110294, + 0.009912681765854359, + 0.009929349645972252, + 0.0029451949521899223, + 0.003321325406432152, + 0.0042907907627522945, + 0.0054491744376719, + 0.005960663314908743, + 0.0031296631786972284, + 0.008353973738849163, + 0.004365627653896809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.07196445763111115 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899164408445358 + ], + "opposite_logprob_tokens": [ + -2.578991651535034 + ], + "logprob_reward": -0.07899164408445358, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899164408445358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14344248175621033, + 0.07893451303243637, + 0.1258191615343094, + 0.06855962425470352, + 0.13670869171619415, + 0.1909371018409729, + 0.060366325080394745, + 0.09201765060424805, + 0.05269896239042282, + 0.05051545798778534 + ], + "credit_map": [ + 0.01885637827217579, + 0.01948867365717888, + 0.017451761290431023, + 0.019917462021112442, + 0.013827524147927761, + 0.013521287590265274, + 0.018562180921435356, + 0.02181723155081272, + 0.012180128134787083, + 0.005533845629543066, + 0.011086074635386467, + 0.005454367026686668, + 0.006202862598001957, + 0.010663879103958607, + 0.016227474436163902, + 0.011585887521505356, + 0.015108020044863224, + 0.010536846704781055, + 0.014572457410395145, + 0.022140972316265106, + 0.011476440355181694, + 0.01390748005360365, + 0.020867720246315002, + 0.01720922626554966, + 0.007459804881364107, + 0.005491609685122967, + 0.010723622515797615, + 0.012609458528459072, + 0.006123338360339403, + 0.006385979242622852, + 0.013034745119512081, + 0.006731073372066021, + 0.017409296706318855, + 0.009236715734004974, + 0.010463421232998371, + 0.02479749545454979, + 0.012620430439710617, + 0.012755275703966618, + 0.024100499227643013, + 0.02532556839287281, + 0.024566294625401497, + 0.02641446143388748, + 0.020493099465966225, + 0.021381884813308716, + 0.018216926604509354, + 0.021950919181108475, + 0.03572076931595802, + 0.022192761301994324, + 0.006322756409645081, + 0.007988269440829754, + 0.008628526702523232, + 0.00472246203571558, + 0.006111993920058012, + 0.010314234532415867, + 0.008922964334487915, + 0.007355127017945051, + 0.012428579851984978, + 0.011443139985203743, + 0.010221540927886963, + 0.00931508094072342, + 0.0096620237454772, + 0.007882079109549522, + 0.01791263371706009, + 0.013152586296200752, + 0.004700632765889168, + 0.004654324613511562, + 0.004179087001830339, + 0.009239493869245052, + 0.004641485400497913, + 0.006321558263152838, + 0.010491321794688702, + 0.008471067063510418, + 0.004799260292202234, + 0.004628498572856188, + 0.005654295906424522, + 0.004261963069438934, + 0.00799529068171978, + 0.004879697225987911, + 0.011916610412299633, + 0.006379847414791584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07899164408445358 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080636642873287 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080636642873287, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080636642873287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07089109718799591, + 0.0514005608856678, + 0.6927527189254761, + 0.029404014348983765, + 0.01158306933939457, + 0.046356141567230225, + 0.022969799116253853, + 0.017166491597890854, + 0.026541072875261307, + 0.030934995040297508 + ], + "credit_map": [ + 0.010640526190400124, + 0.008482600562274456, + 0.013643015176057816, + 0.009503969922661781, + 0.005919987335801125, + 0.006446488201618195, + 0.008648576214909554, + 0.007605934049934149, + 0.009077630005776882, + 0.004465112928301096, + 0.005135362967848778, + 0.006309041753411293, + 0.005043004173785448, + 0.006351842544972897, + 0.007415505591779947, + 0.007603063248097897, + 0.07814347743988037, + 0.06122748553752899, + 0.08646925538778305, + 0.12945394217967987, + 0.05773197114467621, + 0.07309070974588394, + 0.10790388286113739, + 0.09873200953006744, + 0.003238912671804428, + 0.0022307750768959522, + 0.0035704581532627344, + 0.003179847961291671, + 0.003427832620218396, + 0.003742653178051114, + 0.005793514661490917, + 0.004220020025968552, + 0.0015170593978837132, + 0.00106808973941952, + 0.0014520954573526978, + 0.0013966733822599053, + 0.0015683333622291684, + 0.0012086991919204593, + 0.0016731134383007884, + 0.0016990050207823515, + 0.006589075084775686, + 0.00383131206035614, + 0.004421963356435299, + 0.005885131191462278, + 0.004164693411439657, + 0.005909102037549019, + 0.009376312606036663, + 0.0061785513535141945, + 0.0025701255071908236, + 0.002767221536487341, + 0.004120651166886091, + 0.002921857638284564, + 0.002114947885274887, + 0.00266588618978858, + 0.002696095034480095, + 0.0031130139250308275, + 0.0018540862947702408, + 0.002226162003353238, + 0.0018564197234809399, + 0.002527884440496564, + 0.0018416499951854348, + 0.00146502535790205, + 0.0029778899624943733, + 0.0024173730053007603, + 0.003324691206216812, + 0.0021059145219624043, + 0.002169686835259199, + 0.004928238689899445, + 0.003276379080489278, + 0.0022664947900921106, + 0.003980899695307016, + 0.004488769918680191, + 0.0026856721378862858, + 0.0026613418012857437, + 0.0037362403236329556, + 0.0019875457510352135, + 0.006452956236898899, + 0.003076522843912244, + 0.006127963308244944, + 0.0042067538015544415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.011080636642873287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.07264228910207748, + 0.04485951364040375, + 0.3057502210140228, + 0.0269547700881958, + 0.04795970022678375, + 0.16837488114833832, + 0.1454382985830307, + 0.06326854228973389, + 0.04398033022880554, + 0.08077139407396317 + ], + "credit_map": [ + 0.0106737045571208, + 0.008343723602592945, + 0.011531544849276543, + 0.010124956257641315, + 0.005971049424260855, + 0.006627292837947607, + 0.010090848430991173, + 0.009279165416955948, + 0.007475343532860279, + 0.003569267690181732, + 0.0045113833621144295, + 0.008435617201030254, + 0.0033042358700186014, + 0.00480379443615675, + 0.006635244470089674, + 0.0061246249824762344, + 0.033801447600126266, + 0.02668415755033493, + 0.03643099591135979, + 0.05873870477080345, + 0.02723214030265808, + 0.03290575370192528, + 0.04440274089574814, + 0.04555428773164749, + 0.0032836557365953922, + 0.0019721356220543385, + 0.003595191054046154, + 0.0036694430746138096, + 0.0025671273469924927, + 0.0028917568270117044, + 0.005559586454182863, + 0.003415873507037759, + 0.006962412968277931, + 0.00470086419954896, + 0.0047263773158192635, + 0.0055801598355174065, + 0.005861560348421335, + 0.004359912592917681, + 0.00690070167183876, + 0.008867716416716576, + 0.02295885607600212, + 0.018251486122608185, + 0.021865926682949066, + 0.018391301855444908, + 0.013630940578877926, + 0.019273392856121063, + 0.028426870703697205, + 0.025576118379831314, + 0.01401236280798912, + 0.020933402702212334, + 0.017230112105607986, + 0.02124621905386448, + 0.014478919096291065, + 0.02171524427831173, + 0.01842738687992096, + 0.0173946600407362, + 0.005306438077241182, + 0.007874803617596626, + 0.007981346920132637, + 0.012816343456506729, + 0.006449973676353693, + 0.006145579274743795, + 0.008781831711530685, + 0.0079122269526124, + 0.004502980969846249, + 0.0036497251130640507, + 0.004515398293733597, + 0.006470248103141785, + 0.005829211324453354, + 0.004734056070446968, + 0.0073865363374352455, + 0.006892174016684294, + 0.00792490690946579, + 0.006679070182144642, + 0.008132733404636383, + 0.011786768212914467, + 0.008825979195535183, + 0.008362207561731339, + 0.01853923872113228, + 0.01052048895508051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.0028211576864123344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17757368087768555, + 0.03369481489062309, + 0.619877278804779, + 0.030404234305024147, + 0.00999966450035572, + 0.03717176988720894, + 0.03477662429213524, + 0.02770231105387211, + 0.012680535204708576, + 0.016119105741381645 + ], + "credit_map": [ + 0.027605045586824417, + 0.021106407046318054, + 0.03865760937333107, + 0.02135208062827587, + 0.011229411698877811, + 0.015596517361700535, + 0.01856922172009945, + 0.023457404226064682, + 0.0063095297664403915, + 0.0023107726592570543, + 0.004926762077957392, + 0.002519997302442789, + 0.0025158217176795006, + 0.0038777233567088842, + 0.005543900653719902, + 0.005690309684723616, + 0.07114017754793167, + 0.05405253916978836, + 0.06984175741672516, + 0.12121986597776413, + 0.051998913288116455, + 0.06381276249885559, + 0.09919410943984985, + 0.08861710131168365, + 0.00362807372584939, + 0.0026704638730734587, + 0.0037760641425848007, + 0.0035279658623039722, + 0.003373868763446808, + 0.00334740849211812, + 0.0058274297043681145, + 0.004252958111464977, + 0.0012086596107110381, + 0.0008718157769180834, + 0.000742098840419203, + 0.0014293042477220297, + 0.0013290856732055545, + 0.001044961391016841, + 0.001664578914642334, + 0.001709160627797246, + 0.004929448012262583, + 0.005551853682845831, + 0.003970578778535128, + 0.0050856382586061954, + 0.0026628493797034025, + 0.0038736877031624317, + 0.005331433843821287, + 0.005766280461102724, + 0.00315632252022624, + 0.005133128259330988, + 0.003427304793149233, + 0.004674955271184444, + 0.003802219172939658, + 0.0055273366160690784, + 0.004902241751551628, + 0.004153119865804911, + 0.0034882130566984415, + 0.0030013290233910084, + 0.0030158318113535643, + 0.004763959441334009, + 0.003185458015650511, + 0.0023639078717678785, + 0.004129077773541212, + 0.0037545354571193457, + 0.0014192498056218028, + 0.0010848136153072119, + 0.0010544247925281525, + 0.002052693860605359, + 0.001432976103387773, + 0.001290446030907333, + 0.0022413949482142925, + 0.002104535698890686, + 0.0019332461524754763, + 0.001459409249946475, + 0.002064429922029376, + 0.0019999018404632807, + 0.002292562508955598, + 0.001959855668246746, + 0.0025180133525282145, + 0.0018916875123977661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.005947748199105263 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle, framed in minimalist black and white, captures a male athlete in excruciating pain. A harsh, slanting light carves out his tense facial features. Against the high-contrast backdrop, his eyes are tightly closed, his brow knitted in agony. Between the deep wrinkles of a rickety crease, large beads of sweat form, trickling down his jawline, accentuated by his clenched teeth. His head tosses and turns slowly and painfully against a deep, almost indecipherable black background. Each twist is a silent struggle, as if attempting to escape the inescapable pain within him. His short, damp hair clings to his forehead and temples. The entire scene is shrouded in a primal and oppressive atmosphere. The intense visual impact focuses the viewer's attention on the sheer physical suffering, creating a suffocating sense of tension.", + "seed": 5370, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5370_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": -0.0392558351624757, + "vlm_logprob": -0.0392558351624757, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.0392558351624757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059030938893556595 + ], + "opposite_logprob_tokens": [ + -2.9340310096740723 + ], + "logprob_reward": -0.059030938893556595, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.059030938893556595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2162584364414215, + 0.08516747504472733, + 0.023321405053138733, + 0.08863391727209091, + 0.1011698916554451, + 0.09073983877897263, + 0.059220653027296066, + 0.046851325780153275, + 0.050866905599832535, + 0.2377701848745346 + ], + "credit_map": [ + 0.035249095410108566, + 0.01610729657113552, + 0.018241574987769127, + 0.01625421829521656, + 0.019854722544550896, + 0.035863131284713745, + 0.04771672934293747, + 0.026971669867634773, + 0.010263126343488693, + 0.013585729524493217, + 0.006256475113332272, + 0.008532579056918621, + 0.0074668219313025475, + 0.01270170509815216, + 0.007662151474505663, + 0.018698887899518013, + 0.0036164785269647837, + 0.0014782972866669297, + 0.0033425826113671064, + 0.0018465741304680705, + 0.0024694313760846853, + 0.003134689759463072, + 0.004951328504830599, + 0.0024820228572934866, + 0.013590551912784576, + 0.0075751785188913345, + 0.008009640499949455, + 0.007775154896080494, + 0.012216337025165558, + 0.008674778044223785, + 0.019550520926713943, + 0.011241755448281765, + 0.015218137763440609, + 0.012085186317563057, + 0.015116441063582897, + 0.018236832693219185, + 0.008262211456894875, + 0.007999507710337639, + 0.011632976122200489, + 0.012618587352335453, + 0.0077761998400092125, + 0.004577198997139931, + 0.007909853011369705, + 0.006363831460475922, + 0.013805445283651352, + 0.006665010005235672, + 0.030177641659975052, + 0.013464659452438354, + 0.005737765692174435, + 0.0062799532897770405, + 0.007495330646634102, + 0.006727930158376694, + 0.008427117951214314, + 0.008072258904576302, + 0.00879650004208088, + 0.007683794479817152, + 0.0029055660124868155, + 0.004751788917928934, + 0.0031482854392379522, + 0.009156559593975544, + 0.004614245146512985, + 0.012248996645212173, + 0.004614024888724089, + 0.005411859601736069, + 0.0065497346222400665, + 0.005001607816666365, + 0.0079335393384099, + 0.007905242964625359, + 0.005321254953742027, + 0.005951571743935347, + 0.00616583414375782, + 0.006038119085133076, + 0.01907734014093876, + 0.03661588206887245, + 0.02123899944126606, + 0.031110938638448715, + 0.043063290417194366, + 0.030791355296969414, + 0.02424708940088749, + 0.03162528574466705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.059030938893556595 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496591702103615 + ], + "opposite_logprob_tokens": [ + -2.9299659729003906 + ], + "logprob_reward": -0.05496591702103615, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496591702103615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1493396759033203, + 0.04574611037969589, + 0.06708335131406784, + 0.028325865045189857, + 0.04327486455440521, + 0.06219439581036568, + 0.050850387662649155, + 0.1278921514749527, + 0.3899306058883667, + 0.035362645983695984 + ], + "credit_map": [ + 0.022004052996635437, + 0.01602267660200596, + 0.02108129672706127, + 0.008942090906202793, + 0.011283449828624725, + 0.0146090192720294, + 0.02123033069074154, + 0.03416674956679344, + 0.006204541306942701, + 0.004283671732991934, + 0.008906812407076359, + 0.005352016072720289, + 0.004692487418651581, + 0.005439057480543852, + 0.005638363771140575, + 0.005229154601693153, + 0.01036845799535513, + 0.005271248519420624, + 0.008032893761992455, + 0.004904539790004492, + 0.009625784121453762, + 0.008756761439144611, + 0.010167459957301617, + 0.009956197813153267, + 0.004106818698346615, + 0.00268115964718163, + 0.003007186111062765, + 0.0030799664091318846, + 0.004294487182050943, + 0.0024211080744862556, + 0.006232471205294132, + 0.0025026611983776093, + 0.008537545800209045, + 0.005578021518886089, + 0.0049813007935881615, + 0.0035701096057891846, + 0.004930227063596249, + 0.003928298130631447, + 0.006535097490996122, + 0.005214256700128317, + 0.005702424794435501, + 0.0039022061973810196, + 0.005787509493529797, + 0.004651548806577921, + 0.008189219050109386, + 0.0037244963459670544, + 0.019642051309347153, + 0.01059492863714695, + 0.004311293363571167, + 0.005707623902708292, + 0.007848307490348816, + 0.005471528973430395, + 0.009019490331411362, + 0.006018114276230335, + 0.005734172649681568, + 0.006739852484315634, + 0.008274021558463573, + 0.01364395022392273, + 0.007492620497941971, + 0.014993821270763874, + 0.016660811379551888, + 0.03872707486152649, + 0.013684476725757122, + 0.014415355399250984, + 0.0402822308242321, + 0.03928384184837341, + 0.06412655860185623, + 0.04058925062417984, + 0.04570643976330757, + 0.05201268568634987, + 0.059315577149391174, + 0.04861399531364441, + 0.002519795438274741, + 0.005264344625174999, + 0.0026248565409332514, + 0.006093819160014391, + 0.007612956687808037, + 0.0025825335178524256, + 0.0029813675209879875, + 0.005682968068867922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.05496591702103615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002414762973785 + ], + "opposite_logprob_tokens": [ + -2.695024251937866 + ], + "logprob_reward": -0.07002414762973785, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002414762973785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.332144170999527, + 0.22640180587768555, + 0.02838749997317791, + 0.0428970530629158, + 0.022186987102031708, + 0.09665481001138687, + 0.059227362275123596, + 0.10534164309501648, + 0.018181808292865753, + 0.06857684999704361 + ], + "credit_map": [ + 0.06633426249027252, + 0.03165735676884651, + 0.02820259891450405, + 0.014624686911702156, + 0.030035339295864105, + 0.03516867011785507, + 0.09045366197824478, + 0.035667601972818375, + 0.028119666501879692, + 0.03537049517035484, + 0.019933314993977547, + 0.023998985067009926, + 0.017909636721014977, + 0.035437628626823425, + 0.02341824397444725, + 0.04221383482217789, + 0.004071544855833054, + 0.0019717433024197817, + 0.00502666225656867, + 0.0030849643517285585, + 0.003710393561050296, + 0.002611763309687376, + 0.004592802841216326, + 0.003317627590149641, + 0.006523642223328352, + 0.004698676522821188, + 0.006083378568291664, + 0.003654174506664276, + 0.00450492650270462, + 0.005144338123500347, + 0.007519632577896118, + 0.004768282640725374, + 0.0035968711599707603, + 0.002559406915679574, + 0.0025399180594831705, + 0.0023794565349817276, + 0.002259443048387766, + 0.002467984100803733, + 0.002467678626999259, + 0.0039162300527095795, + 0.010481772013008595, + 0.006549892481416464, + 0.010199062526226044, + 0.011087224818766117, + 0.01965409144759178, + 0.007630566600710154, + 0.019147226586937904, + 0.01190498098731041, + 0.00610806280747056, + 0.006511980202049017, + 0.008090720511972904, + 0.006072348915040493, + 0.009440171532332897, + 0.0071051474660634995, + 0.008213710971176624, + 0.007685224525630474, + 0.008636250160634518, + 0.012292030267417431, + 0.016356168314814568, + 0.013983698561787605, + 0.012558439746499062, + 0.017031433060765266, + 0.010632134042680264, + 0.01385149173438549, + 0.002164871199056506, + 0.0016906959936022758, + 0.002778111258521676, + 0.002654923126101494, + 0.003108274657279253, + 0.0013413183623924851, + 0.002105840714648366, + 0.0023377742618322372, + 0.009248785674571991, + 0.010118279606103897, + 0.008989480324089527, + 0.010213869623839855, + 0.00867677666246891, + 0.005882411729544401, + 0.0061513506807386875, + 0.009295904077589512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07002414762973785 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036178866866976023 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036178866866976023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036178866866976023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2337387204170227, + 0.07926754653453827, + 0.04598088562488556, + 0.11843233555555344, + 0.10775856673717499, + 0.06152584031224251, + 0.04562225192785263, + 0.03906618803739548, + 0.05766189098358154, + 0.210945725440979 + ], + "credit_map": [ + 0.04090508818626404, + 0.019723951816558838, + 0.015978507697582245, + 0.012749733403325081, + 0.024381574243307114, + 0.02291332557797432, + 0.06372560560703278, + 0.03336093947291374, + 0.01047500316053629, + 0.007696923334151506, + 0.008265087381005287, + 0.008075909689068794, + 0.006291720550507307, + 0.013466906733810902, + 0.012613050639629364, + 0.01238294504582882, + 0.007619459182024002, + 0.0030034498777240515, + 0.006257400847971439, + 0.0034151053987443447, + 0.006101819686591625, + 0.006231036502867937, + 0.007536089047789574, + 0.005816529039293528, + 0.016758227720856667, + 0.011396445333957672, + 0.01328642200678587, + 0.00999785028398037, + 0.016596077010035515, + 0.01683025248348713, + 0.020168745890259743, + 0.013398314826190472, + 0.0210419911891222, + 0.01388038881123066, + 0.007202576380223036, + 0.012270038947463036, + 0.009075355716049671, + 0.009398733265697956, + 0.020616868510842323, + 0.014272602275013924, + 0.004269883036613464, + 0.00504450174048543, + 0.004798609763383865, + 0.006081727799028158, + 0.008307299576699734, + 0.0030534665565937757, + 0.017598792910575867, + 0.012371559627354145, + 0.003098120214417577, + 0.00483371177688241, + 0.005527812521904707, + 0.007169734220951796, + 0.008249077945947647, + 0.0041906461119651794, + 0.006879949942231178, + 0.005673200357705355, + 0.003123712493106723, + 0.004880174528807402, + 0.004184015095233917, + 0.006807900033891201, + 0.004911570344120264, + 0.007333304733037949, + 0.002778533613309264, + 0.005046975798904896, + 0.006097816396504641, + 0.005469573196023703, + 0.0068945325911045074, + 0.008351732976734638, + 0.00759862968698144, + 0.00675249146297574, + 0.008322548121213913, + 0.008174567483365536, + 0.019927479326725006, + 0.028043555095791817, + 0.016123410314321518, + 0.02996513806283474, + 0.041088782250881195, + 0.027553357183933258, + 0.02083708345890045, + 0.027406923472881317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0036178866866976023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3263235092163086, + 0.13786174356937408, + 0.07917556166648865, + 0.027259385213255882, + 0.02712034434080124, + 0.2146434336900711, + 0.03133952617645264, + 0.08130410313606262, + 0.03720167651772499, + 0.03777073323726654 + ], + "credit_map": [ + 0.05944783240556717, + 0.043137822300195694, + 0.050979480147361755, + 0.01943960040807724, + 0.03427664563059807, + 0.03522319346666336, + 0.0488114133477211, + 0.03500748798251152, + 0.018796076998114586, + 0.018184134736657143, + 0.026543976739048958, + 0.013999206013977528, + 0.01350289024412632, + 0.013065661303699017, + 0.016604343429207802, + 0.01716543734073639, + 0.01161417830735445, + 0.0070573678240180016, + 0.017058104276657104, + 0.008154967799782753, + 0.008934669196605682, + 0.008381442166864872, + 0.011290126480162144, + 0.006684694439172745, + 0.004482013173401356, + 0.0014251640532165766, + 0.0026848330162465572, + 0.002716277725994587, + 0.002917296951636672, + 0.0028374840039759874, + 0.006566129624843597, + 0.003630182007327676, + 0.005239766091108322, + 0.0029376873280853033, + 0.0031494873110204935, + 0.001858966308645904, + 0.0028284417930990458, + 0.002862045541405678, + 0.0033990342635661364, + 0.004844913259148598, + 0.020089559257030487, + 0.016359545290470123, + 0.024944527074694633, + 0.019440384581685066, + 0.025955747812986374, + 0.01431050430983305, + 0.06050281971693039, + 0.03304031118750572, + 0.0031854426488280296, + 0.003182004438713193, + 0.0035026248078793287, + 0.0042485021986067295, + 0.004293748177587986, + 0.003315829439088702, + 0.004935444798320532, + 0.004675928503274918, + 0.00622959528118372, + 0.006674072705209255, + 0.006586837116628885, + 0.021939342841506004, + 0.00970405898988247, + 0.007089299149811268, + 0.011481297202408314, + 0.011599592864513397, + 0.005250368267297745, + 0.00431427825242281, + 0.0040197535417973995, + 0.005032283719629049, + 0.0051617249846458435, + 0.0037693637423217297, + 0.005328683648258448, + 0.004325215704739094, + 0.004079802427440882, + 0.004827212076634169, + 0.005238976329565048, + 0.006846458185464144, + 0.004776611924171448, + 0.0031282182317227125, + 0.003555367700755596, + 0.0053180851973593235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.008640285581350327 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle, framed in minimalist black and white, captures a male athlete in excruciating pain. A harsh, slanting light carves out his tense facial features. Against the high-contrast backdrop, his eyes are tightly closed, his brow knitted in agony. Between the deep wrinkles of a rickety crease, large beads of sweat form, trickling down his jawline, accentuated by his clenched teeth. His head tosses and turns slowly and painfully against a deep, almost indecipherable black background. Each twist is a silent struggle, as if attempting to escape the inescapable pain within him. His short, damp hair clings to his forehead and temples. The entire scene is shrouded in a primal and oppressive atmosphere. The intense visual impact focuses the viewer's attention on the sheer physical suffering, creating a suffocating sense of tension.", + "seed": 5371, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5371_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": -0.1104710859246552, + "vlm_logprob": -0.1104710859246552, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.1104710859246552, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23093363642692566 + ], + "opposite_logprob_tokens": [ + -1.605933666229248 + ], + "logprob_reward": -0.23093363642692566, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23093363642692566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793414056301117, + 0.06599647551774979, + 0.15458902716636658, + 0.07840865105390549, + 0.054920852184295654, + 0.062436293810606, + 0.05759059265255928, + 0.22293336689472198, + 0.027648083865642548, + 0.09613524377346039 + ], + "credit_map": [ + 0.013552661053836346, + 0.04164653271436691, + 0.021560395136475563, + 0.027755243703722954, + 0.01469896174967289, + 0.023096496239304543, + 0.012083313427865505, + 0.02494780346751213, + 0.005500906612724066, + 0.004356031306087971, + 0.006722528953105211, + 0.005115883890539408, + 0.0038570112083107233, + 0.01884373091161251, + 0.005488555412739515, + 0.01611183024942875, + 0.016602668911218643, + 0.01423331443220377, + 0.02047874592244625, + 0.021692153066396713, + 0.027329951524734497, + 0.012434795498847961, + 0.02451968379318714, + 0.01729772612452507, + 0.0073012723587453365, + 0.00834526214748621, + 0.008693572133779526, + 0.01037822850048542, + 0.010975291021168232, + 0.010065842419862747, + 0.014203725382685661, + 0.008445460349321365, + 0.009272273629903793, + 0.0053644669242203236, + 0.011111919768154621, + 0.005456887185573578, + 0.0042944056913256645, + 0.006316406186670065, + 0.004755824338644743, + 0.008348668925464153, + 0.0072610522620379925, + 0.009011374786496162, + 0.006921862252056599, + 0.005673335399478674, + 0.0063138664700090885, + 0.009776458144187927, + 0.00714973034337163, + 0.010328619740903378, + 0.00647992268204689, + 0.010165849700570107, + 0.007992531172931194, + 0.008184629492461681, + 0.005708637181669474, + 0.006690834648907185, + 0.005599114578217268, + 0.006769075058400631, + 0.02337372489273548, + 0.016945943236351013, + 0.02628898248076439, + 0.018599258735775948, + 0.040431950241327286, + 0.0376240536570549, + 0.020617155358195305, + 0.03905230760574341, + 0.0033989506773650646, + 0.002848783740773797, + 0.00353892776183784, + 0.00343069969676435, + 0.0027082408778369427, + 0.0044526043348014355, + 0.002840739442035556, + 0.0044291382655501366, + 0.011127438396215439, + 0.010449560359120369, + 0.011163941584527493, + 0.014676062390208244, + 0.009300315752625465, + 0.013184452429413795, + 0.013299576006829739, + 0.01293389406055212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.23093363642692566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813568115234375 + ], + "opposite_logprob_tokens": [ + -1.4063568115234375 + ], + "logprob_reward": -0.2813568115234375, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813568115234375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09631834179162979, + 0.1502678245306015, + 0.0779414102435112, + 0.10718531161546707, + 0.12357206642627716, + 0.06716537475585938, + 0.11314693838357925, + 0.12056644260883331, + 0.0420367605984211, + 0.10179953277111053 + ], + "credit_map": [ + 0.006965558510273695, + 0.010619044303894043, + 0.011769764125347137, + 0.01198029238730669, + 0.008283862844109535, + 0.018732227385044098, + 0.008596093393862247, + 0.019371502101421356, + 0.013113743625581264, + 0.017500873655080795, + 0.01627136394381523, + 0.014384117908775806, + 0.009056021459400654, + 0.03397952392697334, + 0.011819944716989994, + 0.03414224088191986, + 0.010524812154471874, + 0.008885935880243778, + 0.013820620253682137, + 0.014575847424566746, + 0.007851817645132542, + 0.006368418224155903, + 0.008220063522458076, + 0.007693891413509846, + 0.017142340540885925, + 0.012245986610651016, + 0.01653546839952469, + 0.014564820565283298, + 0.009856123477220535, + 0.009228833951056004, + 0.01912607066333294, + 0.008485664613544941, + 0.020435474812984467, + 0.01646599732339382, + 0.022185595706105232, + 0.012979505583643913, + 0.010294080711901188, + 0.01341012492775917, + 0.011952120810747147, + 0.015849165618419647, + 0.004731251858174801, + 0.004505207762122154, + 0.005110679194331169, + 0.004673812538385391, + 0.004044653847813606, + 0.018411977216601372, + 0.008543628267943859, + 0.01714416593313217, + 0.01450798287987709, + 0.013977832160890102, + 0.014522348530590534, + 0.014055491425096989, + 0.00977469701319933, + 0.016904329881072044, + 0.012444262392818928, + 0.016959985718131065, + 0.015514422208070755, + 0.008413231000304222, + 0.014530811458826065, + 0.009841358289122581, + 0.022088155150413513, + 0.020034821704030037, + 0.007469876669347286, + 0.02267376147210598, + 0.0046830205246806145, + 0.005783825181424618, + 0.004951717332005501, + 0.004859811160713434, + 0.004978736396878958, + 0.006523000076413155, + 0.004190101288259029, + 0.006066550966352224, + 0.012304314412176609, + 0.01457883045077324, + 0.011670173145830631, + 0.017941560596227646, + 0.008491241373121738, + 0.012456297874450684, + 0.01667662337422371, + 0.007680492475628853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.2813568115234375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708129078149796 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.033708129078149796, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708129078149796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10365024954080582, + 0.11050023883581161, + 0.08964915573596954, + 0.10426986217498779, + 0.06419774889945984, + 0.07455164194107056, + 0.04176607355475426, + 0.10202893614768982, + 0.04480979964137077, + 0.26457634568214417 + ], + "credit_map": [ + 0.006792804226279259, + 0.01019612979143858, + 0.013712388463318348, + 0.011345668695867062, + 0.00889995414763689, + 0.019777264446020126, + 0.009582171216607094, + 0.023343857377767563, + 0.011129185557365417, + 0.010115841403603554, + 0.010851224884390831, + 0.012425193563103676, + 0.010816645808517933, + 0.0213283933699131, + 0.01732867769896984, + 0.016505060717463493, + 0.014231352135539055, + 0.006741282530128956, + 0.015868699178099632, + 0.011962074786424637, + 0.010831854306161404, + 0.008119137026369572, + 0.014974347315728664, + 0.0069204033352434635, + 0.01641344465315342, + 0.01377259474247694, + 0.01447268296033144, + 0.016793303191661835, + 0.011133164167404175, + 0.008472270332276821, + 0.012216663919389248, + 0.010995721444487572, + 0.020488837733864784, + 0.004776624031364918, + 0.020277762785553932, + 0.004671887028962374, + 0.002265606075525284, + 0.002568215597420931, + 0.005298079922795296, + 0.0038507264107465744, + 0.006102351471781731, + 0.008999050594866276, + 0.00988099817186594, + 0.00545912142843008, + 0.00525682233273983, + 0.013168935663998127, + 0.010429933667182922, + 0.015254427678883076, + 0.005507452413439751, + 0.0040877447463572025, + 0.005408850964158773, + 0.005280935205519199, + 0.0034739819820970297, + 0.0068147447891533375, + 0.004523748066276312, + 0.006668612826615572, + 0.010851443745195866, + 0.007828958332538605, + 0.012316952459514141, + 0.010928996838629246, + 0.014810259453952312, + 0.01846322789788246, + 0.007530024740844965, + 0.019299067556858063, + 0.004231973085552454, + 0.006265647243708372, + 0.0030424112919718027, + 0.0069381240755319595, + 0.004508921876549721, + 0.007844486273825169, + 0.004034234676510096, + 0.007943996228277683, + 0.030567187815904617, + 0.03287696838378906, + 0.027270028367638588, + 0.0375349186360836, + 0.023346712812781334, + 0.0268380269408226, + 0.052710577845573425, + 0.03343190997838974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 3 + ], + "selected_reward": -0.033708129078149796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641829524189234 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641829524189234, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641829524189234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11611634492874146, + 0.19317527115345, + 0.0692899152636528, + 0.05680115520954132, + 0.09736286103725433, + 0.02370145544409752, + 0.10858606547117233, + 0.0772276222705841, + 0.03901474550366402, + 0.2187245488166809 + ], + "credit_map": [ + 0.0073503004387021065, + 0.02573913335800171, + 0.019445639103651047, + 0.016224510967731476, + 0.009869903326034546, + 0.014623777009546757, + 0.007701725699007511, + 0.015161357820034027, + 0.01589950919151306, + 0.012134552001953125, + 0.02963307499885559, + 0.011775068007409573, + 0.014293774031102657, + 0.05503476411104202, + 0.01822035200893879, + 0.03618418052792549, + 0.009912392124533653, + 0.0062909373082220554, + 0.011061295866966248, + 0.010412650182843208, + 0.00897142942994833, + 0.0065694451332092285, + 0.011039678007364273, + 0.005032092332839966, + 0.006024177186191082, + 0.0059403181076049805, + 0.006003273651003838, + 0.0067337085492908955, + 0.005862507037818432, + 0.004294305574148893, + 0.014091214165091515, + 0.00785165186971426, + 0.014969675801694393, + 0.010870186612010002, + 0.016795247793197632, + 0.010546296834945679, + 0.008183942176401615, + 0.012215645983815193, + 0.013556096702814102, + 0.010225776582956314, + 0.00251774862408638, + 0.003308284794911742, + 0.004000072367489338, + 0.0023713666014373302, + 0.0018375631188973784, + 0.0036975143011659384, + 0.002430233871564269, + 0.0035386704839766026, + 0.011582275852560997, + 0.012447649613022804, + 0.01567225717008114, + 0.014940163120627403, + 0.00968892127275467, + 0.01600859872996807, + 0.01317874900996685, + 0.01506744883954525, + 0.00743933254852891, + 0.006050517782568932, + 0.010540531016886234, + 0.01060907356441021, + 0.012166083790361881, + 0.011215395294129848, + 0.007338223047554493, + 0.011868459172546864, + 0.0034491433762013912, + 0.004351649899035692, + 0.004365407861769199, + 0.0037239640951156616, + 0.005390499718487263, + 0.006174375768750906, + 0.005122609902173281, + 0.006437097210437059, + 0.024836603552103043, + 0.0309353768825531, + 0.022385118529200554, + 0.026088163256645203, + 0.02352420799434185, + 0.02733606845140457, + 0.039632245898246765, + 0.02398676797747612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004641829524189234 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017150230705738068 + ], + "opposite_logprob_tokens": [ + -6.376715183258057 + ], + "logprob_reward": -0.0017150230705738068, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017150230705738068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16937169432640076, + 0.15687841176986694, + 0.0749080628156662, + 0.0839560404419899, + 0.14484702050685883, + 0.0886496901512146, + 0.12913788855075836, + 0.031145047396421432, + 0.019855370745062828, + 0.1012507751584053 + ], + "credit_map": [ + 0.009988801553845406, + 0.02105376310646534, + 0.01828492432832718, + 0.015150438994169235, + 0.010657304897904396, + 0.0372920036315918, + 0.010643668472766876, + 0.04630078747868538, + 0.009609349071979523, + 0.005563699174672365, + 0.016913555562496185, + 0.00629831338301301, + 0.007022957317531109, + 0.059622474014759064, + 0.007604085840284824, + 0.044243983924388885, + 0.01063668355345726, + 0.007769517134875059, + 0.011664038524031639, + 0.008216784335672855, + 0.01220735814422369, + 0.005777453538030386, + 0.011360864154994488, + 0.0072753592394292355, + 0.013013502582907677, + 0.011701932176947594, + 0.012631116434931755, + 0.01437209453433752, + 0.007741731591522694, + 0.007847416214644909, + 0.008216443471610546, + 0.008431813679635525, + 0.04177555814385414, + 0.009294603951275349, + 0.04451358690857887, + 0.011794582940638065, + 0.00881256628781557, + 0.009613292291760445, + 0.007194825913757086, + 0.011848002672195435, + 0.009328084997832775, + 0.01265606191009283, + 0.012264701537787914, + 0.0068671805784106255, + 0.007394229061901569, + 0.0142029644921422, + 0.010144917294383049, + 0.015791548416018486, + 0.016717083752155304, + 0.008689993992447853, + 0.012506559491157532, + 0.009810672141611576, + 0.005614796653389931, + 0.02825162187218666, + 0.014944057911634445, + 0.03260310739278793, + 0.002387174405157566, + 0.003242093836888671, + 0.004245138261467218, + 0.0038938040379434824, + 0.00398597028106451, + 0.004624881315976381, + 0.0024124174378812313, + 0.0063535673543810844, + 0.0017015927005559206, + 0.0026848551351577044, + 0.0023053179029375315, + 0.0025939045008271933, + 0.0011207959614694118, + 0.003633406711742282, + 0.0019410321256145835, + 0.0038744660560041666, + 0.013538692146539688, + 0.010479569435119629, + 0.013377810828387737, + 0.015238329768180847, + 0.007463185582309961, + 0.012133436277508736, + 0.015071938745677471, + 0.013947808183729649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0017150230705738068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up shot from a slightly overhead angle, framed in minimalist black and white, captures a male athlete in excruciating pain. A harsh, slanting light carves out his tense facial features. Against the high-contrast backdrop, his eyes are tightly closed, his brow knitted in agony. Between the deep wrinkles of a rickety crease, large beads of sweat form, trickling down his jawline, accentuated by his clenched teeth. His head tosses and turns slowly and painfully against a deep, almost indecipherable black background. Each twist is a silent struggle, as if attempting to escape the inescapable pain within him. His short, damp hair clings to his forehead and temples. The entire scene is shrouded in a primal and oppressive atmosphere. The intense visual impact focuses the viewer's attention on the sheer physical suffering, creating a suffocating sense of tension.", + "seed": 5372, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_13/index_3755_seed_5372_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": -0.0485131241613999, + "vlm_logprob": -0.0485131241613999, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.0485131241613999, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1327376514673233 + ], + "opposite_logprob_tokens": [ + -2.132737636566162 + ], + "logprob_reward": -0.1327376514673233, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1327376514673233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138986274600029, + 0.08574534952640533, + 0.08020604401826859, + 0.13722839951515198, + 0.3084995746612549, + 0.03584897518157959, + 0.05992073565721512, + 0.050540391355752945, + 0.06467031687498093, + 0.038353968411684036 + ], + "credit_map": [ + 0.021918457001447678, + 0.015375646762549877, + 0.019082140177488327, + 0.010817032307386398, + 0.02313237078487873, + 0.01936214789748192, + 0.012054317630827427, + 0.017244158312678337, + 0.013431594707071781, + 0.004884682130068541, + 0.011067931540310383, + 0.01121241319924593, + 0.010554826818406582, + 0.013472710736095905, + 0.009620696306228638, + 0.01150048989802599, + 0.009763473644852638, + 0.004965777974575758, + 0.00909896194934845, + 0.006561082322150469, + 0.00790692400187254, + 0.02111050859093666, + 0.006978639867156744, + 0.013820678927004337, + 0.013993554748594761, + 0.01869840919971466, + 0.019347159191966057, + 0.017726892605423927, + 0.014347188174724579, + 0.015200037509202957, + 0.019197002053260803, + 0.01871815323829651, + 0.02043731138110161, + 0.034339677542448044, + 0.044152963906526566, + 0.0422537662088871, + 0.034222375601530075, + 0.051919106394052505, + 0.0277304295450449, + 0.053443942219018936, + 0.003077977802604437, + 0.004780012182891369, + 0.00564745394513011, + 0.0053557464852929115, + 0.003268555039539933, + 0.0053391349501907825, + 0.004172999411821365, + 0.004207095131278038, + 0.006616990547627211, + 0.006368107628077269, + 0.006992724724113941, + 0.007231513503938913, + 0.0056632827036082745, + 0.012579618021845818, + 0.004578821361064911, + 0.009889673441648483, + 0.007070725318044424, + 0.005420757923275232, + 0.006173126865178347, + 0.0037846756167709827, + 0.006914793513715267, + 0.007806119043380022, + 0.006426582112908363, + 0.006943613290786743, + 0.014858315698802471, + 0.004602213855832815, + 0.010975495912134647, + 0.004944712854921818, + 0.007950886152684689, + 0.005894232075661421, + 0.010159485973417759, + 0.005284978076815605, + 0.004338645841926336, + 0.0033269282430410385, + 0.0053431009873747826, + 0.004266970790922642, + 0.005446774419397116, + 0.007183803711086512, + 0.0028192256577312946, + 0.005628518760204315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.1327376514673233 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1691138595342636, + 0.06889114528894424, + 0.1565839648246765, + 0.05241404101252556, + 0.1639614999294281, + 0.09711864590644836, + 0.09100399166345596, + 0.05008403956890106, + 0.08050009608268738, + 0.07032876461744308 + ], + "credit_map": [ + 0.03231855109333992, + 0.016355929896235466, + 0.027035916224122047, + 0.013925786130130291, + 0.025872517377138138, + 0.022456059232354164, + 0.015495550818741322, + 0.015653563663363457, + 0.011721014976501465, + 0.005054007284343243, + 0.010405955836176872, + 0.007280428893864155, + 0.008165739476680756, + 0.011658819392323494, + 0.006984303239732981, + 0.007620873395353556, + 0.019997937604784966, + 0.014488349668681622, + 0.01968151144683361, + 0.011182178743183613, + 0.016613129526376724, + 0.03355826437473297, + 0.011121436953544617, + 0.029941150918602943, + 0.005496135912835598, + 0.010368579998612404, + 0.007991708815097809, + 0.006238899193704128, + 0.005944154225289822, + 0.004998530261218548, + 0.005438328254967928, + 0.005937702488154173, + 0.011304400861263275, + 0.017885996028780937, + 0.02255062572658062, + 0.024903500452637672, + 0.017665497958660126, + 0.025113528594374657, + 0.016939939931035042, + 0.027598008513450623, + 0.0070576854050159454, + 0.01279377844184637, + 0.013601988554000854, + 0.012411496601998806, + 0.009418617933988571, + 0.01571265235543251, + 0.013357847929000854, + 0.01276456844061613, + 0.010324792936444283, + 0.011822396889328957, + 0.010194512084126472, + 0.01082240417599678, + 0.009095875546336174, + 0.018009312450885773, + 0.006745082326233387, + 0.01398961991071701, + 0.007814631797373295, + 0.004589139483869076, + 0.006748472806066275, + 0.0040609086863696575, + 0.006758336443454027, + 0.0068734497763216496, + 0.0077972873114049435, + 0.0054418169893324375, + 0.01517027709633112, + 0.009285415522754192, + 0.012060103937983513, + 0.009613323956727982, + 0.011944097466766834, + 0.0057253120467066765, + 0.008665564469993114, + 0.008036001585423946, + 0.010501673445105553, + 0.006724487524479628, + 0.010145436972379684, + 0.007864528335630894, + 0.00943893101066351, + 0.010291533544659615, + 0.004902532324194908, + 0.01045964565128088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.08904317021369934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203761704266071 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203761704266071, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203761704266071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.197678342461586, + 0.10708827525377274, + 0.10521003603935242, + 0.04238777607679367, + 0.33068785071372986, + 0.06868457049131393, + 0.05336196348071098, + 0.0371343269944191, + 0.029405036941170692, + 0.028361795470118523 + ], + "credit_map": [ + 0.02615346387028694, + 0.021580545231699944, + 0.026473144069314003, + 0.021709736436605453, + 0.030950747430324554, + 0.021089229732751846, + 0.018922971561551094, + 0.030798517167568207, + 0.013808309100568295, + 0.007378614041954279, + 0.01466408558189869, + 0.015867140144109726, + 0.010918260551989079, + 0.016197415068745613, + 0.011814797297120094, + 0.016439661383628845, + 0.013342312537133694, + 0.011461344547569752, + 0.014156271703541279, + 0.009045316837728024, + 0.011707520112395287, + 0.021314239129424095, + 0.009962697513401508, + 0.014220332726836205, + 0.005268272943794727, + 0.0071465144865214825, + 0.0064193918369710445, + 0.004105918575078249, + 0.004435037262737751, + 0.004540111869573593, + 0.004478934686630964, + 0.005993598606437445, + 0.021708859130740166, + 0.035445984452962875, + 0.047332763671875, + 0.05469796434044838, + 0.03370794281363487, + 0.05634469538927078, + 0.030689304694533348, + 0.050760332494974136, + 0.005785706918686628, + 0.008078765124082565, + 0.011244293302297592, + 0.01268619205802679, + 0.005503786727786064, + 0.012430351227521896, + 0.005145674105733633, + 0.00780980521813035, + 0.007657504640519619, + 0.005683821626007557, + 0.007101498078554869, + 0.007534803356975317, + 0.005835601128637791, + 0.00940368790179491, + 0.0030701509676873684, + 0.007074899040162563, + 0.0077845449559390545, + 0.002594700548797846, + 0.005382182076573372, + 0.0033963588066399097, + 0.003709228243678808, + 0.004930973052978516, + 0.007038492243736982, + 0.002297850325703621, + 0.004786140751093626, + 0.003905846271663904, + 0.004182657226920128, + 0.0026010475121438503, + 0.004775224719196558, + 0.003234727308154106, + 0.002861092332750559, + 0.0030583026818931103, + 0.0036406994331628084, + 0.0016718938713893294, + 0.003930729813873768, + 0.002578892046585679, + 0.0028643589466810226, + 0.00757543696090579, + 0.0015744094271212816, + 0.004525377415120602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.014203761704266071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23864610493183136, + 0.0851048156619072, + 0.17094430327415466, + 0.04062534123659134, + 0.08446203917264938, + 0.1350087821483612, + 0.05343940109014511, + 0.04736200347542763, + 0.042479969561100006, + 0.10192718356847763 + ], + "credit_map": [ + 0.046439558267593384, + 0.017223717644810677, + 0.036516785621643066, + 0.017673872411251068, + 0.03125984966754913, + 0.03261017054319382, + 0.02029709704220295, + 0.03662506863474846, + 0.014200528152287006, + 0.005490906536579132, + 0.012000850401818752, + 0.011183924973011017, + 0.011772120371460915, + 0.01414272841066122, + 0.008479848504066467, + 0.007833912968635559, + 0.020993225276470184, + 0.019361378625035286, + 0.022948823869228363, + 0.0124199902638793, + 0.0175533015280962, + 0.03680002689361572, + 0.009682795964181423, + 0.031184764578938484, + 0.005173497833311558, + 0.00657992996275425, + 0.005496510770171881, + 0.004319299478083849, + 0.004618452861905098, + 0.0035554824862629175, + 0.005156693048775196, + 0.005725472699850798, + 0.007451786193996668, + 0.011050873436033726, + 0.013403196819126606, + 0.0143443513661623, + 0.006759554613381624, + 0.012730402871966362, + 0.008070852607488632, + 0.010651024989783764, + 0.008620387874543667, + 0.016351686790585518, + 0.02173398621380329, + 0.02574102394282818, + 0.011366987600922585, + 0.019960973411798477, + 0.01779918000102043, + 0.01343455445021391, + 0.007426220923662186, + 0.0061149816028773785, + 0.006351794116199017, + 0.00799460057169199, + 0.004708119202405214, + 0.010683187283575535, + 0.0037043720949441195, + 0.006456124130636454, + 0.006930666044354439, + 0.003794144606217742, + 0.006186603102833033, + 0.0037895196583122015, + 0.005678483285009861, + 0.008180923759937286, + 0.009546078741550446, + 0.0032555852085351944, + 0.008611710742115974, + 0.0037836560513824224, + 0.005610241089016199, + 0.003952768165618181, + 0.005906485486775637, + 0.006652908865362406, + 0.004209223203361034, + 0.0037529750261455774, + 0.011985738761723042, + 0.008482588455080986, + 0.013360362499952316, + 0.017289869487285614, + 0.011778860352933407, + 0.019582953304052353, + 0.006771130487322807, + 0.012675676494836807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.004640049301087856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the image rendered in black and white with high contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are large beads of sweat visible on the athlete's face and jawline?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background deep and almost indecipherable black?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001940988120622933 + ], + "opposite_logprob_tokens": [ + -6.251941204071045 + ], + "logprob_reward": -0.001940988120622933, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001940988120622933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12116397172212601, + 0.1359141767024994, + 0.1446208357810974, + 0.02865660935640335, + 0.17222602665424347, + 0.06746728718280792, + 0.13413573801517487, + 0.09167832136154175, + 0.04687828570604324, + 0.05725870281457901 + ], + "credit_map": [ + 0.022676317021250725, + 0.009259101934731007, + 0.01988510973751545, + 0.01176183857023716, + 0.015044035390019417, + 0.014174537733197212, + 0.007720886263996363, + 0.020642142742872238, + 0.022477785125374794, + 0.009195947088301182, + 0.01963210478425026, + 0.016300708055496216, + 0.013071628287434578, + 0.022359604015946388, + 0.013153938576579094, + 0.01972246542572975, + 0.01840585470199585, + 0.012831125408411026, + 0.017868483439087868, + 0.011490071192383766, + 0.015147312544286251, + 0.03083753027021885, + 0.011942615732550621, + 0.02609783411026001, + 0.005012993700802326, + 0.004322083201259375, + 0.004078864119946957, + 0.002462563570588827, + 0.00369742838665843, + 0.00180431071203202, + 0.004281867761164904, + 0.002996498951688409, + 0.010546717792749405, + 0.018320031464099884, + 0.025984985753893852, + 0.03152046352624893, + 0.0186979528516531, + 0.025164855644106865, + 0.017175201326608658, + 0.02481582760810852, + 0.006582778878509998, + 0.009976109489798546, + 0.011863384395837784, + 0.013161763548851013, + 0.004665918182581663, + 0.008394565433263779, + 0.007326046004891396, + 0.00549672357738018, + 0.017111672088503838, + 0.013021213002502918, + 0.01943453773856163, + 0.03276755288243294, + 0.009558872319757938, + 0.019994350150227547, + 0.008808962069451809, + 0.01343858428299427, + 0.01700526475906372, + 0.00678718276321888, + 0.01310014445334673, + 0.010009054094552994, + 0.010836335830390453, + 0.014083298854529858, + 0.011006733402609825, + 0.008850312791764736, + 0.007560239173471928, + 0.004326543305069208, + 0.006061780732125044, + 0.003853331785649061, + 0.006916898302733898, + 0.0068434299901127815, + 0.007298415061086416, + 0.004017647821456194, + 0.007595307193696499, + 0.005572764668613672, + 0.008331017568707466, + 0.010505923070013523, + 0.005973227322101593, + 0.008651189506053925, + 0.003981598652899265, + 0.0066476743668317795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.001940988120622933 + } + ], + "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/0000152/group_14/index_tensor([5414, 5414])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_tensor([5414, 5414])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7e426fd2ced598ad1230745b86f9797f0ef8a61b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_tensor([5414, 5414])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, captured from a subtle overhead angle, immerses the viewer in a violently boiling pool of incandescent molten lava. The surface of the molten, viscous, orange-red magma churns and swirls, specks of cooling, darkening black scum floating on it, while cracks reveal a brighter, more dazzling yellow light. Vast bubbles rise slowly and menacingly from the depths, pushing the surface of the magma into domes. Then, reaching their limit, they violently burst, hurling countless sparks and strands of molten lava into the air. Shot at a high frame rate, the entire eruption is captured in surreal slow motion, with every drop of molten lava, shimmering with an inner light, traced in the air with exquisite clarity. Meanwhile, thick, sulfurous black and gray smoke billows from the surface of the magma pool, sharply outlined by the bright light from the molten rock below, twisting and spiraling upward. The entire scene is illuminated entirely by the glow of the lava itself, casting a warm yet menacing shimmer of light and shadow, creating an apocalyptic atmosphere that is raw, violent, and awe-inspiring.", + "seed": 29924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_5414_seed_29924_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": -0.057363017648458484, + "vlm_logprob": -0.057363017648458484, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.057363017648458484, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05812736600637436 + ], + "opposite_logprob_tokens": [ + -2.9331274032592773 + ], + "logprob_reward": -0.05812736600637436, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05812736600637436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1296404004096985, + 0.04950864985585213, + 0.14094223082065582, + 0.12099673599004745, + 0.10106052458286285, + 0.05586416646838188, + 0.10226395726203918, + 0.06854621320962906, + 0.12090519815683365, + 0.11027192324399948 + ], + "credit_map": [ + 0.023760007694363594, + 0.011944265104830265, + 0.01696515455842018, + 0.015280422754585743, + 0.02429841086268425, + 0.013158334419131279, + 0.010297759436070919, + 0.013936054892838001, + 0.009013409726321697, + 0.0041560218669474125, + 0.007116248365491629, + 0.006101842969655991, + 0.005866951309144497, + 0.0062464517541229725, + 0.004519039765000343, + 0.006488690618425608, + 0.014037900604307652, + 0.009271184913814068, + 0.018672090023756027, + 0.013826748356223106, + 0.012533075176179409, + 0.022801559418439865, + 0.028037359938025475, + 0.02176233008503914, + 0.010280342772603035, + 0.010848366655409336, + 0.008279162459075451, + 0.010974645614624023, + 0.02128632925450802, + 0.02674134448170662, + 0.019343312829732895, + 0.013243242166936398, + 0.008549508638679981, + 0.026046516373753548, + 0.0030498839914798737, + 0.0055199917405843735, + 0.008669824339449406, + 0.038275714963674545, + 0.0035753720439970493, + 0.007373718544840813, + 0.003991208039224148, + 0.004373783711344004, + 0.0054452503100037575, + 0.007687270641326904, + 0.010835097171366215, + 0.007492280099540949, + 0.008680606260895729, + 0.007358675356954336, + 0.00985938310623169, + 0.008397392928600311, + 0.013391606509685516, + 0.010591601021587849, + 0.01599838398396969, + 0.014393377117812634, + 0.015711352229118347, + 0.013920867815613747, + 0.0045675430446863174, + 0.0068213134072721004, + 0.007795402780175209, + 0.007108757272362709, + 0.014020155183970928, + 0.01358846016228199, + 0.005807827226817608, + 0.008836760185658932, + 0.007843445986509323, + 0.019892172887921333, + 0.0072402143850922585, + 0.011535145342350006, + 0.034502871334552765, + 0.019676433876156807, + 0.010471891611814499, + 0.009743034839630127, + 0.006176072172820568, + 0.024787576869130135, + 0.007292686030268669, + 0.012799681164324284, + 0.03549768030643463, + 0.009172655642032623, + 0.008167804218828678, + 0.006377787329256535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.05812736600637436 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700482726097107 + ], + "opposite_logprob_tokens": [ + -2.6950483322143555 + ], + "logprob_reward": -0.0700482726097107, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700482726097107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1378966122865677, + 0.027910497039556503, + 0.08210822194814682, + 0.25137510895729065, + 0.17882664501667023, + 0.028211096301674843, + 0.10705015063285828, + 0.02859628014266491, + 0.11206149309873581, + 0.04596388339996338 + ], + "credit_map": [ + 0.02915886417031288, + 0.01546113658696413, + 0.017918996512889862, + 0.01599653623998165, + 0.016768528148531914, + 0.01468865480273962, + 0.015018447302281857, + 0.012885444797575474, + 0.004869387019425631, + 0.0022832839749753475, + 0.003748444141820073, + 0.0031363768503069878, + 0.004646364599466324, + 0.0035706274211406708, + 0.0027062154840677977, + 0.002949794288724661, + 0.009038964286446571, + 0.004692840855568647, + 0.01218255702406168, + 0.008028174750506878, + 0.006686039734631777, + 0.013915041461586952, + 0.013990896753966808, + 0.013573704287409782, + 0.015820128843188286, + 0.028781218454241753, + 0.014401683583855629, + 0.018572667613625526, + 0.06921641528606415, + 0.0464245080947876, + 0.032984305173158646, + 0.025174187496304512, + 0.015676800161600113, + 0.050155382603406906, + 0.004618297331035137, + 0.005753279197961092, + 0.013966451399028301, + 0.07051216065883636, + 0.0064179557375609875, + 0.011726311407983303, + 0.0019012561533600092, + 0.0023823187220841646, + 0.0030976550187915564, + 0.004223175812512636, + 0.005064566154032946, + 0.003423880785703659, + 0.004486516583710909, + 0.0036317261401563883, + 0.01131981797516346, + 0.009028470143675804, + 0.01567162573337555, + 0.01300856750458479, + 0.01650327444076538, + 0.015213618054986, + 0.012125024572014809, + 0.014179753139615059, + 0.002277829684317112, + 0.0034165906254202127, + 0.002904278691858053, + 0.0034115121234208345, + 0.006696986500173807, + 0.004287476651370525, + 0.00309062865562737, + 0.0025109744165092707, + 0.006962289102375507, + 0.017546428367495537, + 0.006858199834823608, + 0.010227163322269917, + 0.02942647971212864, + 0.020644865930080414, + 0.009317836724221706, + 0.011078223586082458, + 0.00463931355625391, + 0.0050188349559903145, + 0.0060811652801930904, + 0.007802037056535482, + 0.006952214520424604, + 0.00627388060092926, + 0.005308373365551233, + 0.0038880633655935526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.0700482726097107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038093168288469315 + ], + "opposite_logprob_tokens": [ + -3.288093090057373 + ], + "logprob_reward": -0.038093168288469315, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038093168288469315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1036020889878273, + 0.046473149210214615, + 0.19130800664424896, + 0.2270440012216568, + 0.11443115770816803, + 0.1308007836341858, + 0.023237524554133415, + 0.04388171061873436, + 0.059893205761909485, + 0.0593283474445343 + ], + "credit_map": [ + 0.020571168512105942, + 0.01123946812003851, + 0.015790583565831184, + 0.011497030965983868, + 0.0129804452881217, + 0.01024469081312418, + 0.012326641008257866, + 0.008952058851718903, + 0.0061139194294810295, + 0.0030958144925534725, + 0.007290922570973635, + 0.0039008241146802902, + 0.0064492556266486645, + 0.006537930574268103, + 0.004934453871101141, + 0.008150029927492142, + 0.020039310678839684, + 0.011317468248307705, + 0.035665690898895264, + 0.03654035925865173, + 0.01736164651811123, + 0.023623278364539146, + 0.01898297294974327, + 0.027777284383773804, + 0.014042433351278305, + 0.027383960783481598, + 0.011197760701179504, + 0.018050942569971085, + 0.06591250002384186, + 0.040149886161088943, + 0.027792371809482574, + 0.02251412160694599, + 0.013089397922158241, + 0.02225964330136776, + 0.006737091112881899, + 0.009888123720884323, + 0.00906309299170971, + 0.03275028616189957, + 0.011736982502043247, + 0.008906543254852295, + 0.00935082696378231, + 0.012533948756754398, + 0.012942329980432987, + 0.013190770521759987, + 0.025012439116835594, + 0.016272904351353645, + 0.019455794245004654, + 0.022041769698262215, + 0.0019666990265250206, + 0.0019284229492768645, + 0.0029711024835705757, + 0.003026076126843691, + 0.0038688809145241976, + 0.0032699499279260635, + 0.003077759640291333, + 0.0031286331359297037, + 0.0032212280202656984, + 0.004544409457594156, + 0.006970433983951807, + 0.004317758604884148, + 0.0069074067287147045, + 0.008095766417682171, + 0.0049401759169995785, + 0.0048845321871340275, + 0.0049479249864816666, + 0.006567426957190037, + 0.004949662368744612, + 0.005608281120657921, + 0.013766270130872726, + 0.009870924055576324, + 0.008111217990517616, + 0.006071498617529869, + 0.009028555825352669, + 0.005244633182883263, + 0.00888089369982481, + 0.008479022420942783, + 0.006373634561896324, + 0.008898498490452766, + 0.005750124808400869, + 0.00667298398911953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.038093168288469315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1226513609290123, + 0.04602237418293953, + 0.07328969240188599, + 0.2682802081108093, + 0.10042142868041992, + 0.07604538649320602, + 0.07582513988018036, + 0.06266192346811295, + 0.028098737820982933, + 0.14670367538928986 + ], + "credit_map": [ + 0.01513723749667406, + 0.012212381698191166, + 0.011334455572068691, + 0.012897883541882038, + 0.017884647473692894, + 0.01959339901804924, + 0.016505271196365356, + 0.017086084932088852, + 0.006100560538470745, + 0.004647447261959314, + 0.004615257028490305, + 0.004585263319313526, + 0.007540496997535229, + 0.008806300349533558, + 0.004517697729170322, + 0.005209350027143955, + 0.007629546336829662, + 0.005571332760155201, + 0.008573736064136028, + 0.010555332526564598, + 0.009812416508793831, + 0.010228612460196018, + 0.00968807190656662, + 0.011230643838644028, + 0.015338758938014507, + 0.03185337781906128, + 0.013417018577456474, + 0.02010795660316944, + 0.07240486890077591, + 0.05149032175540924, + 0.03581492602825165, + 0.027852976694703102, + 0.010026692412793636, + 0.02646722085773945, + 0.003408371703699231, + 0.0041920882649719715, + 0.011143332347273827, + 0.03473654016852379, + 0.00368869979865849, + 0.006758491508662701, + 0.00736142648383975, + 0.0070213512517511845, + 0.006343807093799114, + 0.011567853391170502, + 0.012609419412910938, + 0.00922246091067791, + 0.009209880605340004, + 0.01270918920636177, + 0.007084087934345007, + 0.006160722579807043, + 0.011133931577205658, + 0.007963428273797035, + 0.013200543820858002, + 0.010275368578732014, + 0.009260119870305061, + 0.010746941901743412, + 0.005736052058637142, + 0.006469316780567169, + 0.0060703386552631855, + 0.007811499293893576, + 0.00995378103107214, + 0.012185218743979931, + 0.005716552026569843, + 0.00871916301548481, + 0.003074661362916231, + 0.0025838909205049276, + 0.0031549434643238783, + 0.004261439200490713, + 0.005011497996747494, + 0.003685359377413988, + 0.0035010778810828924, + 0.0028258669190108776, + 0.011281903833150864, + 0.028007274493575096, + 0.011111801490187645, + 0.012248042039573193, + 0.04319620504975319, + 0.019480004906654358, + 0.011265448294579983, + 0.010113006457686424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291413754224777 + ], + "opposite_logprob_tokens": [ + -2.2379140853881836 + ], + "logprob_reward": -0.11291413754224777, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291413754224777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292544215917587, + 0.04221112281084061, + 0.06502087414264679, + 0.09147956222295761, + 0.25798356533050537, + 0.042026273906230927, + 0.2005777657032013, + 0.07589569687843323, + 0.05486602708697319, + 0.06701364368200302 + ], + "credit_map": [ + 0.01897701807320118, + 0.0107705844566226, + 0.013459594920277596, + 0.010057665407657623, + 0.017246657982468605, + 0.010648918338119984, + 0.010786532424390316, + 0.010978487320244312, + 0.007978439331054688, + 0.003139522857964039, + 0.0042503164149820805, + 0.004575521685183048, + 0.005307887215167284, + 0.0052841720171272755, + 0.005676292348653078, + 0.0059989700093865395, + 0.00639173062518239, + 0.006961148697882891, + 0.006414866540580988, + 0.012728441506624222, + 0.010264436714351177, + 0.007475466467440128, + 0.008668136782944202, + 0.006116650532931089, + 0.008317023515701294, + 0.012047392316162586, + 0.005282939877361059, + 0.00734154274687171, + 0.02118447795510292, + 0.018841488286852837, + 0.006319020409137011, + 0.012145684100687504, + 0.022775426506996155, + 0.07760993391275406, + 0.0057843709364533424, + 0.005466618109494448, + 0.01624310202896595, + 0.10597742348909378, + 0.006318208761513233, + 0.017808500677347183, + 0.003041911171749234, + 0.003419804386794567, + 0.003745772410184145, + 0.00674580130726099, + 0.007616983260959387, + 0.004981111269444227, + 0.00690593384206295, + 0.005568956024944782, + 0.02233194373548031, + 0.017890822142362595, + 0.028979839757084846, + 0.022092372179031372, + 0.03697007894515991, + 0.027964072301983833, + 0.021007535979151726, + 0.023341119289398193, + 0.007520800456404686, + 0.010033303871750832, + 0.008428334258496761, + 0.006007895804941654, + 0.011664925143122673, + 0.01534710731357336, + 0.006432721856981516, + 0.010460617952048779, + 0.0035848976112902164, + 0.008492560125887394, + 0.004149758256971836, + 0.005760399159044027, + 0.015322020277380943, + 0.006885158829391003, + 0.005612869746983051, + 0.005058365408331156, + 0.005793845281004906, + 0.011236830614507198, + 0.006350421346724033, + 0.006592220161110163, + 0.01660781539976597, + 0.009226549416780472, + 0.005953178275376558, + 0.005252786912024021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.11291413754224777 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a subtle overhead angle, immerses the viewer in a violently boiling pool of incandescent molten lava. The surface of the molten, viscous, orange-red magma churns and swirls, specks of cooling, darkening black scum floating on it, while cracks reveal a brighter, more dazzling yellow light. Vast bubbles rise slowly and menacingly from the depths, pushing the surface of the magma into domes. Then, reaching their limit, they violently burst, hurling countless sparks and strands of molten lava into the air. Shot at a high frame rate, the entire eruption is captured in surreal slow motion, with every drop of molten lava, shimmering with an inner light, traced in the air with exquisite clarity. Meanwhile, thick, sulfurous black and gray smoke billows from the surface of the magma pool, sharply outlined by the bright light from the molten rock below, twisting and spiraling upward. The entire scene is illuminated entirely by the glow of the lava itself, casting a warm yet menacing shimmer of light and shadow, creating an apocalyptic atmosphere that is raw, violent, and awe-inspiring.", + "seed": 29925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_5414_seed_29925_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": -0.10395072745159269, + "vlm_logprob": -0.10395072745159269, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.10395072745159269, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25995367765426636 + ], + "opposite_logprob_tokens": [ + -1.5099537372589111 + ], + "logprob_reward": -0.25995367765426636, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25995367765426636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11994770914316177, + 0.12130381911993027, + 0.0536310188472271, + 0.04710595682263374, + 0.087053582072258, + 0.16834919154644012, + 0.04453522339463234, + 0.12040742486715317, + 0.15944726765155792, + 0.07821878790855408 + ], + "credit_map": [ + 0.022574132308363914, + 0.014264539815485477, + 0.013551362790167332, + 0.011851340532302856, + 0.017678415402770042, + 0.018705865368247032, + 0.00920199230313301, + 0.012120055966079235, + 0.01413224171847105, + 0.01652310974895954, + 0.014102938584983349, + 0.011757778003811836, + 0.015303000807762146, + 0.01713874377310276, + 0.01584937982261181, + 0.016496628522872925, + 0.006295827217400074, + 0.011090063489973545, + 0.005395950749516487, + 0.005808988586068153, + 0.007028156891465187, + 0.00826988834887743, + 0.005088514182716608, + 0.004653629846870899, + 0.006176571827381849, + 0.005406315438449383, + 0.006366275250911713, + 0.005849821958690882, + 0.006300660315901041, + 0.006073503755033016, + 0.005264407955110073, + 0.0056684017181396484, + 0.010535744950175285, + 0.009923707693815231, + 0.0172421932220459, + 0.007661243434995413, + 0.01300356350839138, + 0.010712601244449615, + 0.006709523964673281, + 0.011265008710324764, + 0.01732933335006237, + 0.01879260689020157, + 0.02913322113454342, + 0.013915217481553555, + 0.029102245345711708, + 0.01734992302954197, + 0.019889524206519127, + 0.022837111726403236, + 0.004624619148671627, + 0.005383644253015518, + 0.006231436971575022, + 0.00391381187364459, + 0.006517524365335703, + 0.005871807225048542, + 0.005034301429986954, + 0.006958077196031809, + 0.010256501846015453, + 0.010618661530315876, + 0.014726126566529274, + 0.021792355924844742, + 0.017171630635857582, + 0.01534572895616293, + 0.018957896158099174, + 0.011538519524037838, + 0.019655540585517883, + 0.031556811183691025, + 0.01770511455833912, + 0.019355006515979767, + 0.02010604366660118, + 0.022560659795999527, + 0.014268022030591965, + 0.014240073040127754, + 0.006724679842591286, + 0.010127930901944637, + 0.014048412442207336, + 0.009657141752541065, + 0.01158309355378151, + 0.007145792245864868, + 0.00880003534257412, + 0.01013170462101698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.25995367765426636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297664046287537 + ], + "opposite_logprob_tokens": [ + -2.237976551055908 + ], + "logprob_reward": -0.11297664046287537, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297664046287537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12524811923503876, + 0.13121958076953888, + 0.05400167405605316, + 0.07418591529130936, + 0.07154133170843124, + 0.1546420305967331, + 0.051545996218919754, + 0.15145425498485565, + 0.09497860074043274, + 0.09118252247571945 + ], + "credit_map": [ + 0.022879483178257942, + 0.0163725558668375, + 0.013787108473479748, + 0.010400130413472652, + 0.019469600170850754, + 0.019030869007110596, + 0.01023652870208025, + 0.013071841560304165, + 0.0164144579321146, + 0.01945537142455578, + 0.014252668246626854, + 0.01082679070532322, + 0.017218613997101784, + 0.0184481181204319, + 0.015673065558075905, + 0.01893049292266369, + 0.006011907011270523, + 0.008765277452766895, + 0.006806381046772003, + 0.006044459063559771, + 0.007839547470211983, + 0.007979474030435085, + 0.005590079352259636, + 0.004964548163115978, + 0.009226340800523758, + 0.00899261049926281, + 0.010952172800898552, + 0.007888139225542545, + 0.010204488411545753, + 0.013566112145781517, + 0.005985557101666927, + 0.007370492909103632, + 0.008267248049378395, + 0.009582574479281902, + 0.006793837994337082, + 0.00678799906745553, + 0.01586732640862465, + 0.009141535498201847, + 0.006865994073450565, + 0.00823481660336256, + 0.01887625828385353, + 0.01817220449447632, + 0.029078340157866478, + 0.01223300863057375, + 0.021022304892539978, + 0.016655033454298973, + 0.01490273792296648, + 0.023702146485447884, + 0.006914122030138969, + 0.007554341107606888, + 0.007947721518576145, + 0.003990170080214739, + 0.006771356798708439, + 0.006657535210251808, + 0.003158983075991273, + 0.008551769889891148, + 0.01284305565059185, + 0.015118405222892761, + 0.014557036571204662, + 0.025910846889019012, + 0.023847831413149834, + 0.018272610381245613, + 0.02409769594669342, + 0.016806766390800476, + 0.010010809637606144, + 0.01512661948800087, + 0.013628464192152023, + 0.01309159491211176, + 0.010730529204010963, + 0.011599267832934856, + 0.010709154419600964, + 0.010082162916660309, + 0.010341054759919643, + 0.007087223697453737, + 0.012526360340416431, + 0.01045881025493145, + 0.012677808292210102, + 0.022687146440148354, + 0.006514201406389475, + 0.00888991728425026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.11297664046287537 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801908880472183 + ], + "opposite_logprob_tokens": [ + -3.529801845550537 + ], + "logprob_reward": -0.029801908880472183, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801908880472183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13517992198467255, + 0.06267749518156052, + 0.2017025202512741, + 0.045837342739105225, + 0.07116834074258804, + 0.054477475583553314, + 0.04231896251440048, + 0.24790774285793304, + 0.06619296222925186, + 0.07253725826740265 + ], + "credit_map": [ + 0.01635393127799034, + 0.03471137583255768, + 0.016097398474812508, + 0.007594776339828968, + 0.015466318465769291, + 0.01704179309308529, + 0.018201811239123344, + 0.009712516330182552, + 0.007627222687005997, + 0.007964548654854298, + 0.006723453756421804, + 0.006796042900532484, + 0.008125589229166508, + 0.010790339671075344, + 0.009152092970907688, + 0.005498202051967382, + 0.02119695581495762, + 0.028299806639552116, + 0.02686881273984909, + 0.02016773633658886, + 0.03811866417527199, + 0.031130312010645866, + 0.016251353546977043, + 0.019668854773044586, + 0.007790517061948776, + 0.0052512590773403645, + 0.002929563634097576, + 0.0032742542680352926, + 0.006187914405018091, + 0.014244419522583485, + 0.002758841263130307, + 0.003400574205443263, + 0.006696453783661127, + 0.00984205026179552, + 0.005774553399533033, + 0.007216459605842829, + 0.013079908676445484, + 0.010594786144793034, + 0.008791182190179825, + 0.009172945283353329, + 0.005259059369564056, + 0.006759984884411097, + 0.009112817235291004, + 0.005288316402584314, + 0.008127212524414062, + 0.00586408656090498, + 0.007507409434765577, + 0.00655859149992466, + 0.004224252887070179, + 0.009801287204027176, + 0.004099245183169842, + 0.003434054320678115, + 0.007327921688556671, + 0.006016414146870375, + 0.003280501812696457, + 0.004135288298130035, + 0.021366173401474953, + 0.02225215919315815, + 0.025052962824702263, + 0.04216986522078514, + 0.03918316215276718, + 0.02924152836203575, + 0.038734886795282364, + 0.0299069844186306, + 0.006812736392021179, + 0.010994400829076767, + 0.009346234612166882, + 0.007130886428058147, + 0.008650212548673153, + 0.010283182375133038, + 0.006782076321542263, + 0.006193230859935284, + 0.008185489103198051, + 0.006239569745957851, + 0.008396863006055355, + 0.005626800004392862, + 0.01496477797627449, + 0.018989453092217445, + 0.00446788826957345, + 0.005666421260684729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.029801908880472183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12110432237386703, + 0.045757051557302475, + 0.12117811292409897, + 0.06816127896308899, + 0.048250988125801086, + 0.10763493180274963, + 0.049451254308223724, + 0.22671379148960114, + 0.10071185231208801, + 0.11103634536266327 + ], + "credit_map": [ + 0.014446224085986614, + 0.021305767819285393, + 0.016761858016252518, + 0.00946583691984415, + 0.017174234613776207, + 0.016163254156708717, + 0.012483774684369564, + 0.01330338791012764, + 0.006231305189430714, + 0.0066874101758003235, + 0.00637392932549119, + 0.0051000663079321384, + 0.005605804268270731, + 0.005543336737900972, + 0.004067924804985523, + 0.006147278938442469, + 0.014873955398797989, + 0.01547679491341114, + 0.024823013693094254, + 0.012627220712602139, + 0.01190472673624754, + 0.023028433322906494, + 0.009441454894840717, + 0.00900252815335989, + 0.010814572684466839, + 0.007502682972699404, + 0.00507169496268034, + 0.005242036655545235, + 0.00906292349100113, + 0.019498292356729507, + 0.005957860499620438, + 0.005011215806007385, + 0.004456183407455683, + 0.00571478670462966, + 0.010709733702242374, + 0.005848165601491928, + 0.006221524905413389, + 0.004364992957562208, + 0.006255910266190767, + 0.0046796961687505245, + 0.013207276351749897, + 0.014186165295541286, + 0.01964709907770157, + 0.009815956465899944, + 0.013865119777619839, + 0.010721113532781601, + 0.01054804865270853, + 0.015644172206521034, + 0.004883690737187862, + 0.005629941821098328, + 0.007681088987737894, + 0.004686497151851654, + 0.00766721460968256, + 0.006170935928821564, + 0.005131418351083994, + 0.007600476033985615, + 0.018810387700796127, + 0.019438723102211952, + 0.022640010342001915, + 0.041924361139535904, + 0.033995695412158966, + 0.027213899418711662, + 0.03805266320705414, + 0.024638082832098007, + 0.011212510988116264, + 0.015363194979727268, + 0.012372373603284359, + 0.012752886861562729, + 0.012071776203811169, + 0.014549055136740208, + 0.01236581988632679, + 0.010024248622357845, + 0.009820092469453812, + 0.015211964957416058, + 0.01841862127184868, + 0.013667136430740356, + 0.020677166059613228, + 0.013796316459774971, + 0.010022335685789585, + 0.009422720409929752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292222887277603 + ], + "opposite_logprob_tokens": [ + -2.237922191619873 + ], + "logprob_reward": -0.11292222887277603, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292222887277603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13975967466831207, + 0.09337372332811356, + 0.07999326288700104, + 0.04843563586473465, + 0.05262002348899841, + 0.05174354463815689, + 0.13486559689044952, + 0.2469967007637024, + 0.09514689445495605, + 0.05706487596035004 + ], + "credit_map": [ + 0.022870106622576714, + 0.016559099778532982, + 0.026263032108545303, + 0.014037002809345722, + 0.017284594476222992, + 0.017983827739953995, + 0.011012880131602287, + 0.01374913565814495, + 0.014093443751335144, + 0.013340274803340435, + 0.012490319088101387, + 0.008591988123953342, + 0.010487998835742474, + 0.013129444792866707, + 0.01018370222300291, + 0.011056549847126007, + 0.009893720969557762, + 0.012506358325481415, + 0.00824077520519495, + 0.008883803151547909, + 0.012145747430622578, + 0.01349475048482418, + 0.007309322711080313, + 0.007518785074353218, + 0.006889788433909416, + 0.006269877310842276, + 0.005238341633230448, + 0.0042393081821501255, + 0.006934622768312693, + 0.01155071146786213, + 0.0028859234880656004, + 0.0044270665384829044, + 0.004954506643116474, + 0.006246435921639204, + 0.009774631820619106, + 0.006431678310036659, + 0.007227784488350153, + 0.004711827263236046, + 0.007062919437885284, + 0.006210236810147762, + 0.004577102605253458, + 0.008386852219700813, + 0.008376289159059525, + 0.005422369111329317, + 0.007990951649844646, + 0.005572040099650621, + 0.0055571445263922215, + 0.005860792472958565, + 0.01552158035337925, + 0.032495465129613876, + 0.013556119054555893, + 0.009417467750608921, + 0.021370677277445793, + 0.019240979105234146, + 0.011579087004065514, + 0.011684225872159004, + 0.017380811274051666, + 0.023406831547617912, + 0.02697679027915001, + 0.04840446263551712, + 0.028757084161043167, + 0.02752511575818062, + 0.04424498975276947, + 0.03030061163008213, + 0.012100192718207836, + 0.019016459584236145, + 0.011381225660443306, + 0.009817971847951412, + 0.012217630632221699, + 0.01598997972905636, + 0.007287521380931139, + 0.007335916627198458, + 0.0053748381324112415, + 0.0059004430659115314, + 0.006471348460763693, + 0.00750653725117445, + 0.012256239540874958, + 0.01139773242175579, + 0.004757327493280172, + 0.003400413552299142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.11292222887277603 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a subtle overhead angle, immerses the viewer in a violently boiling pool of incandescent molten lava. The surface of the molten, viscous, orange-red magma churns and swirls, specks of cooling, darkening black scum floating on it, while cracks reveal a brighter, more dazzling yellow light. Vast bubbles rise slowly and menacingly from the depths, pushing the surface of the magma into domes. Then, reaching their limit, they violently burst, hurling countless sparks and strands of molten lava into the air. Shot at a high frame rate, the entire eruption is captured in surreal slow motion, with every drop of molten lava, shimmering with an inner light, traced in the air with exquisite clarity. Meanwhile, thick, sulfurous black and gray smoke billows from the surface of the magma pool, sharply outlined by the bright light from the molten rock below, twisting and spiraling upward. The entire scene is illuminated entirely by the glow of the lava itself, casting a warm yet menacing shimmer of light and shadow, creating an apocalyptic atmosphere that is raw, violent, and awe-inspiring.", + "seed": 29926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_5414_seed_29926_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": -0.07044283337891102, + "vlm_logprob": -0.07044283337891102, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.07044283337891102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1195053681731224 + ], + "opposite_logprob_tokens": [ + -2.2445054054260254 + ], + "logprob_reward": -0.1195053681731224, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1195053681731224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0806322768330574, + 0.09685257822275162, + 0.07811814546585083, + 0.13726449012756348, + 0.10659468173980713, + 0.10404413193464279, + 0.10063757747411728, + 0.03545733541250229, + 0.1725674569606781, + 0.08783134818077087 + ], + "credit_map": [ + 0.010016868822276592, + 0.008161604404449463, + 0.007512349169701338, + 0.008144808001816273, + 0.009329453110694885, + 0.014233354479074478, + 0.009892208501696587, + 0.013341630809009075, + 0.008508848026394844, + 0.010733981616795063, + 0.005110290367156267, + 0.0069282082840800285, + 0.023685326799750328, + 0.016266779974102974, + 0.012553418055176735, + 0.013065723702311516, + 0.010601376183331013, + 0.009300853125751019, + 0.009702184237539768, + 0.007391021121293306, + 0.012138381600379944, + 0.008951381780207157, + 0.01034043449908495, + 0.00969250500202179, + 0.014665715396404266, + 0.012365403585135937, + 0.01120714470744133, + 0.03601118177175522, + 0.015159261412918568, + 0.013446299359202385, + 0.01891133561730385, + 0.015498142689466476, + 0.012397256679832935, + 0.007458721287548542, + 0.01014360599219799, + 0.008484069257974625, + 0.016139818355441093, + 0.015895769000053406, + 0.012310685589909554, + 0.02376476302742958, + 0.015552283264696598, + 0.016285374760627747, + 0.009399528615176678, + 0.011611123569309711, + 0.02119944803416729, + 0.014593271538615227, + 0.008112149313092232, + 0.007290951907634735, + 0.01947648450732231, + 0.02107211947441101, + 0.011040028184652328, + 0.008456718176603317, + 0.01221736241132021, + 0.0107210548594594, + 0.008521542884409428, + 0.00913227442651987, + 0.004427782725542784, + 0.0064667584374547005, + 0.002973063150420785, + 0.0028012346010655165, + 0.004744136705994606, + 0.005089337006211281, + 0.0048012505285441875, + 0.004153773188591003, + 0.020313119515776634, + 0.0308249369263649, + 0.02504759654402733, + 0.020228277891874313, + 0.020456330850720406, + 0.020608222112059593, + 0.013962462544441223, + 0.021126506850123405, + 0.009366560727357864, + 0.014685893431305885, + 0.010184993036091328, + 0.008333086967468262, + 0.013921137899160385, + 0.011447080411016941, + 0.007469725329428911, + 0.012422870844602585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 4 + ], + "selected_reward": -0.1195053681731224 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907424658536911 + ], + "opposite_logprob_tokens": [ + -2.46407413482666 + ], + "logprob_reward": -0.08907424658536911, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907424658536911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08024905622005463, + 0.03977946564555168, + 0.09426964819431305, + 0.026302309706807137, + 0.15240764617919922, + 0.09536454826593399, + 0.09250356256961823, + 0.07002826780080795, + 0.21007853746414185, + 0.13901695609092712 + ], + "credit_map": [ + 0.010733901523053646, + 0.012180688790977001, + 0.004793783184140921, + 0.0059057241305708885, + 0.014970436692237854, + 0.012795155867934227, + 0.007284144405275583, + 0.011585221625864506, + 0.0035099510569125414, + 0.004254122730344534, + 0.0033212376292794943, + 0.003810638329014182, + 0.007967022247612476, + 0.005647958256304264, + 0.005775311961770058, + 0.005493222735822201, + 0.010982914827764034, + 0.010324012488126755, + 0.008955893106758595, + 0.01383369229733944, + 0.009954637847840786, + 0.011846000328660011, + 0.01490604318678379, + 0.013466442935168743, + 0.004767812788486481, + 0.002672194503247738, + 0.003140921937301755, + 0.002903621643781662, + 0.003690643236041069, + 0.003807833418250084, + 0.0025364654138684273, + 0.002782816532999277, + 0.02002672478556633, + 0.012342525646090508, + 0.016565799713134766, + 0.015797480940818787, + 0.025953875854611397, + 0.021149544045329094, + 0.016385165974497795, + 0.02418653294444084, + 0.011701813898980618, + 0.009110468439757824, + 0.012704512104392052, + 0.011835356242954731, + 0.017480716109275818, + 0.0105128837749362, + 0.011293050833046436, + 0.010725744068622589, + 0.013073776848614216, + 0.018884096294641495, + 0.010787851177155972, + 0.008337310515344143, + 0.014494952745735645, + 0.010981610044836998, + 0.008593257516622543, + 0.00735070463269949, + 0.008525438606739044, + 0.010379026643931866, + 0.007356154266744852, + 0.00836940761655569, + 0.009268113411962986, + 0.00957197044044733, + 0.008079090155661106, + 0.008479065261781216, + 0.017899032682180405, + 0.03228726610541344, + 0.026810042560100555, + 0.022928515449166298, + 0.02799180895090103, + 0.028451010584831238, + 0.02331913448870182, + 0.030391719192266464, + 0.018496409058570862, + 0.03383483737707138, + 0.010877888649702072, + 0.011057968251407146, + 0.019064510241150856, + 0.01931394264101982, + 0.012432335875928402, + 0.01393906306475401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.08907424658536911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048652395606040955 + ], + "opposite_logprob_tokens": [ + -3.048652410507202 + ], + "logprob_reward": -0.048652395606040955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048652395606040955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07520534843206406, + 0.12816503643989563, + 0.028301332145929337, + 0.06264423578977585, + 0.09692957997322083, + 0.22959083318710327, + 0.10398466885089874, + 0.09547725319862366, + 0.06975863873958588, + 0.10994308441877365 + ], + "credit_map": [ + 0.007326584775000811, + 0.010492848232388496, + 0.0065728710032999516, + 0.008609694428741932, + 0.007969233207404613, + 0.010661371983587742, + 0.010365109890699387, + 0.01320762000977993, + 0.012468292377889156, + 0.009611902758479118, + 0.009929295629262924, + 0.012903903611004353, + 0.015972299501299858, + 0.0169161856174469, + 0.022978467866778374, + 0.027384670451283455, + 0.00445443345233798, + 0.0020573607180267572, + 0.0030801875982433558, + 0.0027893262449651957, + 0.004269953817129135, + 0.0036147162318229675, + 0.004028089344501495, + 0.004007259849458933, + 0.006494036875665188, + 0.006026310846209526, + 0.00596899027004838, + 0.014544318430125713, + 0.00749828340485692, + 0.007191050797700882, + 0.008250916376709938, + 0.006670320872217417, + 0.010653727687895298, + 0.007856743410229683, + 0.012184263207018375, + 0.015692738816142082, + 0.016446705907583237, + 0.011351614259183407, + 0.011878808960318565, + 0.010864967480301857, + 0.03698140010237694, + 0.03004448674619198, + 0.02592872641980648, + 0.027310267090797424, + 0.047795798629522324, + 0.029031595215201378, + 0.017786363139748573, + 0.014712179079651833, + 0.012805243022739887, + 0.02113047055900097, + 0.009674099273979664, + 0.0078378701582551, + 0.021474873647093773, + 0.013671793043613434, + 0.010663500986993313, + 0.006726808380335569, + 0.013214418664574623, + 0.022092357277870178, + 0.011397325433790684, + 0.008049731142818928, + 0.009664330631494522, + 0.012444349005818367, + 0.007741512730717659, + 0.0108732208609581, + 0.006153061054646969, + 0.013375453650951385, + 0.008556708693504333, + 0.009211127646267414, + 0.010129219852387905, + 0.007455108221620321, + 0.008217009715735912, + 0.006660941988229752, + 0.01437520794570446, + 0.025561334565281868, + 0.009398783557116985, + 0.008939850144088268, + 0.01592724770307541, + 0.015137487091124058, + 0.007505720481276512, + 0.01309744082391262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.048652395606040955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952725186944008 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952725186944008, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952725186944008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097601056098938, + 0.06751230359077454, + 0.0401502400636673, + 0.1771203577518463, + 0.10341144353151321, + 0.2132650762796402, + 0.066185861825943, + 0.07100311666727066, + 0.057936493307352066, + 0.09365501254796982 + ], + "credit_map": [ + 0.010689180344343185, + 0.02182641066610813, + 0.008665149100124836, + 0.015011087991297245, + 0.01141715981066227, + 0.014572287909686565, + 0.011868476867675781, + 0.015710333362221718, + 0.008457750082015991, + 0.0072878203354775906, + 0.005869436077773571, + 0.008488763123750687, + 0.013272511772811413, + 0.008717759512364864, + 0.005778891034424305, + 0.009639366529881954, + 0.007817110978066921, + 0.005513777956366539, + 0.004501822870224714, + 0.004050752613693476, + 0.004771646577864885, + 0.0048961639404296875, + 0.00417779665440321, + 0.0044211638160049915, + 0.018506914377212524, + 0.017461242154240608, + 0.01598436012864113, + 0.04621008783578873, + 0.018477411940693855, + 0.018281016498804092, + 0.02194388210773468, + 0.020255422219634056, + 0.014181624166667461, + 0.010379079729318619, + 0.00912762712687254, + 0.010153068229556084, + 0.013499762862920761, + 0.015861691907048225, + 0.013573565520346165, + 0.016635019332170486, + 0.033937450498342514, + 0.029283346608281136, + 0.023480607196688652, + 0.029304608702659607, + 0.03708653524518013, + 0.022061483934521675, + 0.020599128678441048, + 0.017511913552880287, + 0.008663767017424107, + 0.01138648483902216, + 0.008880200795829296, + 0.006476256065070629, + 0.010535286739468575, + 0.0068319011479616165, + 0.007147281430661678, + 0.006264678668230772, + 0.011526785790920258, + 0.01366477645933628, + 0.006774516776204109, + 0.00701400451362133, + 0.007001686841249466, + 0.008408415131270885, + 0.0072045400738716125, + 0.009408383630216122, + 0.005611972417682409, + 0.006857144180685282, + 0.006438412703573704, + 0.007604492828249931, + 0.007217913866043091, + 0.008290300145745277, + 0.0066123828291893005, + 0.009303868748247623, + 0.01353779248893261, + 0.015519443899393082, + 0.01221771165728569, + 0.009049713611602783, + 0.01085523422807455, + 0.01449625100940466, + 0.010927880182862282, + 0.007050974294543266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.005952725186944008 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902943134307861 + ], + "opposite_logprob_tokens": [ + -2.464029312133789 + ], + "logprob_reward": -0.08902943134307861, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902943134307861, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10095234960317612, + 0.07202807068824768, + 0.10134653002023697, + 0.07306288182735443, + 0.14571839570999146, + 0.06691458821296692, + 0.08373576402664185, + 0.08961740881204605, + 0.12061487138271332, + 0.146009162068367 + ], + "credit_map": [ + 0.010949971154332161, + 0.01340207364410162, + 0.009921126998960972, + 0.01060726772993803, + 0.011084373109042645, + 0.017284754663705826, + 0.011899183504283428, + 0.01580359973013401, + 0.009281699545681477, + 0.005690309219062328, + 0.005069824866950512, + 0.006537190638482571, + 0.013068979606032372, + 0.009854946285486221, + 0.010959407314658165, + 0.011565710417926311, + 0.014587977901101112, + 0.011347095482051373, + 0.011038461700081825, + 0.014392364770174026, + 0.013058825396001339, + 0.01075362041592598, + 0.01513579674065113, + 0.01103238295763731, + 0.00811304897069931, + 0.006718261633068323, + 0.006985270418226719, + 0.01957220397889614, + 0.008364874869585037, + 0.007098384667187929, + 0.009890389628708363, + 0.006320447660982609, + 0.02564014308154583, + 0.013531336560845375, + 0.01425281073898077, + 0.013951113447546959, + 0.021771680563688278, + 0.02026131935417652, + 0.016520613804459572, + 0.019789395853877068, + 0.010131636634469032, + 0.0092640221118927, + 0.00660680141299963, + 0.008100880309939384, + 0.014071318320930004, + 0.007644108030945063, + 0.005914667155593634, + 0.0051811495795845985, + 0.011423001997172832, + 0.014453991316258907, + 0.00878109410405159, + 0.008029659278690815, + 0.014289471320807934, + 0.0125137222930789, + 0.007879172451794147, + 0.006365646608173847, + 0.011978761292994022, + 0.01893833838403225, + 0.010580415837466717, + 0.008857340551912785, + 0.009154635481536388, + 0.01039282325655222, + 0.008774993009865284, + 0.01094009354710579, + 0.011730252765119076, + 0.02422540821135044, + 0.012330632656812668, + 0.010078530758619308, + 0.022889677435159683, + 0.01894395239651203, + 0.008745532482862473, + 0.011670880950987339, + 0.017947854474186897, + 0.032502442598342896, + 0.011028899811208248, + 0.013490709476172924, + 0.01971304416656494, + 0.02103537507355213, + 0.012678094208240509, + 0.017612747848033905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.08902943134307861 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, captured from a subtle overhead angle, immerses the viewer in a violently boiling pool of incandescent molten lava. The surface of the molten, viscous, orange-red magma churns and swirls, specks of cooling, darkening black scum floating on it, while cracks reveal a brighter, more dazzling yellow light. Vast bubbles rise slowly and menacingly from the depths, pushing the surface of the magma into domes. Then, reaching their limit, they violently burst, hurling countless sparks and strands of molten lava into the air. Shot at a high frame rate, the entire eruption is captured in surreal slow motion, with every drop of molten lava, shimmering with an inner light, traced in the air with exquisite clarity. Meanwhile, thick, sulfurous black and gray smoke billows from the surface of the magma pool, sharply outlined by the bright light from the molten rock below, twisting and spiraling upward. The entire scene is illuminated entirely by the glow of the lava itself, casting a warm yet menacing shimmer of light and shadow, creating an apocalyptic atmosphere that is raw, violent, and awe-inspiring.", + "seed": 29927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_14/index_5414_seed_29927_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": -0.1893251646310091, + "vlm_logprob": -0.1893251646310091, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.1893251646310091, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.636698305606842 + ], + "opposite_logprob_tokens": [ + -0.761698305606842 + ], + "logprob_reward": -0.636698305606842, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.636698305606842, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26539069414138794, + 0.14202935993671417, + 0.10181605070829391, + 0.032576341181993484, + 0.18950168788433075, + 0.0745956301689148, + 0.08531086146831512, + 0.02675670199096203, + 0.03689206391572952, + 0.045130617916584015 + ], + "credit_map": [ + 0.05027168244123459, + 0.0920504778623581, + 0.018634648993611336, + 0.013160309754312038, + 0.03136176988482475, + 0.029246192425489426, + 0.012997875921428204, + 0.017667753621935844, + 0.019409721717238426, + 0.019567664712667465, + 0.01713639684021473, + 0.012447908520698547, + 0.020641593262553215, + 0.02753561921417713, + 0.011439179070293903, + 0.013851293362677097, + 0.01147446595132351, + 0.0106532983481884, + 0.01982925646007061, + 0.015075105242431164, + 0.011339101940393448, + 0.010778553783893585, + 0.010727052576839924, + 0.01193921733647585, + 0.003927886951714754, + 0.004598379135131836, + 0.0046910918317735195, + 0.004375773016363382, + 0.004091316368430853, + 0.003171417396515608, + 0.0041578239761292934, + 0.003562650876119733, + 0.03465403616428375, + 0.018582122400403023, + 0.02190924435853958, + 0.01431693509221077, + 0.022199777886271477, + 0.04652392119169235, + 0.01278606429696083, + 0.01852959766983986, + 0.009475566446781158, + 0.007345329504460096, + 0.014214136637747288, + 0.009416342712938786, + 0.008046937175095081, + 0.008702083490788937, + 0.007352492306381464, + 0.010042749345302582, + 0.014980634674429893, + 0.007513845805078745, + 0.011739936657249928, + 0.007583512458950281, + 0.011218991130590439, + 0.014621442183852196, + 0.0060709817335009575, + 0.01158151775598526, + 0.003987429663538933, + 0.0024695349857211113, + 0.005073040723800659, + 0.003015973139554262, + 0.0024531276430934668, + 0.003482447937130928, + 0.0023781040217727423, + 0.0038970429450273514, + 0.003262173617258668, + 0.007154638413339853, + 0.0064050909131765366, + 0.003775673918426037, + 0.004614999983459711, + 0.004581244196742773, + 0.00324401562102139, + 0.0038542274851351976, + 0.008107485249638557, + 0.004807445220649242, + 0.004675508942455053, + 0.004269553814083338, + 0.00551253417506814, + 0.01102942693978548, + 0.0026162846479564905, + 0.004112381022423506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.636698305606842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285822212696075 + ], + "opposite_logprob_tokens": [ + -2.0178582668304443 + ], + "logprob_reward": -0.14285822212696075, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14285822212696075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28201860189437866, + 0.11675812304019928, + 0.05763452500104904, + 0.038894135504961014, + 0.0901811420917511, + 0.03962724655866623, + 0.07491336017847061, + 0.13450004160404205, + 0.13283076882362366, + 0.03264200687408447 + ], + "credit_map": [ + 0.05588453263044357, + 0.10361889749765396, + 0.015406960621476173, + 0.013262316584587097, + 0.0333520844578743, + 0.030769234523177147, + 0.012824292294681072, + 0.016900286078453064, + 0.01621345616877079, + 0.015782639384269714, + 0.013798223808407784, + 0.010447564534842968, + 0.015179190784692764, + 0.02383705973625183, + 0.010198035277426243, + 0.011301958933472633, + 0.006681221071630716, + 0.007308445405215025, + 0.008937732316553593, + 0.007599385920912027, + 0.008010167628526688, + 0.007373461965471506, + 0.005853889975696802, + 0.00587021978572011, + 0.006006880197674036, + 0.006069319322705269, + 0.005810640752315521, + 0.004841807298362255, + 0.005434667691588402, + 0.0033394719939678907, + 0.0038477913476526737, + 0.003543558530509472, + 0.02027190290391445, + 0.0077914223074913025, + 0.007877674885094166, + 0.009391858242452145, + 0.008316842839121819, + 0.022252939641475677, + 0.007704176940023899, + 0.006574327126145363, + 0.006138649769127369, + 0.010013596154749393, + 0.0052381581626832485, + 0.004557773470878601, + 0.004128478933125734, + 0.00385641329921782, + 0.0028737313114106655, + 0.002820447087287903, + 0.012410040013492107, + 0.007235449738800526, + 0.013501999899744987, + 0.009768655523657799, + 0.009581193327903748, + 0.011119309812784195, + 0.004624709486961365, + 0.006672002840787172, + 0.017415247857570648, + 0.012203628197312355, + 0.02302212454378605, + 0.01974988915026188, + 0.01238230150192976, + 0.01823425479233265, + 0.010653058998286724, + 0.020839527249336243, + 0.016614122316241264, + 0.023264767602086067, + 0.018297327682375908, + 0.015147155150771141, + 0.023635370656847954, + 0.017483536154031754, + 0.0077771772630512714, + 0.010611315257847309, + 0.00485065346583724, + 0.003934208769351244, + 0.0035980541724711657, + 0.002933724783360958, + 0.0037796352989971638, + 0.008303288370370865, + 0.002215478802099824, + 0.0030269636772572994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.14285822212696075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306305572390556 + ], + "opposite_logprob_tokens": [ + -3.168063163757324 + ], + "logprob_reward": -0.04306305572390556, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306305572390556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1838453859090805, + 0.07237221300601959, + 0.1909564733505249, + 0.19097359478473663, + 0.10609042644500732, + 0.04487022012472153, + 0.05204339697957039, + 0.0758768692612648, + 0.042899761348962784, + 0.040071651339530945 + ], + "credit_map": [ + 0.03612380102276802, + 0.053710997104644775, + 0.018222499638795853, + 0.012851018458604813, + 0.016936274245381355, + 0.02355223521590233, + 0.008484059013426304, + 0.013964510522782803, + 0.010106861591339111, + 0.010117178782820702, + 0.008703251369297504, + 0.006295762024819851, + 0.009080111980438232, + 0.01490883156657219, + 0.00636378675699234, + 0.006796430796384811, + 0.021551644429564476, + 0.01960849203169346, + 0.036653757095336914, + 0.029797155410051346, + 0.01972969062626362, + 0.02088613249361515, + 0.019411735236644745, + 0.023317880928516388, + 0.026918720453977585, + 0.044303830713033676, + 0.022499866783618927, + 0.02131536230444908, + 0.037091437727212906, + 0.014287418685853481, + 0.014701168052852154, + 0.009855794720351696, + 0.015085568651556969, + 0.01581825502216816, + 0.012166428379714489, + 0.011853515170514584, + 0.008375911973416805, + 0.022868096828460693, + 0.009218281134963036, + 0.01070436555892229, + 0.006080000661313534, + 0.0067705512046813965, + 0.0072580683045089245, + 0.005272327456623316, + 0.005220744758844376, + 0.004287723917514086, + 0.005170282442122698, + 0.004810522310435772, + 0.006685555446892977, + 0.005667493212968111, + 0.006783340126276016, + 0.005320928059518337, + 0.006623474415391684, + 0.008174238726496696, + 0.004381795413792133, + 0.008406572043895721, + 0.011934343725442886, + 0.00820731371641159, + 0.010043867863714695, + 0.00666297972202301, + 0.008807055652141571, + 0.013360990211367607, + 0.005606540944427252, + 0.011253776960074902, + 0.005790242459625006, + 0.008321401663124561, + 0.005386422388255596, + 0.004691338166594505, + 0.0056482115760445595, + 0.008152189664542675, + 0.00216094427742064, + 0.002749013714492321, + 0.005370700731873512, + 0.007472441531717777, + 0.005210400093346834, + 0.004582589492201805, + 0.006333036348223686, + 0.00505776097998023, + 0.002499116351827979, + 0.003545604180544615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04306305572390556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108134351670742 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.01108134351670742, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108134351670742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18207770586013794, + 0.08667220920324326, + 0.06720580160617828, + 0.07011718302965164, + 0.1597082018852234, + 0.05473377928137779, + 0.06819745153188705, + 0.061652638018131256, + 0.11018869280815125, + 0.13944631814956665 + ], + "credit_map": [ + 0.032988160848617554, + 0.05044092237949371, + 0.011929105967283249, + 0.010615522041916847, + 0.01951216161251068, + 0.03163917735219002, + 0.012402660213410854, + 0.01254997868090868, + 0.00953669659793377, + 0.011597205884754658, + 0.009720667265355587, + 0.010436946526169777, + 0.01777147315442562, + 0.00944898184388876, + 0.010079213418066502, + 0.008081028237938881, + 0.007518674712628126, + 0.008580282330513, + 0.011510409414768219, + 0.009128623642027378, + 0.009127771481871605, + 0.006574063561856747, + 0.0070776077918708324, + 0.0076883714646101, + 0.010005209594964981, + 0.013363602571189404, + 0.010083760134875774, + 0.007743325084447861, + 0.011303046718239784, + 0.005865234881639481, + 0.00691711250692606, + 0.004835893400013447, + 0.02225366421043873, + 0.024416325613856316, + 0.024562345817685127, + 0.014793464913964272, + 0.018731608986854553, + 0.02805294096469879, + 0.011417618952691555, + 0.015480229631066322, + 0.007542557083070278, + 0.010682472959160805, + 0.006875182501971722, + 0.0053420718759298325, + 0.007192247547209263, + 0.005281445570290089, + 0.006939387414604425, + 0.004878414329141378, + 0.015260636806488037, + 0.007901476696133614, + 0.008981356397271156, + 0.006885233335196972, + 0.009028133936226368, + 0.008177404291927814, + 0.0054358504712581635, + 0.006527351215481758, + 0.0082078967243433, + 0.005185778718441725, + 0.0073809255845844746, + 0.007362872827798128, + 0.005165103357285261, + 0.010190777480602264, + 0.005592484027147293, + 0.01256680116057396, + 0.010942709632217884, + 0.019177012145519257, + 0.014491644687950611, + 0.01236662920564413, + 0.02055731974542141, + 0.01406679954379797, + 0.007672338746488094, + 0.010914240032434464, + 0.020663555711507797, + 0.021955937147140503, + 0.01352479588240385, + 0.011204729788005352, + 0.02080349810421467, + 0.026242343708872795, + 0.011237798258662224, + 0.013813657686114311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.01108134351670742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 pool of molten lava?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle positioned subtly overhead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are dark black scum specks visible floating on the surface of the magma?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do vast bubbles rise from the depths and burst violently to hurl sparks and lava strands into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured in slow motion with high frame rate clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292489618062973 + ], + "opposite_logprob_tokens": [ + -2.237924814224243 + ], + "logprob_reward": -0.11292489618062973, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292489618062973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15200990438461304, + 0.11651396006345749, + 0.07392307370901108, + 0.08002261072397232, + 0.18638451397418976, + 0.11341182142496109, + 0.07779821753501892, + 0.06382802128791809, + 0.05613749846816063, + 0.0799703523516655 + ], + "credit_map": [ + 0.021179446950554848, + 0.015490137040615082, + 0.021569540724158287, + 0.016451578587293625, + 0.014799762517213821, + 0.035538896918296814, + 0.010948033072054386, + 0.0160325039178133, + 0.01772978901863098, + 0.0139927351847291, + 0.015043983235955238, + 0.0127872908487916, + 0.009654105640947819, + 0.025238102301955223, + 0.009835949167609215, + 0.012232008390128613, + 0.00889680627733469, + 0.007084238342940807, + 0.012714598327875137, + 0.011425124481320381, + 0.009201524779200554, + 0.01021196972578764, + 0.0073020607233047485, + 0.007086750119924545, + 0.012176203541457653, + 0.011353038251399994, + 0.009959935210645199, + 0.010600118897855282, + 0.009947918355464935, + 0.00870394054800272, + 0.009350763633847237, + 0.007930691353976727, + 0.019172577187418938, + 0.03772922605276108, + 0.019387967884540558, + 0.024685880169272423, + 0.016842257231473923, + 0.024929670616984367, + 0.019804520532488823, + 0.023832421749830246, + 0.014520804397761822, + 0.019432250410318375, + 0.012430576607584953, + 0.016273552551865578, + 0.011213093064725399, + 0.016584092751145363, + 0.012757303193211555, + 0.010200145654380322, + 0.01582169160246849, + 0.008215264417231083, + 0.008535443805158138, + 0.007387430407106876, + 0.008745360188186169, + 0.010722092352807522, + 0.006608950439840555, + 0.011761987581849098, + 0.007888799533247948, + 0.005294805392622948, + 0.00942486897110939, + 0.007481968030333519, + 0.006146897096186876, + 0.010460419580340385, + 0.004998814780265093, + 0.01213145162910223, + 0.006370349321514368, + 0.008050392381846905, + 0.008688423782587051, + 0.007840973325073719, + 0.010146621614694595, + 0.00682243425399065, + 0.003552108071744442, + 0.0046661971136927605, + 0.015051026828587055, + 0.008173546753823757, + 0.009218184277415276, + 0.005789336282759905, + 0.00794240739196539, + 0.024199353531003, + 0.004617139231413603, + 0.004979361314326525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.11292489618062973 + } + ], + "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/0000152/group_15/index_tensor([30, 30])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_tensor([30, 30])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fce22d79478ac49300d1194e2fcf5078cb7475a1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_tensor([30, 30])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with a sleek, steady camera, this cinematic close-up features a young woman jogging effortlessly through a sun-drenched park. Her chestnut hair is pulled back into a high, bouncing ponytail, with a few locks framing her exercise-flushed face. She wears a pair of stylish, minimalist white wireless earbuds, which fit securely and comfortably in her ears. A bright, happy smile illuminates her features, and her eyes crinkle at the corners as she loses herself in the music. The camera focuses firmly on her joyful expression as she moves with steady, rhythmic grace. The background melts into a soft-focus blur of the park's emerald green lawn and tall, leafy trees, all bathed in the warm golden glow of the afternoon sun, creating a beautiful, soft lens flare. The entire scene is imbued with an energetic, uplifting atmosphere of quiet personal freedom and pure bliss.", + "seed": 20838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_30_seed_20838_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": -0.031044705491513013, + "vlm_logprob": -0.031044705491513013, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.031044705491513013, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.060891278088092804 + ], + "opposite_logprob_tokens": [ + -2.9358913898468018 + ], + "logprob_reward": -0.060891278088092804, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060891278088092804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10736209899187088, + 0.04780948534607887, + 0.06741345673799515, + 0.1029086783528328, + 0.026970993727445602, + 0.034249722957611084, + 0.0584585927426815, + 0.0219324491918087, + 0.4993843734264374, + 0.033510103821754456 + ], + "credit_map": [ + 0.01135991606861353, + 0.012219395488500595, + 0.014750375412404537, + 0.020491531118750572, + 0.015372330322861671, + 0.013444003649055958, + 0.013485032133758068, + 0.0062395245768129826, + 0.006601613946259022, + 0.0035286908969283104, + 0.005910203792154789, + 0.003146137809380889, + 0.009243638254702091, + 0.005202016327530146, + 0.008130159229040146, + 0.006047032307833433, + 0.00763119850307703, + 0.008280646987259388, + 0.0071651223115623, + 0.008192186243832111, + 0.010646169073879719, + 0.008540579117834568, + 0.00913927424699068, + 0.00781828910112381, + 0.009379088878631592, + 0.014339710585772991, + 0.01348841655999422, + 0.010449045337736607, + 0.0064770132303237915, + 0.029511533677577972, + 0.008382786065340042, + 0.010881084948778152, + 0.003808731911703944, + 0.0026489945594221354, + 0.0032141529954969883, + 0.002225967589765787, + 0.005414353683590889, + 0.0027004063595086336, + 0.004268491640686989, + 0.002689896384254098, + 0.005114040803164244, + 0.003916253335773945, + 0.0057450151070952415, + 0.004951357841491699, + 0.004141971468925476, + 0.0024672294966876507, + 0.003577374154701829, + 0.004336484242230654, + 0.008907762356102467, + 0.005088349338620901, + 0.008607377298176289, + 0.004993784241378307, + 0.005833073519170284, + 0.01167357712984085, + 0.005777816753834486, + 0.007576858624815941, + 0.0026623536832630634, + 0.0025353229139000177, + 0.0025685965083539486, + 0.0022650035098195076, + 0.0030143409967422485, + 0.0034077914897352457, + 0.0028875013813376427, + 0.002591540105640888, + 0.04057889059185982, + 0.021148603409528732, + 0.11211694777011871, + 0.036537691950798035, + 0.0320790633559227, + 0.08460605889558792, + 0.02922518365085125, + 0.14309199154376984, + 0.003899468807503581, + 0.0028948779217898846, + 0.004127216525375843, + 0.0037113854195922613, + 0.005136560183018446, + 0.003494631964713335, + 0.005682039074599743, + 0.0045639281161129475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.060891278088092804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.007656395435333252 + ], + "opposite_logprob_tokens": [ + -4.882656574249268 + ], + "logprob_reward": -0.007656395435333252, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007656395435333252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12603463232517242, + 0.08301178365945816, + 0.2680501639842987, + 0.11026367545127869, + 0.020274974405765533, + 0.1029016450047493, + 0.026795728132128716, + 0.02113344706594944, + 0.11825791001319885, + 0.12327605485916138 + ], + "credit_map": [ + 0.018649641424417496, + 0.011177405714988708, + 0.02837250381708145, + 0.02124423161149025, + 0.009597073309123516, + 0.011341597884893417, + 0.014390747994184494, + 0.011261425912380219, + 0.013936974108219147, + 0.005777670070528984, + 0.015337448567152023, + 0.007217284291982651, + 0.0066343327052891254, + 0.011817923747003078, + 0.012875447049736977, + 0.009414706379175186, + 0.06886958330869675, + 0.019836552441120148, + 0.09466402977705002, + 0.014773637987673283, + 0.016260594129562378, + 0.017078151926398277, + 0.025350872427225113, + 0.011216752231121063, + 0.006353011820465326, + 0.010464801453053951, + 0.02341855876147747, + 0.011327655054628849, + 0.0055764466524124146, + 0.036092955619096756, + 0.007128231693059206, + 0.009902018122375011, + 0.003010132582858205, + 0.0017952237976714969, + 0.002167025115340948, + 0.00329137546941638, + 0.003682794515043497, + 0.001904626959003508, + 0.003047793172299862, + 0.0013760019792243838, + 0.018630586564540863, + 0.017228424549102783, + 0.010890508070588112, + 0.008540628477931023, + 0.018480604514479637, + 0.008350365795195103, + 0.011767018586397171, + 0.009013504721224308, + 0.003248421009629965, + 0.0018387292511761189, + 0.0027761822566390038, + 0.0026930710300803185, + 0.0028335563838481903, + 0.004448048770427704, + 0.004545436706393957, + 0.004412280395627022, + 0.003457730170339346, + 0.0018925246549770236, + 0.002636412624269724, + 0.0020825681276619434, + 0.0028834277763962746, + 0.003043253207579255, + 0.0029534802306443453, + 0.0021840508561581373, + 0.010501793585717678, + 0.004571941215544939, + 0.028649453073740005, + 0.007751591969281435, + 0.00914745219051838, + 0.02056991495192051, + 0.0065324632450938225, + 0.03053329326212406, + 0.013329927809536457, + 0.010966625064611435, + 0.011799447238445282, + 0.011339359916746616, + 0.014086917974054813, + 0.02822938561439514, + 0.015287618152797222, + 0.01823677122592926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.007656395435333252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.062052879482507706 + ], + "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": [ + 0.09720391780138016, + 0.05410340800881386, + 0.14228279888629913, + 0.07578226178884506, + 0.01551816239953041, + 0.025862757116556168, + 0.02844424545764923, + 0.0220208577811718, + 0.4938463270664215, + 0.044935230165719986 + ], + "credit_map": [ + 0.010029435157775879, + 0.008175816386938095, + 0.021294720470905304, + 0.013977942056953907, + 0.010246657766401768, + 0.008929156698286533, + 0.012598187662661076, + 0.011951997876167297, + 0.012167369946837425, + 0.004523367155343294, + 0.01339650247246027, + 0.004356238525360823, + 0.0061996085569262505, + 0.0036524110473692417, + 0.006507319398224354, + 0.0033005853183567524, + 0.029740264639258385, + 0.011169115081429482, + 0.04399876669049263, + 0.012502743862569332, + 0.012450999580323696, + 0.010770872235298157, + 0.011875662952661514, + 0.009774378500878811, + 0.004005143418908119, + 0.006726445164531469, + 0.019125960767269135, + 0.005380573682487011, + 0.005739456042647362, + 0.02323833853006363, + 0.0058400011621415615, + 0.005726340692490339, + 0.0023348466493189335, + 0.001360742375254631, + 0.0016600266098976135, + 0.002175972331315279, + 0.002831540536135435, + 0.0014389599673449993, + 0.002515847096219659, + 0.0012002272997051477, + 0.003497922560200095, + 0.0028974537272006273, + 0.003154121106490493, + 0.0030116229318082333, + 0.004742852412164211, + 0.003055847715586424, + 0.003033296437934041, + 0.002469640225172043, + 0.0037092326674610376, + 0.0029198448173701763, + 0.002233210951089859, + 0.00213409261777997, + 0.004048710223287344, + 0.004559711087495089, + 0.0060143619775772095, + 0.002825080882757902, + 0.0016862541669979692, + 0.0023340980987995863, + 0.0015603473875671625, + 0.001826380961574614, + 0.00521583529189229, + 0.00343350600451231, + 0.004373512696474791, + 0.0015909228241071105, + 0.041311051696538925, + 0.019081220030784607, + 0.11730354279279709, + 0.031481772661209106, + 0.031686242669820786, + 0.08366373181343079, + 0.030646393075585365, + 0.13867239654064178, + 0.005046682897955179, + 0.004051019437611103, + 0.003936546389013529, + 0.005355656612664461, + 0.004159120377153158, + 0.013271168805658817, + 0.0028292215429246426, + 0.006285814568400383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.062052879482507706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0013372297398746014 + ], + "opposite_logprob_tokens": [ + -6.626337051391602 + ], + "logprob_reward": -0.0013372297398746014, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013372297398746014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09827495366334915, + 0.050018154084682465, + 0.1226525753736496, + 0.059069279581308365, + 0.043346937745809555, + 0.11285620182752609, + 0.04972529038786888, + 0.021025046706199646, + 0.39581912755966187, + 0.047212373465299606 + ], + "credit_map": [ + 0.013040540739893913, + 0.009127859957516193, + 0.01885334961116314, + 0.021096227690577507, + 0.008175082504749298, + 0.008359619416296482, + 0.012861779890954494, + 0.006760501768440008, + 0.007430775091052055, + 0.0035057850182056427, + 0.009339526295661926, + 0.002758990041911602, + 0.004914112389087677, + 0.00798037089407444, + 0.004475680645555258, + 0.009612919762730598, + 0.032809361815452576, + 0.00772560341283679, + 0.04426438733935356, + 0.00831808615475893, + 0.007581653073430061, + 0.007226174231618643, + 0.009474305436015129, + 0.005253009032458067, + 0.00678080040961504, + 0.007748853415250778, + 0.007748584263026714, + 0.00778107438236475, + 0.007418776396661997, + 0.00783457700163126, + 0.0073041110299527645, + 0.006452508270740509, + 0.005633176304399967, + 0.00425147358328104, + 0.005519969388842583, + 0.006800089497119188, + 0.008005809038877487, + 0.004659696482121944, + 0.0054416717030107975, + 0.0030350571032613516, + 0.017988750711083412, + 0.01652236469089985, + 0.013531866483390331, + 0.013442687690258026, + 0.017733439803123474, + 0.008166727609932423, + 0.012571541592478752, + 0.012898837216198444, + 0.00641495268791914, + 0.003990734927356243, + 0.006606791168451309, + 0.0073897261172533035, + 0.0045211114920675755, + 0.00676942290738225, + 0.00640344200655818, + 0.007629117928445339, + 0.002511924598366022, + 0.0021080845035612583, + 0.0016988249262794852, + 0.0025334053207188845, + 0.003770160023123026, + 0.0027331330347806215, + 0.0037386605981737375, + 0.001930857077240944, + 0.03294369950890541, + 0.015530873090028763, + 0.09522929042577744, + 0.027669336646795273, + 0.026652716100215912, + 0.06482335925102234, + 0.02482689544558525, + 0.10814300924539566, + 0.005784981418401003, + 0.00479923514649272, + 0.004505406599491835, + 0.004323289729654789, + 0.0072894636541605, + 0.0051517426036298275, + 0.00848458707332611, + 0.0068736691027879715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.0013372297398746014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.023285744711756706 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285744711756706, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285744711756706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22847646474838257, + 0.07715414464473724, + 0.19426749646663666, + 0.08434852957725525, + 0.03579824045300484, + 0.09288118779659271, + 0.02582641690969467, + 0.08652011305093765, + 0.10614659637212753, + 0.06858080625534058 + ], + "credit_map": [ + 0.02428102120757103, + 0.022542636841535568, + 0.05367434397339821, + 0.03638594597578049, + 0.022489270195364952, + 0.02169230207800865, + 0.026712223887443542, + 0.02069871313869953, + 0.018900398164987564, + 0.005324181169271469, + 0.019778776913881302, + 0.006047786679118872, + 0.006478739436715841, + 0.006439458578824997, + 0.008136965334415436, + 0.006047838367521763, + 0.05627143010497093, + 0.01110618282109499, + 0.0790044367313385, + 0.011089500039815903, + 0.006454464979469776, + 0.008392751216888428, + 0.014882589690387249, + 0.007066132966428995, + 0.009890136308968067, + 0.009862547740340233, + 0.006779825314879417, + 0.010197112336754799, + 0.007909144274890423, + 0.020352734252810478, + 0.007802930194884539, + 0.011554098688066006, + 0.004833436571061611, + 0.0031903591006994247, + 0.003962590359151363, + 0.0038965714629739523, + 0.0073206983506679535, + 0.003853973001241684, + 0.005522776395082474, + 0.0032178342808037996, + 0.011951540596783161, + 0.010540236718952656, + 0.01187895704060793, + 0.011557742953300476, + 0.018688756972551346, + 0.006593287922441959, + 0.012490203604102135, + 0.009180462919175625, + 0.0037167761474847794, + 0.002312025520950556, + 0.0036178473383188248, + 0.002802192931994796, + 0.0037944363430142403, + 0.003030376974493265, + 0.0039600529707968235, + 0.0025927089154720306, + 0.011597385630011559, + 0.0084829805418849, + 0.008037040010094643, + 0.008742018602788448, + 0.01851755753159523, + 0.009078223258256912, + 0.013175525702536106, + 0.00888938456773758, + 0.013026176020503044, + 0.00525858486071229, + 0.02570182830095291, + 0.00823263917118311, + 0.00879720039665699, + 0.015696261078119278, + 0.00814441405236721, + 0.021289493888616562, + 0.007347800303250551, + 0.006931958720088005, + 0.008441390469670296, + 0.009894609451293945, + 0.009936300106346607, + 0.009026640094816685, + 0.008744603954255581, + 0.00825749896466732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.023285744711756706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a sleek, steady camera, this cinematic close-up features a young woman jogging effortlessly through a sun-drenched park. Her chestnut hair is pulled back into a high, bouncing ponytail, with a few locks framing her exercise-flushed face. She wears a pair of stylish, minimalist white wireless earbuds, which fit securely and comfortably in her ears. A bright, happy smile illuminates her features, and her eyes crinkle at the corners as she loses herself in the music. The camera focuses firmly on her joyful expression as she moves with steady, rhythmic grace. The background melts into a soft-focus blur of the park's emerald green lawn and tall, leafy trees, all bathed in the warm golden glow of the afternoon sun, creating a beautiful, soft lens flare. The entire scene is imbued with an energetic, uplifting atmosphere of quiet personal freedom and pure bliss.", + "seed": 20839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_30_seed_20839_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": -0.04306611535139382, + "vlm_logprob": -0.04306611535139382, + "vlm_yes_no_margin": 4.4500000238418576, + "vlm_token_credit": -0.04306611535139382, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.1510406881570816 + ], + "opposite_logprob_tokens": [ + -2.02604079246521 + ], + "logprob_reward": -0.1510406881570816, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1510406881570816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15159544348716736, + 0.0780743658542633, + 0.07288901507854462, + 0.08721641451120377, + 0.0688808411359787, + 0.10540319234132767, + 0.16077962517738342, + 0.11796874552965164, + 0.10664931684732437, + 0.05054299533367157 + ], + "credit_map": [ + 0.022444508969783783, + 0.01556149311363697, + 0.016969239339232445, + 0.019785568118095398, + 0.01753303036093712, + 0.022469516843557358, + 0.017059994861483574, + 0.019772091880440712, + 0.011235954239964485, + 0.009185846894979477, + 0.011183219030499458, + 0.006323864217847586, + 0.011167109943926334, + 0.008817553520202637, + 0.009841547347605228, + 0.010319272987544537, + 0.010209782049059868, + 0.008241756819188595, + 0.007655628025531769, + 0.013753448612987995, + 0.00887213833630085, + 0.007397638168185949, + 0.007405802607536316, + 0.00935282465070486, + 0.011368363164365292, + 0.015571256168186665, + 0.016798976808786392, + 0.014884413219988346, + 0.0061508952639997005, + 0.007855638861656189, + 0.006519488990306854, + 0.008067382499575615, + 0.005870969966053963, + 0.006363595370203257, + 0.00723973149433732, + 0.009899365715682507, + 0.013922817073762417, + 0.006443037651479244, + 0.011244995519518852, + 0.007896323688328266, + 0.013398532755672932, + 0.010813891887664795, + 0.012842579744756222, + 0.012608838267624378, + 0.01181658823043108, + 0.0177358016371727, + 0.015940751880407333, + 0.010246208868920803, + 0.029188787564635277, + 0.011736372485756874, + 0.04814501851797104, + 0.01925305649638176, + 0.01041044108569622, + 0.020847022533416748, + 0.008613268844783306, + 0.012585652060806751, + 0.019053896889090538, + 0.011919531039893627, + 0.02465015836060047, + 0.013011367991566658, + 0.008172395639121532, + 0.013022423721849918, + 0.012037280015647411, + 0.016101693734526634, + 0.010040681809186935, + 0.008452900685369968, + 0.014082434587180614, + 0.00993565283715725, + 0.01329563744366169, + 0.009790689684450626, + 0.019856173545122147, + 0.021195149049162865, + 0.01058391947299242, + 0.0059543801471591, + 0.007354081142693758, + 0.006719628814607859, + 0.004374124575406313, + 0.0058426083996891975, + 0.005064006429165602, + 0.004650249611586332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.1510406881570816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.005283677484840155 + ], + "opposite_logprob_tokens": [ + -5.255283832550049 + ], + "logprob_reward": -0.005283677484840155, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005283677484840155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1712818741798401, + 0.11736006289720535, + 0.05362749472260475, + 0.029185043647885323, + 0.03973732143640518, + 0.03043527714908123, + 0.03622521087527275, + 0.2940603494644165, + 0.09765920042991638, + 0.13042812049388885 + ], + "credit_map": [ + 0.02179047092795372, + 0.01745598204433918, + 0.017914503812789917, + 0.020379310473799706, + 0.02904394268989563, + 0.024979660287499428, + 0.020335376262664795, + 0.019382638856768608, + 0.01934155821800232, + 0.02005738578736782, + 0.01666870340704918, + 0.010611812584102154, + 0.019591646268963814, + 0.011957723647356033, + 0.011428475379943848, + 0.007702759001404047, + 0.0072356597520411015, + 0.007502089720219374, + 0.004873491823673248, + 0.00733785517513752, + 0.010684809647500515, + 0.00648136530071497, + 0.004995634313672781, + 0.004516590852290392, + 0.003384395269677043, + 0.0035308613441884518, + 0.0049815247766673565, + 0.003856708761304617, + 0.003294462338089943, + 0.00263966154307127, + 0.004102776758372784, + 0.0033946565818041563, + 0.003945128992199898, + 0.004610443953424692, + 0.0039680893532931805, + 0.0032614138908684254, + 0.006859112996608019, + 0.003906704485416412, + 0.00847272202372551, + 0.0047137076035141945, + 0.0027984967455267906, + 0.0024326625280082226, + 0.0034112948924303055, + 0.0029859731439501047, + 0.005170091986656189, + 0.004227946046739817, + 0.005763525143265724, + 0.00364528875797987, + 0.0059148347936570644, + 0.003014537738636136, + 0.010061928071081638, + 0.004391715861856937, + 0.0035798889584839344, + 0.0034502504859119654, + 0.0022909636609256268, + 0.003521094797179103, + 0.08866804838180542, + 0.010511644184589386, + 0.10058436542749405, + 0.025104183703660965, + 0.012416771613061428, + 0.028202062472701073, + 0.011799393221735954, + 0.016773907467722893, + 0.025599544867873192, + 0.0074375225231051445, + 0.020372914150357246, + 0.006216123234480619, + 0.009846695698797703, + 0.005378133151680231, + 0.014065125025808811, + 0.008743150159716606, + 0.013208752498030663, + 0.02006533555686474, + 0.016257252544164658, + 0.020420996472239494, + 0.014379695057868958, + 0.01652413234114647, + 0.017326680943369865, + 0.0122452974319458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.005283677484840155 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.04866442829370499 + ], + "opposite_logprob_tokens": [ + -3.0486643314361572 + ], + "logprob_reward": -0.04866442829370499, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866442829370499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10286925733089447, + 0.2820710837841034, + 0.10024488717317581, + 0.034936342388391495, + 0.04184897989034653, + 0.030487533658742905, + 0.03224332258105278, + 0.06849697977304459, + 0.2369421273469925, + 0.06985947489738464 + ], + "credit_map": [ + 0.018828410655260086, + 0.010815363377332687, + 0.014966804534196854, + 0.009311950765550137, + 0.010598772205412388, + 0.012581706047058105, + 0.011657211929559708, + 0.014109033159911633, + 0.03656745329499245, + 0.030543655157089233, + 0.03063957206904888, + 0.027368390932679176, + 0.053148847073316574, + 0.02772577479481697, + 0.04168788716197014, + 0.03438948094844818, + 0.015608533285558224, + 0.013226982206106186, + 0.009888244792819023, + 0.014370190910995007, + 0.01599765382707119, + 0.011418938636779785, + 0.00936818029731512, + 0.010366166010499, + 0.005485169123858213, + 0.005113681312650442, + 0.005460099317133427, + 0.00492877047508955, + 0.003937643952667713, + 0.0034352338407188654, + 0.0031747757457196712, + 0.0034009674564003944, + 0.004343133419752121, + 0.0077066742815077305, + 0.004214353859424591, + 0.005290827248245478, + 0.005420093424618244, + 0.0038465119432657957, + 0.006687804125249386, + 0.004339584615081549, + 0.0032311405520886183, + 0.0025944411754608154, + 0.003566828789189458, + 0.004109332803636789, + 0.005134018138051033, + 0.002412194386124611, + 0.00672810897231102, + 0.0027114690747112036, + 0.004212882835417986, + 0.0034508388489484787, + 0.005263162776827812, + 0.0030160758178681135, + 0.0037447945214807987, + 0.0027512391097843647, + 0.0030205983202904463, + 0.006783728953450918, + 0.014430051669478416, + 0.004833207931369543, + 0.01391687523573637, + 0.0059201386757195, + 0.005046306177973747, + 0.008299622684717178, + 0.0058607012033462524, + 0.010190073400735855, + 0.046245768666267395, + 0.023998994380235672, + 0.039038579910993576, + 0.016428755596280098, + 0.02641279809176922, + 0.018466178327798843, + 0.03850824758410454, + 0.027842802926898003, + 0.00985594280064106, + 0.012008365243673325, + 0.009083881042897701, + 0.010529248043894768, + 0.0070808385498821735, + 0.007477233652025461, + 0.007411862723529339, + 0.006412106100469828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.04866442829370499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0005614373367279768 + ], + "opposite_logprob_tokens": [ + -7.500561237335205 + ], + "logprob_reward": -0.0005614373367279768, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005614373367279768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12615497410297394, + 0.25239041447639465, + 0.028222711756825447, + 0.04825516417622566, + 0.030817434191703796, + 0.06094781309366226, + 0.06805737316608429, + 0.1289488673210144, + 0.027750208973884583, + 0.22845499217510223 + ], + "credit_map": [ + 0.020897896960377693, + 0.010236518457531929, + 0.01629924401640892, + 0.014456108212471008, + 0.01649315282702446, + 0.01938285119831562, + 0.013305559754371643, + 0.015083643607795238, + 0.037420544773340225, + 0.042532119899988174, + 0.038060542196035385, + 0.024117719382047653, + 0.044414274394512177, + 0.01838173344731331, + 0.022782595828175545, + 0.02468091994524002, + 0.004283682443201542, + 0.00454287976026535, + 0.0026443563401699066, + 0.004308196250349283, + 0.0053383768536150455, + 0.0021674393210560083, + 0.003159665036946535, + 0.0017781163332983851, + 0.008263682015240192, + 0.007060719653964043, + 0.008939603343605995, + 0.006847992539405823, + 0.004096385091543198, + 0.004312204197049141, + 0.004755501169711351, + 0.00397907430306077, + 0.003552958369255066, + 0.005222022067755461, + 0.0037592893932014704, + 0.0045743235386908054, + 0.0032998507376760244, + 0.004023908171802759, + 0.0033350607845932245, + 0.0030500206630676985, + 0.009999401867389679, + 0.007498240098357201, + 0.009281257167458534, + 0.009637489914894104, + 0.005577720236033201, + 0.008142017759382725, + 0.006298931315541267, + 0.0045127528719604015, + 0.01056743785738945, + 0.0054932222701609135, + 0.01904861442744732, + 0.006726924795657396, + 0.005077366717159748, + 0.007390138693153858, + 0.0040581016801297665, + 0.009695562534034252, + 0.04156367853283882, + 0.005545319058001041, + 0.04432346671819687, + 0.010675759986042976, + 0.003110928926616907, + 0.011991962790489197, + 0.005262024234980345, + 0.00647572660818696, + 0.004094181582331657, + 0.004020104184746742, + 0.002518546534702182, + 0.002750062383711338, + 0.0033594470005482435, + 0.0028512231074273586, + 0.004234192427247763, + 0.003922454081475735, + 0.04377906769514084, + 0.02525874599814415, + 0.034430500119924545, + 0.032451480627059937, + 0.020910996943712234, + 0.0295993871986866, + 0.021422164514660835, + 0.0206026379019022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.0005614373367279768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11456327885389328, + 0.11091654747724533, + 0.05982328578829765, + 0.07870831340551376, + 0.054889027029275894, + 0.05844704434275627, + 0.09067098051309586, + 0.1367466002702713, + 0.21746128797531128, + 0.07777363061904907 + ], + "credit_map": [ + 0.02197117544710636, + 0.011017369106411934, + 0.01821746677160263, + 0.012014176696538925, + 0.012308889999985695, + 0.011526569724082947, + 0.011729326099157333, + 0.015778303146362305, + 0.021303387358784676, + 0.016657039523124695, + 0.021815717220306396, + 0.0116009172052145, + 0.010687664151191711, + 0.010174675844609737, + 0.007928190752863884, + 0.010748956352472305, + 0.007869088090956211, + 0.010798237286508083, + 0.005776108242571354, + 0.0070053706876933575, + 0.008881356567144394, + 0.0059462119825184345, + 0.006203946657478809, + 0.007342967204749584, + 0.012795304879546165, + 0.013390464708209038, + 0.013024791143834591, + 0.012041979469358921, + 0.006583736278116703, + 0.0066200559958815575, + 0.007003914099186659, + 0.00724807009100914, + 0.004793214611709118, + 0.007664091885089874, + 0.006052000448107719, + 0.006590777076780796, + 0.007763986475765705, + 0.005103630479425192, + 0.009792988188564777, + 0.007128339726477861, + 0.010385937988758087, + 0.005789094138890505, + 0.009003476239740849, + 0.008135519921779633, + 0.006488865707069635, + 0.006637845188379288, + 0.007079383358359337, + 0.00492691807448864, + 0.011835813522338867, + 0.008862531743943691, + 0.008152969181537628, + 0.007859366945922375, + 0.009386653080582619, + 0.015508194454014301, + 0.013355448842048645, + 0.01570999249815941, + 0.026682615280151367, + 0.013685856014490128, + 0.018832910805940628, + 0.02092822454869747, + 0.011518076993525028, + 0.013722962699830532, + 0.01023702323436737, + 0.021138930693268776, + 0.06562049686908722, + 0.015500413253903389, + 0.05059553682804108, + 0.01341919507831335, + 0.019409092143177986, + 0.012411453761160374, + 0.0232070442289114, + 0.017298072576522827, + 0.012454173527657986, + 0.010274212807416916, + 0.014417899772524834, + 0.009332100860774517, + 0.008718653582036495, + 0.007480406668037176, + 0.006494962610304356, + 0.00860122125595808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.009780345484614372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a sleek, steady camera, this cinematic close-up features a young woman jogging effortlessly through a sun-drenched park. Her chestnut hair is pulled back into a high, bouncing ponytail, with a few locks framing her exercise-flushed face. She wears a pair of stylish, minimalist white wireless earbuds, which fit securely and comfortably in her ears. A bright, happy smile illuminates her features, and her eyes crinkle at the corners as she loses herself in the music. The camera focuses firmly on her joyful expression as she moves with steady, rhythmic grace. The background melts into a soft-focus blur of the park's emerald green lawn and tall, leafy trees, all bathed in the warm golden glow of the afternoon sun, creating a beautiful, soft lens flare. The entire scene is imbued with an energetic, uplifting atmosphere of quiet personal freedom and pure bliss.", + "seed": 20840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_30_seed_20840_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": -0.07137002004310489, + "vlm_logprob": -0.07137002004310489, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.07137002004310489, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.2589755058288574 + ], + "opposite_logprob_tokens": [ + -1.5089755058288574 + ], + "logprob_reward": -0.2589755058288574, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2589755058288574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14048179984092712, + 0.29073190689086914, + 0.04430122300982475, + 0.16060437262058258, + 0.044542551040649414, + 0.19579419493675232, + 0.023311488330364227, + 0.022554896771907806, + 0.02890302985906601, + 0.048774536699056625 + ], + "credit_map": [ + 0.018688835203647614, + 0.004709442146122456, + 0.04739891737699509, + 0.004871632903814316, + 0.006535363383591175, + 0.010087475180625916, + 0.0053492398001253605, + 0.042840879410505295, + 0.11074691265821457, + 0.009622509591281414, + 0.11965563148260117, + 0.007994617335498333, + 0.007855893112719059, + 0.019808771088719368, + 0.0068082488141953945, + 0.008239313028752804, + 0.00856700912117958, + 0.0038146176375448704, + 0.005232417024672031, + 0.00426456006243825, + 0.0053038629703223705, + 0.006227611564099789, + 0.0030572630930691957, + 0.007833884097635746, + 0.022495592013001442, + 0.013935292139649391, + 0.017126701772212982, + 0.01813836395740509, + 0.020647408440709114, + 0.029303574934601784, + 0.02527317777276039, + 0.01368425227701664, + 0.004559014458209276, + 0.005431740544736385, + 0.005141609814018011, + 0.007131170481443405, + 0.004535739775747061, + 0.008552138693630695, + 0.00341776292771101, + 0.005773372016847134, + 0.06885319203138351, + 0.007386643905192614, + 0.010090315714478493, + 0.0036702107172459364, + 0.007449017837643623, + 0.01967577636241913, + 0.006947631016373634, + 0.0717214047908783, + 0.005358041264116764, + 0.0016095397295430303, + 0.0033417772501707077, + 0.001276256050914526, + 0.0027778984513133764, + 0.00405479408800602, + 0.0016711810603737831, + 0.00322199915535748, + 0.002727261744439602, + 0.002060474129393697, + 0.0032166584860533476, + 0.0026066091377288103, + 0.0027588417287915945, + 0.002817218191921711, + 0.002801706548780203, + 0.003566126339137554, + 0.007221819367259741, + 0.0021328097209334373, + 0.004915166646242142, + 0.001956607447937131, + 0.0033169083762913942, + 0.0036265056114643812, + 0.0033878551330417395, + 0.002345356158912182, + 0.007069406099617481, + 0.003066002856940031, + 0.011032305657863617, + 0.006288276519626379, + 0.0032900965306907892, + 0.004280323162674904, + 0.003204505192115903, + 0.010543623939156532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.2589755058288574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0067643700167536736 + ], + "opposite_logprob_tokens": [ + -5.0067644119262695 + ], + "logprob_reward": -0.0067643700167536736, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067643700167536736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8060107231140137, + 0.03935607150197029, + 0.019788827747106552, + 0.01348316390067339, + 0.023621277883648872, + 0.06071152910590172, + 0.011253038421273232, + 0.004402051214128733, + 0.01238276157528162, + 0.008990493603050709 + ], + "credit_map": [ + 0.11475462466478348, + 0.020564673468470573, + 0.29166412353515625, + 0.022203434258699417, + 0.031343892216682434, + 0.045011695474386215, + 0.022870203480124474, + 0.2575981616973877, + 0.014237446710467339, + 0.0014470444293692708, + 0.016176827251911163, + 0.001273275469429791, + 0.0013047991087660193, + 0.0028202044777572155, + 0.0009991738479584455, + 0.0010973032331094146, + 0.0032137867528945208, + 0.0016471046255901456, + 0.002804007614031434, + 0.0017082837875932455, + 0.0021339906379580498, + 0.002494337735697627, + 0.0016888240352272987, + 0.004098494537174702, + 0.0017637748969718814, + 0.001706258743070066, + 0.0013231543125584722, + 0.002044940134510398, + 0.0016354912659153342, + 0.0014809212880209088, + 0.002452920889481902, + 0.0010757028358057141, + 0.0026825524400919676, + 0.003405563300475478, + 0.0025574853643774986, + 0.0037577336188405752, + 0.002033558674156666, + 0.00401671975851059, + 0.001621335162781179, + 0.003546329215168953, + 0.021969474852085114, + 0.002375735668465495, + 0.0032232331577688456, + 0.0010708405170589685, + 0.0016042334027588367, + 0.00637738686054945, + 0.0020003376994282007, + 0.022090287879109383, + 0.0022790238726884127, + 0.0010424606734886765, + 0.0012834003427997231, + 0.0006794826476834714, + 0.001159729203209281, + 0.0019522318616509438, + 0.0008960269042290747, + 0.0019606840796768665, + 0.0005417645443230867, + 0.00036644694046117365, + 0.0007454910082742572, + 0.00044834997970610857, + 0.00043480281601659954, + 0.0005300755728967488, + 0.000529749202542007, + 0.0008053716155700386, + 0.003292528912425041, + 0.0008091085474006832, + 0.004097003489732742, + 0.000618274905718863, + 0.0008447334985248744, + 0.0008743187063373625, + 0.0008141968864947557, + 0.0010325972689315677, + 0.0016681243432685733, + 0.0007763145840726793, + 0.0019347000634297729, + 0.000753828207962215, + 0.001007339684292674, + 0.0010316671105101705, + 0.0005237708101049066, + 0.0012947493232786655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0067643700167536736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.07899495214223862 + ], + "opposite_logprob_tokens": [ + -2.5789949893951416 + ], + "logprob_reward": -0.07899495214223862, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899495214223862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5593991875648499, + 0.10861218720674515, + 0.025157583877444267, + 0.03866605460643768, + 0.03165939450263977, + 0.09015562385320663, + 0.026086878031492233, + 0.03544119745492935, + 0.02313719503581524, + 0.06168471649289131 + ], + "credit_map": [ + 0.07962093502283096, + 0.015031016431748867, + 0.20018185675144196, + 0.015645653009414673, + 0.021350346505641937, + 0.03020685724914074, + 0.016724100336432457, + 0.1806384027004242, + 0.04011394828557968, + 0.004572130274027586, + 0.04400060698390007, + 0.0035805401857942343, + 0.003368626581504941, + 0.007176730316132307, + 0.0027192686684429646, + 0.0030803338158875704, + 0.0036593389231711626, + 0.0021403650753200054, + 0.0034592507872730494, + 0.0025260222610086203, + 0.0033311841543763876, + 0.0026392494328320026, + 0.0036552611272782087, + 0.0037469121161848307, + 0.005233042407780886, + 0.003382552182301879, + 0.004081126768141985, + 0.005349470302462578, + 0.004325580317527056, + 0.005566287320107222, + 0.007005922496318817, + 0.0037220725789666176, + 0.0036432608030736446, + 0.003701763227581978, + 0.004111039452254772, + 0.003652657615020871, + 0.003470382886007428, + 0.006212202366441488, + 0.0022126927506178617, + 0.004655392840504646, + 0.030620861798524857, + 0.004879757296293974, + 0.005421230103820562, + 0.0027869855985045433, + 0.0030376766808331013, + 0.010627657175064087, + 0.0037605762481689453, + 0.02902088314294815, + 0.003961182199418545, + 0.0024014238733798265, + 0.0020024387631565332, + 0.0014305418590083718, + 0.005665143020451069, + 0.005130690056830645, + 0.0019677001982927322, + 0.0035277586430311203, + 0.004269259981811047, + 0.0030382550321519375, + 0.004621000960469246, + 0.003661188529804349, + 0.0036117765121161938, + 0.006343065295368433, + 0.004147909115999937, + 0.005748739466071129, + 0.005463052541017532, + 0.0020111557096242905, + 0.005499368067830801, + 0.0012461801525205374, + 0.0020902082324028015, + 0.001998948398977518, + 0.0026799505576491356, + 0.0021483302116394043, + 0.009009823203086853, + 0.004030521959066391, + 0.009951293468475342, + 0.009230083785951138, + 0.010955963283777237, + 0.005546666216105223, + 0.0036139176227152348, + 0.009346451610326767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07899495214223862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0010422994382679462 + ], + "opposite_logprob_tokens": [ + -6.876042366027832 + ], + "logprob_reward": -0.0010422994382679462, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010422994382679462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32296183705329895, + 0.4278782606124878, + 0.01612105593085289, + 0.012545780278742313, + 0.05033797398209572, + 0.07548747211694717, + 0.01707555539906025, + 0.03981339558959007, + 0.019420552998781204, + 0.01835818402469158 + ], + "credit_map": [ + 0.04600509628653526, + 0.008411750197410583, + 0.11612466722726822, + 0.009025407023727894, + 0.012558553367853165, + 0.017595235258340836, + 0.009502262808382511, + 0.10373882204294205, + 0.16409040987491608, + 0.01418014895170927, + 0.17952412366867065, + 0.011417191475629807, + 0.011423110030591488, + 0.02820151299238205, + 0.009003684855997562, + 0.010038048960268497, + 0.0025523824151605368, + 0.0016754985554143786, + 0.002035296056419611, + 0.002513428684324026, + 0.0013439201284199953, + 0.001737924525514245, + 0.0014991196803748608, + 0.0027634850703179836, + 0.0017752652056515217, + 0.0015384485013782978, + 0.0013278885744512081, + 0.001245731022208929, + 0.002608356298878789, + 0.0015433857915923, + 0.0015987710794433951, + 0.0009079338633455336, + 0.004814106971025467, + 0.007057869341224432, + 0.004898900166153908, + 0.00821426697075367, + 0.00447701383382082, + 0.010599537752568722, + 0.0033505430910736322, + 0.006925729103386402, + 0.02626917138695717, + 0.0030246784444898367, + 0.004058548249304295, + 0.0019559040665626526, + 0.002374491887167096, + 0.007752673700451851, + 0.0025974479503929615, + 0.027454553171992302, + 0.0031533853616565466, + 0.0017119201365858316, + 0.002114779781550169, + 0.001065436052158475, + 0.002647931920364499, + 0.003058153670281172, + 0.0018138266168534756, + 0.0015101211611181498, + 0.004869000054895878, + 0.004315054975450039, + 0.005876917392015457, + 0.004484075121581554, + 0.004655947908759117, + 0.005351586267352104, + 0.004286142531782389, + 0.005974668078124523, + 0.0065238578245043755, + 0.0008726329542696476, + 0.007019923534244299, + 0.0007084631943143904, + 0.0010541147785261273, + 0.0013443378265947104, + 0.0007723362650722265, + 0.0011248852824792266, + 0.0029084531124681234, + 0.0015175612643361092, + 0.00332848378457129, + 0.0016087716212496161, + 0.0018930176738649607, + 0.0022630365565419197, + 0.0009532190742902458, + 0.0038856398314237595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0010422994382679462 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.011072972789406776 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011072972789406776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072972789406776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4487612545490265, + 0.2376449853181839, + 0.022046944126486778, + 0.019678007811307907, + 0.03655339777469635, + 0.1544499546289444, + 0.025464601814746857, + 0.007605370134115219, + 0.019217072054743767, + 0.028578413650393486 + ], + "credit_map": [ + 0.06482390314340591, + 0.011105436831712723, + 0.16444943845272064, + 0.01280719880014658, + 0.018103260546922684, + 0.024420294910669327, + 0.01221460197120905, + 0.14083705842494965, + 0.08965712040662766, + 0.008648109622299671, + 0.09733361005783081, + 0.0068793585523962975, + 0.007417174056172371, + 0.015884635969996452, + 0.0054063210263848305, + 0.006418646778911352, + 0.0036342672538012266, + 0.002309599658474326, + 0.002329173032194376, + 0.0031121731735765934, + 0.002618678379803896, + 0.002940333914011717, + 0.0025306616444140673, + 0.002572055673226714, + 0.0028499290347099304, + 0.0016045747324824333, + 0.00222985097207129, + 0.0028683419805020094, + 0.001953486120328307, + 0.0032832284923642874, + 0.0031070313416421413, + 0.001781564555130899, + 0.004512957762926817, + 0.005485246889293194, + 0.003743621753528714, + 0.007574328687041998, + 0.003043445525690913, + 0.004923696629703045, + 0.002571831224486232, + 0.004698263946920633, + 0.05560988932847977, + 0.006236070767045021, + 0.008181223645806313, + 0.0030854910146445036, + 0.00396902859210968, + 0.015847772359848022, + 0.005796755664050579, + 0.055723730474710464, + 0.005174687597900629, + 0.0028248631861060858, + 0.0034311150666326284, + 0.00174368591979146, + 0.0028410484082996845, + 0.004238416440784931, + 0.0023805503733456135, + 0.0028302341233938932, + 0.0013120983494445682, + 0.000554995087441057, + 0.0010410475078970194, + 0.0006877646665088832, + 0.0007948410348035395, + 0.0013725882163271308, + 0.0007409645477309823, + 0.0011010699672624469, + 0.002399487653747201, + 0.003501413157209754, + 0.002631256589666009, + 0.0017000655643641949, + 0.00241200253367424, + 0.0023412946611642838, + 0.002178091323003173, + 0.002053460106253624, + 0.003797618905082345, + 0.001604453194886446, + 0.007230591494590044, + 0.0022517892066389322, + 0.002021661726757884, + 0.002626172499731183, + 0.0018199706682935357, + 0.007226154673844576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011072972789406776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot with a sleek, steady camera, this cinematic close-up features a young woman jogging effortlessly through a sun-drenched park. Her chestnut hair is pulled back into a high, bouncing ponytail, with a few locks framing her exercise-flushed face. She wears a pair of stylish, minimalist white wireless earbuds, which fit securely and comfortably in her ears. A bright, happy smile illuminates her features, and her eyes crinkle at the corners as she loses herself in the music. The camera focuses firmly on her joyful expression as she moves with steady, rhythmic grace. The background melts into a soft-focus blur of the park's emerald green lawn and tall, leafy trees, all bathed in the warm golden glow of the afternoon sun, creating a beautiful, soft lens flare. The entire scene is imbued with an energetic, uplifting atmosphere of quiet personal freedom and pure bliss.", + "seed": 20841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_15/index_30_seed_20841_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": -0.0688302357099019, + "vlm_logprob": -0.0688302357099019, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.0688302357099019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.2590235471725464 + ], + "opposite_logprob_tokens": [ + -1.5090235471725464 + ], + "logprob_reward": -0.2590235471725464, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2590235471725464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0934695452451706, + 0.07113929837942123, + 0.20103596150875092, + 0.07133655250072479, + 0.06280067563056946, + 0.10570577532052994, + 0.1477539837360382, + 0.12333564460277557, + 0.10076554864645004, + 0.022657053545117378 + ], + "credit_map": [ + 0.014293774031102657, + 0.011709639802575111, + 0.012835441157221794, + 0.007075967732816935, + 0.014643935486674309, + 0.016459790989756584, + 0.008624338544905186, + 0.00782665703445673, + 0.009201043285429478, + 0.0089913010597229, + 0.010631285607814789, + 0.006025781389325857, + 0.010923529043793678, + 0.012211968190968037, + 0.005207217298448086, + 0.007947162725031376, + 0.016338951885700226, + 0.009666112251579762, + 0.06988733261823654, + 0.007978228852152824, + 0.017532141879200935, + 0.05118253827095032, + 0.007623046170920134, + 0.020827602595090866, + 0.008410455659031868, + 0.009188658557832241, + 0.0058929952792823315, + 0.005031964275985956, + 0.011068454012274742, + 0.009557168930768967, + 0.012338913045823574, + 0.009847933426499367, + 0.016241421923041344, + 0.00443508755415678, + 0.017658114433288574, + 0.004698001313954592, + 0.003724997863173485, + 0.005082752089947462, + 0.006136768963187933, + 0.0048235272988677025, + 0.014159096404910088, + 0.012743476778268814, + 0.009148724377155304, + 0.027175022289156914, + 0.008069617673754692, + 0.010499805212020874, + 0.009736260399222374, + 0.014173751696944237, + 0.01799190230667591, + 0.02290167845785618, + 0.02212865836918354, + 0.017522644251585007, + 0.017875079065561295, + 0.014717524871230125, + 0.017624646425247192, + 0.016991833224892616, + 0.01658026874065399, + 0.012296551838517189, + 0.008954312652349472, + 0.009870382957160473, + 0.008013073354959488, + 0.024949248880147934, + 0.011700560338795185, + 0.030971236526966095, + 0.009094043634831905, + 0.01309259608387947, + 0.009415924549102783, + 0.00870359968394041, + 0.019342077895998955, + 0.012636356055736542, + 0.020090948790311813, + 0.008389994502067566, + 0.0032817069441080093, + 0.0024033631198108196, + 0.004909854847937822, + 0.0025122244842350483, + 0.003174061654135585, + 0.0018514262046664953, + 0.0022341294679790735, + 0.002290286123752594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.2590235471725464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0067628310061991215 + ], + "opposite_logprob_tokens": [ + -5.006762981414795 + ], + "logprob_reward": -0.0067628310061991215, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067628310061991215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15152068436145782, + 0.29445114731788635, + 0.11596732586622238, + 0.08957042545080185, + 0.04045888036489487, + 0.041530169546604156, + 0.07922278344631195, + 0.10440550744533539, + 0.0431939959526062, + 0.039679039269685745 + ], + "credit_map": [ + 0.02384762093424797, + 0.016596974804997444, + 0.026144860312342644, + 0.012853260152041912, + 0.01541138906031847, + 0.02488625980913639, + 0.013527346774935722, + 0.018252979964017868, + 0.04760819301009178, + 0.039570312947034836, + 0.03126603364944458, + 0.020901992917060852, + 0.03359568864107132, + 0.07914617657661438, + 0.01848168671131134, + 0.023881088942289352, + 0.011358375661075115, + 0.0076390961185097694, + 0.03809361904859543, + 0.008094524964690208, + 0.012156479060649872, + 0.01816927082836628, + 0.007523005362600088, + 0.01293296366930008, + 0.009987778030335903, + 0.010015622712671757, + 0.005735567305237055, + 0.00469673378393054, + 0.016435643658041954, + 0.01724713295698166, + 0.014798786491155624, + 0.010653164237737656, + 0.011126200668513775, + 0.003029328305274248, + 0.0102553004398942, + 0.002953960560262203, + 0.0027950042858719826, + 0.0032340222969651222, + 0.0035367358941584826, + 0.003528324654325843, + 0.004461323376744986, + 0.004464637953788042, + 0.0022886835504323244, + 0.007432202808558941, + 0.004110014531761408, + 0.007047918159514666, + 0.005180895794183016, + 0.006544492207467556, + 0.010468834079802036, + 0.00863924901932478, + 0.008967973291873932, + 0.006266192067414522, + 0.013939553871750832, + 0.009530291892588139, + 0.012657280080020428, + 0.008753402158617973, + 0.010418304242193699, + 0.012520146556198597, + 0.008066486567258835, + 0.011129572987556458, + 0.009078054688870907, + 0.014548094943165779, + 0.022549845278263092, + 0.016095006838440895, + 0.006859158631414175, + 0.00466063292697072, + 0.006341161672025919, + 0.005482343025505543, + 0.004179719369858503, + 0.005655951332300901, + 0.005657518282532692, + 0.004357511643320322, + 0.0046954648569226265, + 0.003792861010879278, + 0.005085890181362629, + 0.003922627307474613, + 0.0043733553029596806, + 0.007322652265429497, + 0.0038662313017994165, + 0.0066199577413499355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0067628310061991215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.07002059370279312 + ], + "opposite_logprob_tokens": [ + -2.6950206756591797 + ], + "logprob_reward": -0.07002059370279312, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002059370279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060211002826690674, + 0.48184439539909363, + 0.05552962049841881, + 0.027376467362046242, + 0.1288447231054306, + 0.03109852224588394, + 0.12369400262832642, + 0.03576065972447395, + 0.03138529136776924, + 0.02425527013838291 + ], + "credit_map": [ + 0.009144443087279797, + 0.0073366714641451836, + 0.006406561005860567, + 0.006088738329708576, + 0.011657696217298508, + 0.007514915894716978, + 0.0070186518132686615, + 0.005043328739702702, + 0.07728160917758942, + 0.05694098770618439, + 0.06049995869398117, + 0.038571469485759735, + 0.058424025774002075, + 0.11148256808519363, + 0.03429719805717468, + 0.04434658959507942, + 0.005599240306764841, + 0.003970286808907986, + 0.018201559782028198, + 0.003526340937241912, + 0.005654554348438978, + 0.00918116420507431, + 0.0033397767692804337, + 0.006056704092770815, + 0.002851275261491537, + 0.0032256690319627523, + 0.0017742475029081106, + 0.0014113347278907895, + 0.003128715790808201, + 0.005460234358906746, + 0.004770409781485796, + 0.004754583351314068, + 0.03428724780678749, + 0.009193001314997673, + 0.03436755761504173, + 0.009565334767103195, + 0.0069890148006379604, + 0.010650432668626308, + 0.011650348082184792, + 0.012141797691583633, + 0.004479362163692713, + 0.003562235739082098, + 0.0021333852782845497, + 0.004392572212964296, + 0.0030058061238378286, + 0.003952840343117714, + 0.0064115081913769245, + 0.003160815453156829, + 0.012758065015077591, + 0.012196142226457596, + 0.013857514597475529, + 0.008935818448662758, + 0.029887232929468155, + 0.01418520137667656, + 0.01971285790205002, + 0.012161177583038807, + 0.004904225468635559, + 0.002983850659802556, + 0.002556083258241415, + 0.0038832894060760736, + 0.003570105880498886, + 0.005187645088881254, + 0.007433455437421799, + 0.0052420045249164104, + 0.003766168374568224, + 0.0035582752898335457, + 0.004023815039545298, + 0.0025885896757245064, + 0.004778479225933552, + 0.0046204570680856705, + 0.005286200903356075, + 0.0027633090503513813, + 0.004079325590282679, + 0.002900106133893132, + 0.004715799819678068, + 0.0019122576341032982, + 0.002271703677251935, + 0.003115294501185417, + 0.002469034865498543, + 0.0027917507104575634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.07002059370279312 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.0007179781678132713 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007179781678132713, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007179781678132713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1298409402370453, + 0.160917729139328, + 0.2323315441608429, + 0.09050010144710541, + 0.05302836373448372, + 0.02812184952199459, + 0.1020612046122551, + 0.055038850754499435, + 0.07091517746448517, + 0.07724418491125107 + ], + "credit_map": [ + 0.015572653152048588, + 0.013664516620337963, + 0.027765687555074692, + 0.010803230106830597, + 0.01633516512811184, + 0.013968734070658684, + 0.013936979696154594, + 0.01779397763311863, + 0.027329986914992332, + 0.019759155809879303, + 0.031098151579499245, + 0.016470521688461304, + 0.014871903695166111, + 0.01907440833747387, + 0.012825357727706432, + 0.019488252699375153, + 0.020869728177785873, + 0.012221330776810646, + 0.07360055297613144, + 0.008566001430153847, + 0.02094525657594204, + 0.06600387394428253, + 0.008804270066320896, + 0.021320542320609093, + 0.015925079584121704, + 0.011209465563297272, + 0.010555761866271496, + 0.004994618706405163, + 0.009950821287930012, + 0.014735114760696888, + 0.011593441478908062, + 0.011535797268152237, + 0.012868471443653107, + 0.004577487241476774, + 0.01153328362852335, + 0.0028901989571750164, + 0.0036604818888008595, + 0.005634240340441465, + 0.0053734686225652695, + 0.0064907316118478775, + 0.004647642374038696, + 0.005556568503379822, + 0.0028346616309136152, + 0.0029020891524851322, + 0.0029186601750552654, + 0.0033982563763856888, + 0.0035616219975054264, + 0.00230234838090837, + 0.010261690244078636, + 0.010836734436452389, + 0.011561539955437183, + 0.009227072820067406, + 0.02296384423971176, + 0.013268576003611088, + 0.015364431776106358, + 0.008577327243983746, + 0.007767592556774616, + 0.005313152447342873, + 0.005044431425631046, + 0.004814555868506432, + 0.004474766086786985, + 0.011555318720638752, + 0.006422697566449642, + 0.009646338410675526, + 0.010095156729221344, + 0.007287549786269665, + 0.01019410602748394, + 0.007443010341376066, + 0.007189262192696333, + 0.013156813569366932, + 0.009067916311323643, + 0.006481363903731108, + 0.010309557430446148, + 0.009711967781186104, + 0.019035743549466133, + 0.006742632482200861, + 0.007879105396568775, + 0.009768764488399029, + 0.005480051040649414, + 0.008316357620060444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0007179781678132713 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 jogging through a park?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have chestnut hair styled in a high ponytail with loose strands framing her face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing white wireless earbuds in her ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a soft-focus blur of green lawn and trees bathed in warm golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a cinematic close-up with a steady camera focus on the woman's 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": [ + -0.007626228500157595 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007626228500157595, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626228500157595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05746149271726608, + 0.20664872229099274, + 0.48461201786994934, + 0.025222154334187508, + 0.06707706302404404, + 0.022873811423778534, + 0.03109268844127655, + 0.027938924729824066, + 0.04120425134897232, + 0.035868868231773376 + ], + "credit_map": [ + 0.008135322481393814, + 0.00688323937356472, + 0.0074369823560118675, + 0.0047286683693528175, + 0.0074029723182320595, + 0.010086441412568092, + 0.006029950454831123, + 0.0067579178139567375, + 0.03362783044576645, + 0.023780254647135735, + 0.029298000037670135, + 0.01807124726474285, + 0.025679944083094597, + 0.0392901748418808, + 0.015583468601107597, + 0.021317802369594574, + 0.041035301983356476, + 0.02526480332016945, + 0.16405507922172546, + 0.017244242131710052, + 0.0435604602098465, + 0.1371540129184723, + 0.016274018213152885, + 0.04002409428358078, + 0.003147035837173462, + 0.0033387006260454655, + 0.003307610284537077, + 0.001585413352586329, + 0.00284943962469697, + 0.004562085960060358, + 0.003876789938658476, + 0.002555077662691474, + 0.01988876424729824, + 0.003842519596219063, + 0.01874767616391182, + 0.0032030984293669462, + 0.0025941573549062014, + 0.006104578264057636, + 0.005118978675454855, + 0.007577292155474424, + 0.003484816290438175, + 0.004071686416864395, + 0.002363500650972128, + 0.0017904400592669845, + 0.002436256967484951, + 0.004259547218680382, + 0.002349328948184848, + 0.0021182354539632797, + 0.0033491977956146, + 0.003199209226295352, + 0.0030632854904979467, + 0.0026876647025346756, + 0.007597117684781551, + 0.003362919669598341, + 0.005504358094185591, + 0.002328934846445918, + 0.003640069393441081, + 0.00331317912787199, + 0.00203341175802052, + 0.003235254203900695, + 0.0025793532840907574, + 0.00402183597907424, + 0.004634961020201445, + 0.004480861127376556, + 0.003908612299710512, + 0.00525471568107605, + 0.003568279091268778, + 0.004286262206733227, + 0.007753180805593729, + 0.005748196505010128, + 0.007134007290005684, + 0.0035509990993887186, + 0.0034020182210952044, + 0.004165475256741047, + 0.005081938114017248, + 0.003169021336361766, + 0.004055636003613472, + 0.0074853491969406605, + 0.0025135299656540155, + 0.005995902232825756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.007626228500157595 + } + ], + "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/0000152/group_2/index_tensor([9151, 9151])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_tensor([9151, 9151])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..923090edc56ea78d833d4f84fd1dc6fe408046f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_tensor([9151, 9151])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From an epic perspective, the camera rotates with slow, majestic movements, capturing a cosmic spectacle utterly immersed in the unbridled brushstrokes and impasto of Van Gogh's \"The Starry Night.\" At the center of the image lies Earth, yet it has been transformed beyond all recognition: the once azure ocean has become a vast, seething vortex, where molten crimson and glacial blue violently intertwine, collide, and tumble, forming a dynamic, charged storm of color that foreshadows a profound future transformation. Above the planet's sky, vast, vibrant ribbons of auroras dance gracefully in hues of emerald green, brilliant purple, and rosy pink. Their vibrant forms, like the breath of the universe, unfold against the inky blackness. Far from a static darkness, the cosmic space is teeming with Van Gogh's vitality. Stars are no longer simple points of light, but luminous yellow swirls and pulsating clusters, each ray a vibrant, thick coat of paint squeezed straight from a tube. As the camera slowly rotates, this chaotic and magnificent scene is presented in a panoramic view. The turbulent earth and the dancing auroras together form a visual symphony full of primitive vitality.", + "seed": 2626, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2626_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": -0.05449063372798264, + "vlm_logprob": -0.05449063372798264, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.05449063372798264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009563583880662918 + ], + "opposite_logprob_tokens": [ + -5.134563446044922 + ], + "logprob_reward": -0.009563583880662918, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009563583880662918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1268608719110489, + 0.07876770943403244, + 0.07384956628084183, + 0.05224473774433136, + 0.03621811047196388, + 0.05767730250954628, + 0.2631798982620239, + 0.09621189534664154, + 0.11319808661937714, + 0.10179179161787033 + ], + "credit_map": [ + 0.01456964761018753, + 0.011252013966441154, + 0.014839904382824898, + 0.01887236163020134, + 0.014946231618523598, + 0.015464131720364094, + 0.020122548565268517, + 0.016794046387076378, + 0.012496599927544594, + 0.012099039740860462, + 0.012329395860433578, + 0.007597795687615871, + 0.008060749620199203, + 0.007249980233609676, + 0.007867430336773396, + 0.011066725477576256, + 0.013175171799957752, + 0.008696464821696281, + 0.010651253163814545, + 0.009046635590493679, + 0.005271547939628363, + 0.007447793614119291, + 0.007890647277235985, + 0.011670059524476528, + 0.009265050292015076, + 0.005154614802449942, + 0.007885265164077282, + 0.005018988158553839, + 0.006384941749274731, + 0.007711726240813732, + 0.004311299882829189, + 0.006512860767543316, + 0.004442689474672079, + 0.0042223138734698296, + 0.004989556968212128, + 0.004801821894943714, + 0.005778610706329346, + 0.004399754572659731, + 0.00460846396163106, + 0.002974901581183076, + 0.007465468719601631, + 0.00767631595954299, + 0.010437258519232273, + 0.007031583692878485, + 0.009220411069691181, + 0.005373136606067419, + 0.006309890188276768, + 0.004163243342190981, + 0.029917296022176743, + 0.028276214376091957, + 0.018682556226849556, + 0.02642260491847992, + 0.05736669525504112, + 0.06219504773616791, + 0.017224757000803947, + 0.02309476211667061, + 0.025511596351861954, + 0.006417260505259037, + 0.008409571833908558, + 0.00801229476928711, + 0.004457985050976276, + 0.018958263099193573, + 0.008907279931008816, + 0.015537652187049389, + 0.01855023205280304, + 0.009379466995596886, + 0.015100019052624702, + 0.014747437089681625, + 0.008638089522719383, + 0.016983142122626305, + 0.016040969640016556, + 0.013758744113147259, + 0.01299452967941761, + 0.01222955621778965, + 0.01678110472857952, + 0.013048999011516571, + 0.01177775114774704, + 0.01349404826760292, + 0.010634714737534523, + 0.010831106454133987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.009563583880662918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822702983394265 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.002822702983394265, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822702983394265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07741429656744003, + 0.04935568571090698, + 0.051060907542705536, + 0.050226613879203796, + 0.022396190091967583, + 0.17674554884433746, + 0.17698368430137634, + 0.0778038501739502, + 0.10117567330598831, + 0.21683751046657562 + ], + "credit_map": [ + 0.012426870875060558, + 0.01271426398307085, + 0.008812022395431995, + 0.00920095108449459, + 0.0066880397498607635, + 0.008259891532361507, + 0.007320414762943983, + 0.011991845443844795, + 0.006361199077218771, + 0.01061785314232111, + 0.003819537116214633, + 0.004931046161800623, + 0.005986033007502556, + 0.0050838422030210495, + 0.004171197768300772, + 0.00838498305529356, + 0.006767562590539455, + 0.006842828821390867, + 0.008173109963536263, + 0.005427272990345955, + 0.007398128043860197, + 0.0047589740715920925, + 0.004867494571954012, + 0.006825540214776993, + 0.009237764403223991, + 0.007227352354675531, + 0.005820719990879297, + 0.0035663852468132973, + 0.008500082418322563, + 0.007537040859460831, + 0.004171403124928474, + 0.004165870603173971, + 0.0023600657004863024, + 0.0023372292052954435, + 0.0032440489158034325, + 0.004297245293855667, + 0.002332798670977354, + 0.0023812069557607174, + 0.0029284614138305187, + 0.002515136729925871, + 0.016999270766973495, + 0.0201001837849617, + 0.029814474284648895, + 0.024694403633475304, + 0.018454622477293015, + 0.018508968874812126, + 0.024730529636144638, + 0.023443102836608887, + 0.02273072488605976, + 0.019179031252861023, + 0.01581599935889244, + 0.02085745334625244, + 0.030224855989217758, + 0.038854483515024185, + 0.014012914150953293, + 0.015308230184018612, + 0.018532520160079002, + 0.007618485018610954, + 0.007851258851587772, + 0.006334446370601654, + 0.00362442247569561, + 0.014160221442580223, + 0.006914797239005566, + 0.012767712585628033, + 0.01404380053281784, + 0.009908120147883892, + 0.014837252907454967, + 0.011422032490372658, + 0.009542376734316349, + 0.015873873606324196, + 0.013610231690108776, + 0.011937995441257954, + 0.04272544011473656, + 0.01988164708018303, + 0.031012805178761482, + 0.01931017078459263, + 0.021452728658914566, + 0.036524105817079544, + 0.02287127822637558, + 0.023059355095028877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.002822702983394265 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054945945739746094 + ], + "opposite_logprob_tokens": [ + -2.929945945739746 + ], + "logprob_reward": -0.054945945739746094, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054945945739746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0956968441605568, + 0.13721702992916107, + 0.08459609001874924, + 0.04085333272814751, + 0.05785631388425827, + 0.05330757051706314, + 0.11651219427585602, + 0.13207687437534332, + 0.09636711329221725, + 0.18551664054393768 + ], + "credit_map": [ + 0.012483022175729275, + 0.013911879621446133, + 0.009670626372098923, + 0.010544104501605034, + 0.011946599930524826, + 0.011744960211217403, + 0.010270019061863422, + 0.015125637874007225, + 0.018061328679323196, + 0.017758021131157875, + 0.030312933027744293, + 0.01572125405073166, + 0.012049700133502483, + 0.012987288646399975, + 0.009837460704147816, + 0.020489055663347244, + 0.010360348038375378, + 0.009094920940697193, + 0.011162257753312588, + 0.00727537926286459, + 0.012970561161637306, + 0.010920255444943905, + 0.009369396604597569, + 0.013442978262901306, + 0.006879299413412809, + 0.005402196664363146, + 0.006145417224615812, + 0.003790320362895727, + 0.0051141418516635895, + 0.006236850284039974, + 0.0031842393800616264, + 0.00410086615011096, + 0.005215155892074108, + 0.006693690549582243, + 0.010766965337097645, + 0.006851362530142069, + 0.009555215947329998, + 0.008469694294035435, + 0.005365863908082247, + 0.004938366822898388, + 0.004450323525816202, + 0.006031534634530544, + 0.009926646016538143, + 0.006788555532693863, + 0.006506435107439756, + 0.0044640121050179005, + 0.009127489291131496, + 0.006012577097862959, + 0.011694078333675861, + 0.013037191703915596, + 0.013232521712779999, + 0.01605124957859516, + 0.015351432375609875, + 0.01613803394138813, + 0.012925590388476849, + 0.01808210089802742, + 0.03843860700726509, + 0.013757671229541302, + 0.0068495674058794975, + 0.010004731826484203, + 0.006743466481566429, + 0.0322093665599823, + 0.010435910895466805, + 0.013637561351060867, + 0.00911999586969614, + 0.01084038708359003, + 0.011563075706362724, + 0.012865007854998112, + 0.010836697183549404, + 0.013859112747013569, + 0.01363116130232811, + 0.013651678338646889, + 0.032041095197200775, + 0.016357790678739548, + 0.024314196780323982, + 0.02075923979282379, + 0.02065950445830822, + 0.028671137988567352, + 0.02082238160073757, + 0.021891308948397636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.054945945739746094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "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": [ + 0.12850765883922577, + 0.07283347100019455, + 0.07318974286317825, + 0.11601865291595459, + 0.03366871550679207, + 0.048244789242744446, + 0.09556134790182114, + 0.1056927889585495, + 0.14525222778320312, + 0.18103060126304626 + ], + "credit_map": [ + 0.015561770647764206, + 0.014450904913246632, + 0.01650196872651577, + 0.014510421082377434, + 0.015390894375741482, + 0.013232696801424026, + 0.014770871959626675, + 0.024088138714432716, + 0.008255556225776672, + 0.01616746187210083, + 0.008457974530756474, + 0.006311458069831133, + 0.010089431889355183, + 0.008155531249940395, + 0.006166303064674139, + 0.00922974944114685, + 0.008493294939398766, + 0.00829842034727335, + 0.010754425078630447, + 0.006851935759186745, + 0.008748150430619717, + 0.008355376310646534, + 0.011501641944050789, + 0.010186504572629929, + 0.019111912697553635, + 0.018717322498559952, + 0.014848721213638783, + 0.006778784096240997, + 0.01797044649720192, + 0.019381875172257423, + 0.009382868185639381, + 0.009826726280152798, + 0.0033600940369069576, + 0.003446341259405017, + 0.006941322702914476, + 0.0036804794799536467, + 0.0050977314822375774, + 0.004652367904782295, + 0.0030852044001221657, + 0.003405173309147358, + 0.0050350516103208065, + 0.008815613575279713, + 0.00743291387334466, + 0.004326529800891876, + 0.007664281874895096, + 0.006660476792603731, + 0.0047987173311412334, + 0.0035112048499286175, + 0.013180317357182503, + 0.01055451761931181, + 0.011199014261364937, + 0.009779496118426323, + 0.012175076641142368, + 0.016077795997262, + 0.010795030742883682, + 0.011800099164247513, + 0.0214497372508049, + 0.012949131429195404, + 0.009989358484745026, + 0.011140668764710426, + 0.00902402400970459, + 0.013960560783743858, + 0.013153186067938805, + 0.014026113785803318, + 0.015618358738720417, + 0.011164274998009205, + 0.018086953088641167, + 0.021881988272070885, + 0.011228621006011963, + 0.02328445017337799, + 0.025289615616202354, + 0.01869797147810459, + 0.0325503870844841, + 0.017099441960453987, + 0.0238221175968647, + 0.019681548699736595, + 0.018292823806405067, + 0.03193167969584465, + 0.01943143643438816, + 0.018221169710159302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0036149173974990845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20150601863861084 + ], + "opposite_logprob_tokens": [ + -1.7015060186386108 + ], + "logprob_reward": -0.20150601863861084, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20150601863861084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07056855410337448, + 0.09660673141479492, + 0.03309894725680351, + 0.060966141521930695, + 0.03551984205842018, + 0.051418256014585495, + 0.1077689528465271, + 0.20093120634555817, + 0.11853480339050293, + 0.2245865911245346 + ], + "credit_map": [ + 0.008974717929959297, + 0.005536806304007769, + 0.012759138830006123, + 0.011595379561185837, + 0.005021247081458569, + 0.009864650666713715, + 0.006567454896867275, + 0.010249156504869461, + 0.013342270627617836, + 0.020842015743255615, + 0.012408546172082424, + 0.008030303753912449, + 0.016336465254426003, + 0.01106414943933487, + 0.006192495115101337, + 0.00839048158377409, + 0.005219147074967623, + 0.002542018424719572, + 0.0053219241090118885, + 0.0028761958237737417, + 0.002802702132612467, + 0.0029312947299331427, + 0.004584400914609432, + 0.00682125985622406, + 0.00956516619771719, + 0.006936176214367151, + 0.009073605760931969, + 0.007648053579032421, + 0.0060052149929106236, + 0.007421884220093489, + 0.006953139323741198, + 0.0073628975078463554, + 0.003295408096164465, + 0.004366376902908087, + 0.00452926941215992, + 0.005153879523277283, + 0.004917483311146498, + 0.004455822054296732, + 0.004656375851482153, + 0.00414522597566247, + 0.004683524835854769, + 0.0053360783495008945, + 0.008601428009569645, + 0.007528284564614296, + 0.004283678252249956, + 0.003690768964588642, + 0.010047189891338348, + 0.0072473022155463696, + 0.010316084139049053, + 0.0104344692081213, + 0.01843453198671341, + 0.013991568237543106, + 0.008343304507434368, + 0.013162771239876747, + 0.016407769173383713, + 0.016678452491760254, + 0.03905096277594566, + 0.015790926292538643, + 0.017809713259339333, + 0.023010265082120895, + 0.012414063327014446, + 0.04471002519130707, + 0.022895487025380135, + 0.02524973638355732, + 0.010217991657555103, + 0.010574410669505596, + 0.01719336025416851, + 0.016007034108042717, + 0.010024556890130043, + 0.02036811225116253, + 0.019539007917046547, + 0.014610324054956436, + 0.03017161414027214, + 0.020944418385624886, + 0.05837872624397278, + 0.02764752507209778, + 0.017415190115571022, + 0.022474905475974083, + 0.021901972591876984, + 0.02565223164856434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.20150601863861084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From an epic perspective, the camera rotates with slow, majestic movements, capturing a cosmic spectacle utterly immersed in the unbridled brushstrokes and impasto of Van Gogh's \"The Starry Night.\" At the center of the image lies Earth, yet it has been transformed beyond all recognition: the once azure ocean has become a vast, seething vortex, where molten crimson and glacial blue violently intertwine, collide, and tumble, forming a dynamic, charged storm of color that foreshadows a profound future transformation. Above the planet's sky, vast, vibrant ribbons of auroras dance gracefully in hues of emerald green, brilliant purple, and rosy pink. Their vibrant forms, like the breath of the universe, unfold against the inky blackness. Far from a static darkness, the cosmic space is teeming with Van Gogh's vitality. Stars are no longer simple points of light, but luminous yellow swirls and pulsating clusters, each ray a vibrant, thick coat of paint squeezed straight from a tube. As the camera slowly rotates, this chaotic and magnificent scene is presented in a panoramic view. The turbulent earth and the dancing auroras together form a visual symphony full of primitive vitality.", + "seed": 2627, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2627_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": -0.13573749992065132, + "vlm_logprob": -0.13573749992065132, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.13573749992065132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037288423627614975 + ], + "opposite_logprob_tokens": [ + -3.4122884273529053 + ], + "logprob_reward": -0.037288423627614975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037288423627614975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11519551277160645, + 0.03476111218333244, + 0.024344364181160927, + 0.03340192139148712, + 0.315219908952713, + 0.05825142562389374, + 0.2042795568704605, + 0.076301708817482, + 0.08135722577571869, + 0.05688722804188728 + ], + "credit_map": [ + 0.016959866508841515, + 0.023888971656560898, + 0.012110949493944645, + 0.010747378692030907, + 0.017074650153517723, + 0.010764437727630138, + 0.014654083177447319, + 0.008995174430310726, + 0.006092422176152468, + 0.006167940329760313, + 0.0027732192538678646, + 0.005454997532069683, + 0.003915442619472742, + 0.00424725329503417, + 0.0030658687464892864, + 0.0030439686961472034, + 0.0034794071689248085, + 0.0036744556855410337, + 0.0020484437700361013, + 0.0032133643981069326, + 0.002350542228668928, + 0.004793476779013872, + 0.0023817059118300676, + 0.0024029689375311136, + 0.009354927577078342, + 0.009198169223964214, + 0.002951796865090728, + 0.0022956575267016888, + 0.002149874810129404, + 0.002674691379070282, + 0.0028116737958043814, + 0.0019651283510029316, + 0.03434399515390396, + 0.04224960878491402, + 0.02722654677927494, + 0.041000109165906906, + 0.022100379690527916, + 0.05498143658041954, + 0.05599158629775047, + 0.037326257675886154, + 0.008922012522816658, + 0.007167486939579248, + 0.00457528792321682, + 0.005726615898311138, + 0.008197254501283169, + 0.007647557184100151, + 0.007641016039997339, + 0.008374195545911789, + 0.044428303837776184, + 0.009966947138309479, + 0.005156633909791708, + 0.004525746218860149, + 0.01387446466833353, + 0.06964477151632309, + 0.03538206219673157, + 0.02130063809454441, + 0.014940665103495121, + 0.010514532215893269, + 0.00767071358859539, + 0.012297690846025944, + 0.008211665786802769, + 0.008530757389962673, + 0.007671931758522987, + 0.006463753990828991, + 0.010985993780195713, + 0.013882859610021114, + 0.009184041060507298, + 0.008223744109272957, + 0.010985154658555984, + 0.01005515456199646, + 0.00871102511882782, + 0.009329253807663918, + 0.007491494994610548, + 0.012290825136005878, + 0.00451808562502265, + 0.00392639497295022, + 0.008358648978173733, + 0.007559848017990589, + 0.006739805452525616, + 0.0060021230019629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.037288423627614975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003617292968556285 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.003617292968556285, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003617292968556285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13108094036579132, + 0.05734574794769287, + 0.13624732196331024, + 0.0827002227306366, + 0.0945790484547615, + 0.05106667801737785, + 0.14633606374263763, + 0.09696996957063675, + 0.14624068140983582, + 0.057433322072029114 + ], + "credit_map": [ + 0.016530852764844894, + 0.01529870368540287, + 0.02055290900170803, + 0.01688319444656372, + 0.025473754853010178, + 0.009621006436645985, + 0.016226038336753845, + 0.010494491085410118, + 0.008345529437065125, + 0.008641098625957966, + 0.0057310787960886955, + 0.011255335062742233, + 0.007624538615345955, + 0.0050607165321707726, + 0.005788631737232208, + 0.004898824729025364, + 0.016220523044466972, + 0.026316620409488678, + 0.007462615612894297, + 0.010574180632829666, + 0.032921891659498215, + 0.02504885196685791, + 0.009733402170240879, + 0.007969247177243233, + 0.012837110087275505, + 0.02740156650543213, + 0.01001941878348589, + 0.005991433747112751, + 0.006335585378110409, + 0.0069414349272847176, + 0.007281223777681589, + 0.005892459768801928, + 0.010964891873300076, + 0.023698335513472557, + 0.006956818047910929, + 0.009336299262940884, + 0.009175263345241547, + 0.017700467258691788, + 0.009299888275563717, + 0.007447096984833479, + 0.005170060787349939, + 0.006555816158652306, + 0.004407734610140324, + 0.0038667600601911545, + 0.005452297627925873, + 0.003958405926823616, + 0.008172320201992989, + 0.013483288697898388, + 0.030231529846787453, + 0.008189352229237556, + 0.0037343602161854506, + 0.004509238991886377, + 0.0101912347599864, + 0.047767091542482376, + 0.025878533720970154, + 0.015834728255867958, + 0.019588101655244827, + 0.012076485902071, + 0.006435105111449957, + 0.009061257354915142, + 0.011757990345358849, + 0.018416427075862885, + 0.007619054988026619, + 0.012015566229820251, + 0.02140401490032673, + 0.02278289571404457, + 0.0133057227358222, + 0.025878161191940308, + 0.02059337869286537, + 0.017672749236226082, + 0.011549621820449829, + 0.013054145500063896, + 0.007524981629103422, + 0.0072441124357283115, + 0.0034756942186504602, + 0.004748356994241476, + 0.007813000120222569, + 0.006706322077661753, + 0.012546801939606667, + 0.0073740556836128235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.003617292968556285 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054956551641225815 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.054956551641225815, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054956551641225815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16868889331817627, + 0.09944487363100052, + 0.06871780008077621, + 0.028049742802977562, + 0.15253101289272308, + 0.05250026285648346, + 0.19835853576660156, + 0.11521212011575699, + 0.053602464497089386, + 0.06289425492286682 + ], + "credit_map": [ + 0.027595501393079758, + 0.04122619330883026, + 0.020511876791715622, + 0.0167049840092659, + 0.02015629969537258, + 0.013077649287879467, + 0.017505286261439323, + 0.011911102570593357, + 0.01979244314134121, + 0.0154221560806036, + 0.005666810087859631, + 0.01129540242254734, + 0.018504459410905838, + 0.012587231583893299, + 0.008108413778245449, + 0.008067956194281578, + 0.009628025814890862, + 0.014252293854951859, + 0.006134539842605591, + 0.005915129091590643, + 0.007864896208047867, + 0.012507077306509018, + 0.006739978212863207, + 0.005675857421010733, + 0.0036896849051117897, + 0.0061495075933635235, + 0.0031807099003344774, + 0.002788288053125143, + 0.003051258623600006, + 0.002609026152640581, + 0.004435079637914896, + 0.0021461863070726395, + 0.015151994302868843, + 0.026839416474103928, + 0.0117754265666008, + 0.017483172938227654, + 0.01186683215200901, + 0.02714364416897297, + 0.02504093013703823, + 0.017229599878191948, + 0.003607372520491481, + 0.009664465673267841, + 0.004268673248589039, + 0.0030202518682926893, + 0.0024612820707261562, + 0.008280882611870766, + 0.008369640447199345, + 0.012827692553400993, + 0.04581330716609955, + 0.011112404987215996, + 0.00527182687073946, + 0.004568094853311777, + 0.014341317117214203, + 0.06610777974128723, + 0.030926160514354706, + 0.02021763101220131, + 0.025572340935468674, + 0.015242268331348896, + 0.010807017795741558, + 0.015327985398471355, + 0.012336363084614277, + 0.01446224469691515, + 0.0111334677785635, + 0.010330431163311005, + 0.007151920348405838, + 0.005837906152009964, + 0.004301389213651419, + 0.008837862871587276, + 0.0098084332421422, + 0.005593847017735243, + 0.006466728635132313, + 0.005604375619441271, + 0.007400230970233679, + 0.008081205189228058, + 0.003829751629382372, + 0.002616163343191147, + 0.016996115446090698, + 0.01567581482231617, + 0.003236930351704359, + 0.005058041773736477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.054956551641225815 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733584217727184 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733584217727184, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733584217727184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14676544070243835, + 0.051016148179769516, + 0.06535658985376358, + 0.04151212424039841, + 0.23840080201625824, + 0.0752367153763771, + 0.06637571007013321, + 0.06035272404551506, + 0.13066312670707703, + 0.12432058155536652 + ], + "credit_map": [ + 0.015599075704813004, + 0.020858293399214745, + 0.02093549072742462, + 0.01889459788799286, + 0.020376048982143402, + 0.01597479172050953, + 0.0199419092386961, + 0.014185243286192417, + 0.00778413750231266, + 0.008447662927210331, + 0.0036485749296844006, + 0.005511516239494085, + 0.011820454150438309, + 0.0054287174716591835, + 0.004314045421779156, + 0.004061042796820402, + 0.01100919023156166, + 0.012441088445484638, + 0.0034513568971306086, + 0.004543766845017672, + 0.01949981227517128, + 0.005742393899708986, + 0.004185891710221767, + 0.004483094438910484, + 0.005323100835084915, + 0.009059949778020382, + 0.004529061261564493, + 0.0030280740465968847, + 0.004883170127868652, + 0.004484752658754587, + 0.0070131090469658375, + 0.003190909279510379, + 0.027376921847462654, + 0.02547447942197323, + 0.023475971072912216, + 0.03531523048877716, + 0.01638467237353325, + 0.03918173536658287, + 0.04236197471618652, + 0.028829826042056084, + 0.008866540156304836, + 0.013714579865336418, + 0.008965344168245792, + 0.003957755863666534, + 0.005998463835567236, + 0.01338062435388565, + 0.008959811180830002, + 0.011393607594072819, + 0.01857644133269787, + 0.005552876275032759, + 0.003092948580160737, + 0.00467478483915329, + 0.007305173669010401, + 0.016942156478762627, + 0.003937987610697746, + 0.006293342914432287, + 0.012581149116158485, + 0.008601957000792027, + 0.003727520816028118, + 0.005989392288029194, + 0.0076015726663172245, + 0.011108221486210823, + 0.004505819175392389, + 0.006237095221877098, + 0.019842784851789474, + 0.01597682386636734, + 0.007640734780579805, + 0.013442975468933582, + 0.02251814305782318, + 0.01813974231481552, + 0.016724204644560814, + 0.016377722844481468, + 0.012196648865938187, + 0.014518667943775654, + 0.008366759866476059, + 0.006439353805035353, + 0.03151290863752365, + 0.027747849002480507, + 0.01166339498013258, + 0.011875009164214134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.006733584217727184 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760916471481323 + ], + "opposite_logprob_tokens": [ + -0.8260916471481323 + ], + "logprob_reward": -0.5760916471481323, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760916471481323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12257308512926102, + 0.12771375477313995, + 0.05501847714185715, + 0.12565350532531738, + 0.1740752011537552, + 0.05634073540568352, + 0.1425911784172058, + 0.06347887217998505, + 0.07776616513729095, + 0.05478901416063309 + ], + "credit_map": [ + 0.016410764306783676, + 0.01829092763364315, + 0.015599190257489681, + 0.016800543293356895, + 0.018515510484576225, + 0.010690638795495033, + 0.013855931349098682, + 0.012409568764269352, + 0.015614795498549938, + 0.019036762416362762, + 0.012633686885237694, + 0.03245330974459648, + 0.014491879381239414, + 0.010144790634512901, + 0.013646001927554607, + 0.009692533873021603, + 0.009703900665044785, + 0.008842460811138153, + 0.0049500735476613045, + 0.00671281386166811, + 0.009067806415259838, + 0.00584509689360857, + 0.004465468693524599, + 0.005430861376225948, + 0.021557854488492012, + 0.03839460015296936, + 0.013663883320987225, + 0.007918321527540684, + 0.011940923519432545, + 0.013439353555440903, + 0.010182034224271774, + 0.008556530810892582, + 0.0190211720764637, + 0.03066342882812023, + 0.011598167009651661, + 0.01698531024158001, + 0.014228629879653454, + 0.0318470224738121, + 0.029365001246333122, + 0.020366478711366653, + 0.006912206765264273, + 0.0059407916851341724, + 0.005485659930855036, + 0.004354647360742092, + 0.0049024163745343685, + 0.007144757080823183, + 0.009032915346324444, + 0.01256734412163496, + 0.03909514844417572, + 0.006698715966194868, + 0.004141448065638542, + 0.003051616484299302, + 0.0104499701410532, + 0.0476299449801445, + 0.01673167198896408, + 0.0147926677018404, + 0.007648852188140154, + 0.00823026243597269, + 0.004376079421490431, + 0.017318136990070343, + 0.004947478882968426, + 0.007700471673160791, + 0.006596029736101627, + 0.00666156318038702, + 0.010132594965398312, + 0.01184114534407854, + 0.007735076360404491, + 0.008312436752021313, + 0.013936372473835945, + 0.01056534145027399, + 0.008168117143213749, + 0.0070750839076936245, + 0.006786218378692865, + 0.007669053506106138, + 0.0038032170850783587, + 0.002529425546526909, + 0.01424320600926876, + 0.011338776908814907, + 0.004292691126465797, + 0.00412642490118742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.5760916471481323 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From an epic perspective, the camera rotates with slow, majestic movements, capturing a cosmic spectacle utterly immersed in the unbridled brushstrokes and impasto of Van Gogh's \"The Starry Night.\" At the center of the image lies Earth, yet it has been transformed beyond all recognition: the once azure ocean has become a vast, seething vortex, where molten crimson and glacial blue violently intertwine, collide, and tumble, forming a dynamic, charged storm of color that foreshadows a profound future transformation. Above the planet's sky, vast, vibrant ribbons of auroras dance gracefully in hues of emerald green, brilliant purple, and rosy pink. Their vibrant forms, like the breath of the universe, unfold against the inky blackness. Far from a static darkness, the cosmic space is teeming with Van Gogh's vitality. Stars are no longer simple points of light, but luminous yellow swirls and pulsating clusters, each ray a vibrant, thick coat of paint squeezed straight from a tube. As the camera slowly rotates, this chaotic and magnificent scene is presented in a panoramic view. The turbulent earth and the dancing auroras together form a visual symphony full of primitive vitality.", + "seed": 2628, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2628_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": -0.2648338370025158, + "vlm_logprob": -0.2648338370025158, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.2648338370025158, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.695521891117096 + ], + "opposite_logprob_tokens": [ + -0.695521891117096 + ], + "logprob_reward": -0.695521891117096, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.695521891117096, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11664324998855591, + 0.16807620227336884, + 0.06207146868109703, + 0.010379839688539505, + 0.026662543416023254, + 0.4364958703517914, + 0.09523814171552658, + 0.02907930128276348, + 0.026206502690911293, + 0.029146894812583923 + ], + "credit_map": [ + 0.015473256818950176, + 0.029952241107821465, + 0.014769298024475574, + 0.008946673013269901, + 0.01334623247385025, + 0.010238966904580593, + 0.010044461116194725, + 0.013872120529413223, + 0.014125630259513855, + 0.05752601474523544, + 0.012847267091274261, + 0.005377497989684343, + 0.012644928880035877, + 0.005249391309916973, + 0.005278443451970816, + 0.055027034133672714, + 0.006592599209398031, + 0.02193625271320343, + 0.0032302732579410076, + 0.002869262360036373, + 0.00903995893895626, + 0.0029924644622951746, + 0.0027659113984555006, + 0.012644746340811253, + 0.001624282798729837, + 0.0010965655092149973, + 0.0009316041832789779, + 0.001074507599696517, + 0.0010473214788362384, + 0.00087001669453457, + 0.0019803193863481283, + 0.0017552225617691875, + 0.003393645631149411, + 0.003099197056144476, + 0.005350430961698294, + 0.002539559733122587, + 0.002560548484325409, + 0.001761116785928607, + 0.00436710799112916, + 0.003590937703847885, + 0.043656446039676666, + 0.17678046226501465, + 0.024598222225904465, + 0.05737245827913284, + 0.040822770446538925, + 0.019560912624001503, + 0.01878580078482628, + 0.054918788373470306, + 0.016789250075817108, + 0.012425525113940239, + 0.016049817204475403, + 0.009866876527667046, + 0.00683458149433136, + 0.007711812388151884, + 0.005741684231907129, + 0.01981859840452671, + 0.005566594190895557, + 0.004465916659682989, + 0.004014450125396252, + 0.002404800383374095, + 0.0027307963464409113, + 0.0033745162654668093, + 0.0027529180515557528, + 0.0037693106569349766, + 0.006477090995758772, + 0.002675434108823538, + 0.0021348497830331326, + 0.0013924483209848404, + 0.0015907012857496738, + 0.001844177138991654, + 0.0016517095500603318, + 0.008440094068646431, + 0.00843601394444704, + 0.002240247093141079, + 0.0018699797801673412, + 0.002357091521844268, + 0.0014792666770517826, + 0.0019904752261936665, + 0.0026489943265914917, + 0.008124826475977898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.695521891117096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078514158725739 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078514158725739, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078514158725739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07030284404754639, + 0.1865352988243103, + 0.036835357546806335, + 0.029126925393939018, + 0.03920866921544075, + 0.39342233538627625, + 0.08324796706438065, + 0.029355991631746292, + 0.0297743771225214, + 0.10219034552574158 + ], + "credit_map": [ + 0.01078027207404375, + 0.009913908317685127, + 0.011701325885951519, + 0.007059903349727392, + 0.005994257051497698, + 0.004805244971066713, + 0.008254732005298138, + 0.011793186888098717, + 0.014911104924976826, + 0.06355741620063782, + 0.014844908379018307, + 0.005648922175168991, + 0.013836597092449665, + 0.005769167561084032, + 0.00525136012583971, + 0.0627157911658287, + 0.004696906544268131, + 0.011678480543196201, + 0.0025985143147408962, + 0.0024790186434984207, + 0.005100285168737173, + 0.0022229370661079884, + 0.0022201216779649258, + 0.0058390856720507145, + 0.004028285853564739, + 0.003624240169301629, + 0.003667139448225498, + 0.0026996030937880278, + 0.0027667724061757326, + 0.0018763565458357334, + 0.005713579710572958, + 0.004750943277031183, + 0.003804735140874982, + 0.003985049668699503, + 0.008194983936846256, + 0.004038502927869558, + 0.0035952902399003506, + 0.002204887568950653, + 0.00763347651809454, + 0.005751739721745253, + 0.04021709784865379, + 0.16007089614868164, + 0.022975677624344826, + 0.055356159806251526, + 0.036988265812397, + 0.016673972830176353, + 0.017099009826779366, + 0.04404118284583092, + 0.015288352966308594, + 0.0115803312510252, + 0.014188120141625404, + 0.008317875675857067, + 0.006666109431535006, + 0.006487214006483555, + 0.004573534242808819, + 0.01614641770720482, + 0.004321020096540451, + 0.00398814445361495, + 0.004987581167370081, + 0.003645871067419648, + 0.0035604347940534353, + 0.0029373997822403908, + 0.0029313445556908846, + 0.0029841919895261526, + 0.007145150098949671, + 0.0036328325513750315, + 0.00210165255703032, + 0.0014484486309811473, + 0.0021753821056336164, + 0.0021723639219999313, + 0.0016941721551120281, + 0.009404370561242104, + 0.03449828922748566, + 0.005729337222874165, + 0.007058913819491863, + 0.004657301586121321, + 0.003490377916023135, + 0.006180957891047001, + 0.005598291289061308, + 0.034976862370967865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.011078514158725739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904709666967392 + ], + "opposite_logprob_tokens": [ + -2.4640471935272217 + ], + "logprob_reward": -0.08904709666967392, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904709666967392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10056443512439728, + 0.22154198586940765, + 0.04665236547589302, + 0.019132498651742935, + 0.026016177609562874, + 0.35044020414352417, + 0.06543517112731934, + 0.058042071759700775, + 0.06419063359498978, + 0.04798448085784912 + ], + "credit_map": [ + 0.013282202184200287, + 0.01405555009841919, + 0.013774924911558628, + 0.007444370537996292, + 0.008212818764150143, + 0.005340980365872383, + 0.011625324375927448, + 0.02682826854288578, + 0.03268559277057648, + 0.09201792627573013, + 0.009213535115122795, + 0.007572769187390804, + 0.014810241758823395, + 0.005579250864684582, + 0.005143254529684782, + 0.05451943352818489, + 0.005533700808882713, + 0.009940855205059052, + 0.004938235506415367, + 0.00405783299356699, + 0.005057442933320999, + 0.005293452646583319, + 0.004256492480635643, + 0.007574353367090225, + 0.0030504337046295404, + 0.0026233126409351826, + 0.00213329354301095, + 0.0015776912914589047, + 0.0019207014702260494, + 0.0016708060866221786, + 0.0028137359768152237, + 0.003342523705214262, + 0.004983164835721254, + 0.0028279204852879047, + 0.0023575089871883392, + 0.002192575251683593, + 0.0027861122507601976, + 0.0017061427934095263, + 0.0056671700440347195, + 0.003495581913739443, + 0.0392354317009449, + 0.12761400640010834, + 0.024457918480038643, + 0.06377921253442764, + 0.03632756322622299, + 0.01893787644803524, + 0.01553055364638567, + 0.024557633325457573, + 0.014996337704360485, + 0.008599676191806793, + 0.009755323641002178, + 0.0037422676105052233, + 0.004850056488066912, + 0.007707591634243727, + 0.00516159925609827, + 0.01062232255935669, + 0.007315595168620348, + 0.00519794225692749, + 0.01191639807075262, + 0.0072416584007442, + 0.00577687518671155, + 0.004092399496585131, + 0.005140185821801424, + 0.011361016891896725, + 0.011171882040798664, + 0.007833056151866913, + 0.00944629404693842, + 0.007173218298703432, + 0.005843947641551495, + 0.00550676416605711, + 0.008096480742096901, + 0.009118989109992981, + 0.01029166392982006, + 0.005315415561199188, + 0.003584768157452345, + 0.004595538601279259, + 0.0033860562834888697, + 0.0044638365507125854, + 0.004049953073263168, + 0.012297247536480427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.08904709666967392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249407142400742 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249407142400742, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249407142400742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06987294554710388, + 0.1289491206407547, + 0.03330441191792488, + 0.009919398464262486, + 0.02441277727484703, + 0.5481623411178589, + 0.058215536177158356, + 0.026806997135281563, + 0.05521100386977196, + 0.04514544829726219 + ], + "credit_map": [ + 0.02003053016960621, + 0.019493240863084793, + 0.004347427748143673, + 0.0031351482030004263, + 0.0054123178124427795, + 0.0035418851766735315, + 0.004211167339235544, + 0.009701233357191086, + 0.01863216422498226, + 0.05384232476353645, + 0.005363159812986851, + 0.0043399520218372345, + 0.008374591358006, + 0.003310395637527108, + 0.003359213238582015, + 0.03172732889652252, + 0.003303257981315255, + 0.01011255569756031, + 0.0027693205047398806, + 0.002103228121995926, + 0.003965562209486961, + 0.0026286656502634287, + 0.0023948848247528076, + 0.006026940420269966, + 0.0012319477973505855, + 0.0009695898625068367, + 0.0011776885949075222, + 0.0009464044123888016, + 0.0013456755550578237, + 0.0010331792291253805, + 0.001741744694299996, + 0.0014731701230630279, + 0.004981569014489651, + 0.002950392896309495, + 0.0021945734042674303, + 0.0021407410968095064, + 0.0020851900335401297, + 0.0018025285098701715, + 0.004559661261737347, + 0.0036981203593313694, + 0.05803433433175087, + 0.2226652055978775, + 0.030734580010175705, + 0.08895841240882874, + 0.051558881998062134, + 0.020146198570728302, + 0.024882540106773376, + 0.05118226632475853, + 0.010036226361989975, + 0.008322048000991344, + 0.009904737584292889, + 0.0049013979732990265, + 0.003915917593985796, + 0.005872448440641165, + 0.0036197220906615257, + 0.011643042787909508, + 0.0034975134767591953, + 0.0028884063940495253, + 0.005696954671293497, + 0.002868249313905835, + 0.0025709550827741623, + 0.0033724980894476175, + 0.0019054507138207555, + 0.004006970673799515, + 0.015174617059528828, + 0.003370081540197134, + 0.005170315504074097, + 0.003556731389835477, + 0.003040770534425974, + 0.002638539532199502, + 0.0033952007070183754, + 0.018864748999476433, + 0.014910471625626087, + 0.0031469331588596106, + 0.0027745477855205536, + 0.0019550970755517483, + 0.002356880344450474, + 0.0030211147386580706, + 0.003080567345023155, + 0.013899839483201504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.005249407142400742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5232722759246826 + ], + "opposite_logprob_tokens": [ + -0.8982722759246826 + ], + "logprob_reward": -0.5232722759246826, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5232722759246826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11362528800964355, + 0.05248778685927391, + 0.043945297598838806, + 0.01825028657913208, + 0.036423467099666595, + 0.30183178186416626, + 0.12197593599557877, + 0.184843048453331, + 0.09776554256677628, + 0.028851477429270744 + ], + "credit_map": [ + 0.01117779966443777, + 0.03382320702075958, + 0.007659874856472015, + 0.00648118369281292, + 0.012725844979286194, + 0.007797805592417717, + 0.009224912151694298, + 0.024734672158956528, + 0.009903188794851303, + 0.008084431290626526, + 0.006527298595756292, + 0.0030990152154117823, + 0.007000660058110952, + 0.003407801501452923, + 0.0052951122634112835, + 0.009170284494757652, + 0.004263833165168762, + 0.008228020742535591, + 0.007578563876450062, + 0.004378238692879677, + 0.003820563666522503, + 0.002130685606971383, + 0.006732643581926823, + 0.006812752224504948, + 0.002684391103684902, + 0.0019249817123636603, + 0.0019331909716129303, + 0.0012953937985002995, + 0.0024727433919906616, + 0.001529355300590396, + 0.0030393609777092934, + 0.003370872000232339, + 0.004630519077181816, + 0.004365593194961548, + 0.006229868624359369, + 0.0020537569653242826, + 0.004202839452773333, + 0.002544156275689602, + 0.00728514464572072, + 0.005111591890454292, + 0.03629164397716522, + 0.12903574109077454, + 0.016809212043881416, + 0.018249299377202988, + 0.025598563253879547, + 0.015761585906147957, + 0.010895992629230022, + 0.049189768731594086, + 0.017995044589042664, + 0.020716816186904907, + 0.01642225682735443, + 0.00984967965632677, + 0.016448477283120155, + 0.017654161900281906, + 0.006485360208898783, + 0.016404159367084503, + 0.024034248664975166, + 0.01625337265431881, + 0.036415014415979385, + 0.023130791261792183, + 0.014863225631415844, + 0.014675106853246689, + 0.01745002530515194, + 0.038021285086870193, + 0.01982966624200344, + 0.010943764820694923, + 0.016673285514116287, + 0.008519109338521957, + 0.007122716400772333, + 0.010191475972533226, + 0.009066276252269745, + 0.015419257804751396, + 0.006235936190932989, + 0.0031263730488717556, + 0.0028479157481342554, + 0.002340886276215315, + 0.0017769865226000547, + 0.0026515531353652477, + 0.002458159113302827, + 0.0074136704206466675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.5232722759246826 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From an epic perspective, the camera rotates with slow, majestic movements, capturing a cosmic spectacle utterly immersed in the unbridled brushstrokes and impasto of Van Gogh's \"The Starry Night.\" At the center of the image lies Earth, yet it has been transformed beyond all recognition: the once azure ocean has become a vast, seething vortex, where molten crimson and glacial blue violently intertwine, collide, and tumble, forming a dynamic, charged storm of color that foreshadows a profound future transformation. Above the planet's sky, vast, vibrant ribbons of auroras dance gracefully in hues of emerald green, brilliant purple, and rosy pink. Their vibrant forms, like the breath of the universe, unfold against the inky blackness. Far from a static darkness, the cosmic space is teeming with Van Gogh's vitality. Stars are no longer simple points of light, but luminous yellow swirls and pulsating clusters, each ray a vibrant, thick coat of paint squeezed straight from a tube. As the camera slowly rotates, this chaotic and magnificent scene is presented in a panoramic view. The turbulent earth and the dancing auroras together form a visual symphony full of primitive vitality.", + "seed": 2629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_2/index_9151_seed_2629_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": -0.09294407069683075, + "vlm_logprob": -0.09294407069683075, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.09294407069683075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014575551263988018 + ], + "opposite_logprob_tokens": [ + -4.764575481414795 + ], + "logprob_reward": -0.014575551263988018, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014575551263988018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17201821506023407, + 0.26414424180984497, + 0.046412691473960876, + 0.09173445403575897, + 0.08583491295576096, + 0.04089836776256561, + 0.05368282273411751, + 0.10740847885608673, + 0.08304145187139511, + 0.0548243448138237 + ], + "credit_map": [ + 0.02359866350889206, + 0.025854140520095825, + 0.014174004085361958, + 0.018181612715125084, + 0.012069701217114925, + 0.026716530323028564, + 0.022691426798701286, + 0.028732135891914368, + 0.025846973061561584, + 0.025890113785862923, + 0.02801128290593624, + 0.03464983403682709, + 0.030771588906645775, + 0.03924563154578209, + 0.033772192895412445, + 0.04595661908388138, + 0.005209026858210564, + 0.006976919248700142, + 0.004838311113417149, + 0.004817300476133823, + 0.004028682596981525, + 0.007923553697764874, + 0.005155004095286131, + 0.0074638924561440945, + 0.007144646253436804, + 0.008181625977158546, + 0.012796907685697079, + 0.014664357528090477, + 0.008889930322766304, + 0.012426652945578098, + 0.011052712798118591, + 0.016577623784542084, + 0.01449548825621605, + 0.01045049075037241, + 0.009959197603166103, + 0.007397498469799757, + 0.009182021021842957, + 0.009587282314896584, + 0.012763469479978085, + 0.011999471113085747, + 0.004489170387387276, + 0.004146126098930836, + 0.005150264594703913, + 0.00544313807040453, + 0.0036618385929614305, + 0.004576138686388731, + 0.00751021783798933, + 0.005921473260968924, + 0.009205841459333897, + 0.006232287734746933, + 0.010715070180594921, + 0.00543640973046422, + 0.0037966922391206026, + 0.005194785539060831, + 0.004723833873867989, + 0.008377902209758759, + 0.011438955552875996, + 0.017649000510573387, + 0.01014223974198103, + 0.015276813879609108, + 0.020081652328372, + 0.01022480707615614, + 0.01176793035119772, + 0.010827072896063328, + 0.011075302958488464, + 0.0134587362408638, + 0.005965590942651033, + 0.006264453753829002, + 0.017328524962067604, + 0.01598670519888401, + 0.006828417535871267, + 0.006133721675723791, + 0.005774920340627432, + 0.008694412186741829, + 0.004992613568902016, + 0.0051021818071603775, + 0.0076214284636080265, + 0.00633664708584547, + 0.006719387602061033, + 0.009582750499248505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.014575551263988018 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040998938493430614 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.0040998938493430614, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040998938493430614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08199185878038406, + 0.11372620612382889, + 0.04670565575361252, + 0.12786708772182465, + 0.20427952706813812, + 0.058124832808971405, + 0.030664309859275818, + 0.05734124779701233, + 0.2571720778942108, + 0.02212715707719326 + ], + "credit_map": [ + 0.011187650263309479, + 0.008752038702368736, + 0.00775308208540082, + 0.006587315816432238, + 0.008081946521997452, + 0.01748736761510372, + 0.009200834669172764, + 0.012941626831889153, + 0.020302869379520416, + 0.02024918794631958, + 0.007789029274135828, + 0.011775200255215168, + 0.009209147654473782, + 0.017693666741251945, + 0.012577513232827187, + 0.014129591174423695, + 0.005568076856434345, + 0.008608170785009861, + 0.004648655652999878, + 0.004428456071764231, + 0.005812898278236389, + 0.007885084487497807, + 0.004529017023742199, + 0.005225295200943947, + 0.009556549601256847, + 0.01337230484932661, + 0.016728565096855164, + 0.017288759350776672, + 0.01418275199830532, + 0.018055565655231476, + 0.014253364875912666, + 0.02442922815680504, + 0.02695711888372898, + 0.020797764882445335, + 0.029548121616244316, + 0.023398779332637787, + 0.017414884641766548, + 0.020268866792321205, + 0.02952372469007969, + 0.03637027367949486, + 0.007668337319046259, + 0.007319280412048101, + 0.00751660717651248, + 0.005980479996651411, + 0.0059827882796525955, + 0.007098378147929907, + 0.009342524223029613, + 0.0072164349257946014, + 0.0047817411832511425, + 0.003889217507094145, + 0.004439754877239466, + 0.002946282271295786, + 0.0036396640352904797, + 0.0046506356447935104, + 0.003235775977373123, + 0.0030812404584139585, + 0.006140686571598053, + 0.011111704632639885, + 0.00605388218536973, + 0.007278351113200188, + 0.006924898829311132, + 0.004914416465908289, + 0.00870492309331894, + 0.006212389096617699, + 0.03508109599351883, + 0.038588665425777435, + 0.020747127011418343, + 0.025791723281145096, + 0.03606317937374115, + 0.046188149601221085, + 0.034914787858724594, + 0.019797353073954582, + 0.002775618340820074, + 0.0035658266860991716, + 0.002055957680568099, + 0.0024070932995527983, + 0.003959746100008488, + 0.003373020561411977, + 0.0020170712377876043, + 0.0019728231709450483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.0040998938493430614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054947640746831894 + ], + "opposite_logprob_tokens": [ + -2.9299476146698 + ], + "logprob_reward": -0.054947640746831894, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054947640746831894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2094758003950119, + 0.11597166955471039, + 0.060025230050086975, + 0.15064775943756104, + 0.15967567265033722, + 0.044692255556583405, + 0.044927895069122314, + 0.03289008140563965, + 0.1433054506778717, + 0.03838815540075302 + ], + "credit_map": [ + 0.031801387667655945, + 0.02458413504064083, + 0.021791690960526466, + 0.021024301648139954, + 0.01622508466243744, + 0.03521042689681053, + 0.029526272788643837, + 0.029312506318092346, + 0.009743883274495602, + 0.013267731294035912, + 0.008971719071269035, + 0.014457641169428825, + 0.006845741532742977, + 0.01819525472819805, + 0.02038334123790264, + 0.024106355383992195, + 0.005843062419444323, + 0.008066002279520035, + 0.00494047487154603, + 0.005990064237266779, + 0.006699264049530029, + 0.009883723221719265, + 0.007644904777407646, + 0.010957734659314156, + 0.01314388308674097, + 0.020485132932662964, + 0.023070428520441055, + 0.025768814608454704, + 0.018269013613462448, + 0.018146023154258728, + 0.015575411729514599, + 0.01618904620409012, + 0.01686668023467064, + 0.01587684452533722, + 0.022721562534570694, + 0.01901988312602043, + 0.015253228135406971, + 0.016039639711380005, + 0.02214435487985611, + 0.03175348788499832, + 0.005825752392411232, + 0.005222339183092117, + 0.006865493021905422, + 0.004672639537602663, + 0.004109376575797796, + 0.0050865295343101025, + 0.007236599922180176, + 0.00567352632060647, + 0.00842360407114029, + 0.005718831438571215, + 0.006607810966670513, + 0.005457596853375435, + 0.004539022222161293, + 0.004683084785938263, + 0.004191773012280464, + 0.00530617032200098, + 0.0031790274661034346, + 0.003546723397448659, + 0.0029505379498004913, + 0.007408604957163334, + 0.0038944620173424482, + 0.003655150765553117, + 0.004757039714604616, + 0.003498537465929985, + 0.020166080445051193, + 0.020444966852664948, + 0.011371780186891556, + 0.013972572050988674, + 0.022241415455937386, + 0.02670237608253956, + 0.01805516518652439, + 0.010351086966693401, + 0.00567425275221467, + 0.005699805915355682, + 0.003954686224460602, + 0.0033663727808743715, + 0.006463733501732349, + 0.006109066307544708, + 0.0036392284091562033, + 0.003481011837720871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.054947640746831894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092176910489798 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.004092176910489798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092176910489798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13002757728099823, + 0.1707880198955536, + 0.04596664384007454, + 0.16940994560718536, + 0.17745932936668396, + 0.05333290621638298, + 0.06150243803858757, + 0.06180163100361824, + 0.04519931226968765, + 0.08451221138238907 + ], + "credit_map": [ + 0.01947700046002865, + 0.01955305226147175, + 0.010224382393062115, + 0.009650436230003834, + 0.00824552122503519, + 0.020012840628623962, + 0.01931624859571457, + 0.02354808896780014, + 0.021937645971775055, + 0.02137427031993866, + 0.0176271740347147, + 0.02315574139356613, + 0.015130558051168919, + 0.024894505739212036, + 0.02079005353152752, + 0.02587808109819889, + 0.003680960275232792, + 0.0062124826945364475, + 0.004819720517843962, + 0.0036346102133393288, + 0.006531593389809132, + 0.00777071388438344, + 0.004087206441909075, + 0.009229353629052639, + 0.020124908536672592, + 0.02858259342610836, + 0.01295775081962347, + 0.01785251870751381, + 0.028121085837483406, + 0.02414417639374733, + 0.018125338479876518, + 0.0195015799254179, + 0.01840524934232235, + 0.018279915675520897, + 0.02619641087949276, + 0.022021131590008736, + 0.01787327229976654, + 0.017671719193458557, + 0.026834335178136826, + 0.030177289620041847, + 0.0073898774571716785, + 0.006778893526643515, + 0.008344755508005619, + 0.005507532507181168, + 0.005578548181802034, + 0.005666900891810656, + 0.007664433680474758, + 0.0064019630663096905, + 0.009959304705262184, + 0.007561258506029844, + 0.003427791642025113, + 0.006482736673206091, + 0.005580555181950331, + 0.00945279560983181, + 0.006508197635412216, + 0.012529796920716763, + 0.006921823136508465, + 0.010756080038845539, + 0.005878507625311613, + 0.0073822070844471455, + 0.010740336030721664, + 0.005690766032785177, + 0.008288626559078693, + 0.006143283098936081, + 0.0046917409636080265, + 0.010239848867058754, + 0.004250326193869114, + 0.004284583497792482, + 0.006437645759433508, + 0.007689564488828182, + 0.003346837591379881, + 0.004258766304701567, + 0.010455827228724957, + 0.010364102199673653, + 0.007494212593883276, + 0.00908560398966074, + 0.010452985763549805, + 0.008785272017121315, + 0.012124921195209026, + 0.015749288722872734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.004092176910489798 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 transformed Earth featuring a vortex of molten crimson and glacial blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are there vast ribbons of auroras in emerald green, purple, and rosy pink above the planet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the stars appear as luminous yellow swirls and pulsating clusters with thick paint textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style characterized by unbridled brushstrokes and impasto similar to Van Gogh's 'The Starry Night'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera rotate slowly to present the scene in a panoramic view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387005090713501 + ], + "opposite_logprob_tokens": [ + -1.137005090713501 + ], + "logprob_reward": -0.387005090713501, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387005090713501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08929876983165741, + 0.17245298624038696, + 0.04088698700070381, + 0.2407573163509369, + 0.07329732924699783, + 0.06064087897539139, + 0.10193140804767609, + 0.09359771013259888, + 0.09245855361223221, + 0.034678008407354355 + ], + "credit_map": [ + 0.011201188899576664, + 0.014078139327466488, + 0.006473393179476261, + 0.007900496013462543, + 0.0054466272704303265, + 0.013503932394087315, + 0.013336977921426296, + 0.017358025535941124, + 0.022645920515060425, + 0.017320973798632622, + 0.018793335184454918, + 0.024576829746365547, + 0.012013512663543224, + 0.025519853457808495, + 0.022618521004915237, + 0.028964050114154816, + 0.004650901071727276, + 0.005893400404602289, + 0.0040432061068713665, + 0.005142473615705967, + 0.004440778400748968, + 0.006540945265442133, + 0.004635469987988472, + 0.005539816804230213, + 0.018717888742685318, + 0.01999524235725403, + 0.03652124106884003, + 0.037037644535303116, + 0.020731471478939056, + 0.030696162953972816, + 0.02955845557153225, + 0.04749921336770058, + 0.013735475018620491, + 0.007766540627926588, + 0.009617063216865063, + 0.005892110988497734, + 0.007842464372515678, + 0.0072050453163683414, + 0.010527442209422588, + 0.01071118749678135, + 0.0053426050581038, + 0.0051063233986496925, + 0.004840149078518152, + 0.013593788258731365, + 0.00700786616653204, + 0.008264562115073204, + 0.006960589904338121, + 0.009525001049041748, + 0.018918819725513458, + 0.010115056298673153, + 0.019800661131739616, + 0.009424831718206406, + 0.006992587354034185, + 0.010894572362303734, + 0.008954797871410847, + 0.01683008298277855, + 0.008340937085449696, + 0.016541901975870132, + 0.008587312884628773, + 0.010459115728735924, + 0.018748806789517403, + 0.010896106250584126, + 0.009570913389325142, + 0.010452625341713428, + 0.011240900494158268, + 0.01304781436920166, + 0.011451092548668385, + 0.015896273776888847, + 0.009621677920222282, + 0.009139293804764748, + 0.013084911741316319, + 0.00897658709436655, + 0.004072711803019047, + 0.005485164932906628, + 0.002792784245684743, + 0.0029066638089716434, + 0.005109195597469807, + 0.004387780092656612, + 0.00408891960978508, + 0.0058347913436591625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 6 + ], + "selected_reward": -0.387005090713501 + } + ], + "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/0000152/group_3/index_tensor([283, 283])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_tensor([283, 283])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..46eb76b99b1903d534c7905cc7326ec04dea04dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_tensor([283, 283])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, shot from an extremely low angle, captures a vast, deep blue wave, as vast as a mountain, surging upward with unstoppable force. The core of this majestic body of water is a deep indigo, while its sharp, crumbling peak, shimmering with sunlight filtering through gaps in the dark clouds, appears a translucent turquoise. Its edges tumble with countless tiny, dazzling white foam and mist. The crest curves inward in an elegant yet fatal arc, forming a vast tunnel that slowly and inexorably closes. Its interior, smooth as glass, reflects the gloomy light of the sky, creating a claustrophobic and majestic visual space. The surrounding sea surges and tumbles, merging with the wave's massive base. The entire scene, illuminated by a dramatic, cold palette of light and shadow, is imbued with an atmosphere of tension and grandeur, perfectly showcasing the awe-inspiring raw power of nature.", + "seed": 20669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20669_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": -0.19822053890675306, + "vlm_logprob": -0.19822053890675306, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.19822053890675306, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6414128541946411 + ], + "opposite_logprob_tokens": [ + -0.7664128541946411 + ], + "logprob_reward": -0.6414128541946411, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6414128541946411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3450251817703247, + 0.06849467009305954, + 0.032395388931035995, + 0.062117915600538254, + 0.12357822060585022, + 0.05272461846470833, + 0.04875022545456886, + 0.05135134235024452, + 0.15538759529590607, + 0.060174793004989624 + ], + "credit_map": [ + 0.06937646865844727, + 0.019253650680184364, + 0.020901644602417946, + 0.024997036904096603, + 0.03744073584675789, + 0.12776988744735718, + 0.01650235243141651, + 0.028783457353711128, + 0.006382379215210676, + 0.006784782744944096, + 0.015779729932546616, + 0.008889754302799702, + 0.007819585502147675, + 0.009958912618458271, + 0.005230540409684181, + 0.007648997008800507, + 0.0037099665496498346, + 0.005098794121295214, + 0.003027792787179351, + 0.006857618689537048, + 0.003498688107356429, + 0.003989044576883316, + 0.002756609348580241, + 0.003456875216215849, + 0.00565137155354023, + 0.009843050502240658, + 0.00860607624053955, + 0.0066209761425852776, + 0.0074485600925982, + 0.010412763804197311, + 0.0057830810546875, + 0.007752040401101112, + 0.013066696934401989, + 0.013139564543962479, + 0.014395524747669697, + 0.01320191565901041, + 0.016757560893893242, + 0.02313530072569847, + 0.013612648472189903, + 0.016269013285636902, + 0.006363310851156712, + 0.009051942266523838, + 0.005947479046881199, + 0.007162000518292189, + 0.005789745599031448, + 0.005499659106135368, + 0.006201605778187513, + 0.006708880886435509, + 0.007246099412441254, + 0.006302328314632177, + 0.006046675145626068, + 0.005356453359127045, + 0.007637255825102329, + 0.004337242804467678, + 0.004102335311472416, + 0.007721843663603067, + 0.006450355984270573, + 0.006423804443329573, + 0.007241055369377136, + 0.0071197920478880405, + 0.0070460787974298, + 0.006989624351263046, + 0.005306261591613293, + 0.004774372559040785, + 0.021574705839157104, + 0.023542147129774094, + 0.020404983311891556, + 0.030220214277505875, + 0.015008136630058289, + 0.01709955558180809, + 0.013917521573603153, + 0.01362033560872078, + 0.010697840712964535, + 0.011920701712369919, + 0.006675877142697573, + 0.005909006576985121, + 0.009264676831662655, + 0.005039342679083347, + 0.0053263441659510136, + 0.005341007839888334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.6414128541946411 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813296318054199 + ], + "opposite_logprob_tokens": [ + -1.40632963180542 + ], + "logprob_reward": -0.2813296318054199, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813296318054199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11122219264507294, + 0.018909508362412453, + 0.020713787525892258, + 0.08335499465465546, + 0.02828298509120941, + 0.016820047050714493, + 0.043750982731580734, + 0.30790868401527405, + 0.3176687955856323, + 0.05136801302433014 + ], + "credit_map": [ + 0.022598199546337128, + 0.008109754882752895, + 0.006319722160696983, + 0.007091291714459658, + 0.011368720792233944, + 0.042523182928562164, + 0.004979288671165705, + 0.008232041262090206, + 0.0017136145615950227, + 0.0025853519327938557, + 0.0024971633683890104, + 0.003177190665155649, + 0.0020707729272544384, + 0.003066176315769553, + 0.0016126441769301891, + 0.002186595229431987, + 0.0032131539192050695, + 0.0023054080083966255, + 0.0016310359351336956, + 0.006290214601904154, + 0.0014318936737254262, + 0.0024112502578645945, + 0.002137460047379136, + 0.0012933718971908092, + 0.007692366372793913, + 0.009846643544733524, + 0.012060461565852165, + 0.008834699168801308, + 0.013078641146421432, + 0.010638760402798653, + 0.00960861798375845, + 0.011594807729125023, + 0.0031745925080031157, + 0.004670529626309872, + 0.00287887966260314, + 0.002984717721119523, + 0.0035661980509757996, + 0.004737823270261288, + 0.0031602629460394382, + 0.0031099829357117414, + 0.002021758584305644, + 0.0023335779551416636, + 0.0017646647756919265, + 0.0019042369676753879, + 0.0014312900602817535, + 0.0020370185375213623, + 0.002209445694461465, + 0.0031180563382804394, + 0.00676591694355011, + 0.006016668863594532, + 0.007969926111400127, + 0.00501467427238822, + 0.004295345861464739, + 0.004272916819900274, + 0.003253950271755457, + 0.006161590106785297, + 0.045355599373579025, + 0.036542635411024094, + 0.03197544813156128, + 0.038170021027326584, + 0.055126916617155075, + 0.04621313884854317, + 0.026309022679924965, + 0.028215907514095306, + 0.04537423700094223, + 0.047177158296108246, + 0.04339884966611862, + 0.057797644287347794, + 0.03130365535616875, + 0.034993842244148254, + 0.02696308307349682, + 0.03066035732626915, + 0.008173651061952114, + 0.0061336420476436615, + 0.008824079297482967, + 0.005097678396850824, + 0.006719698663800955, + 0.004750712774693966, + 0.004062286112457514, + 0.007606267463415861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.2813296318054199 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202469028532505 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202469028532505, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202469028532505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10838235914707184, + 0.014675440266728401, + 0.010387109592556953, + 0.03958127647638321, + 0.02744828164577484, + 0.02251068502664566, + 0.05670016631484032, + 0.11580215394496918, + 0.5152861475944519, + 0.08922640234231949 + ], + "credit_map": [ + 0.022290512919425964, + 0.005985755007714033, + 0.005572819616645575, + 0.007858562283217907, + 0.011433986946940422, + 0.04281659796833992, + 0.00472706975415349, + 0.007697049528360367, + 0.0011961928103119135, + 0.001330816769041121, + 0.0030769493896514177, + 0.0022411025129258633, + 0.0011370787397027016, + 0.0026529966853559017, + 0.0012815033551305532, + 0.0017588012851774693, + 0.0012296788627281785, + 0.0011335592716932297, + 0.00095037353457883, + 0.0019090810092166066, + 0.00113640446215868, + 0.001760632498189807, + 0.0013211274053901434, + 0.0009462528978474438, + 0.0035790971014648676, + 0.005328206345438957, + 0.0048148357309401035, + 0.0042296163737773895, + 0.005992500577121973, + 0.005362611263990402, + 0.004600158892571926, + 0.005674253683537245, + 0.0027823406271636486, + 0.0026669250801205635, + 0.004916249308735132, + 0.0031829786021262407, + 0.0037418354768306017, + 0.0049160015769302845, + 0.002378042321652174, + 0.0028639084193855524, + 0.002634635893628001, + 0.003295373171567917, + 0.002967902459204197, + 0.0021231716964393854, + 0.0031554836314171553, + 0.002731016371399164, + 0.002277444116771221, + 0.003325657220557332, + 0.008435089141130447, + 0.00918598659336567, + 0.005471154116094112, + 0.005998528096824884, + 0.006777784787118435, + 0.005168168339878321, + 0.006103818770498037, + 0.009559639729559422, + 0.017409905791282654, + 0.01673082448542118, + 0.015511704608798027, + 0.01532483659684658, + 0.015838999301195145, + 0.013402219861745834, + 0.008846797049045563, + 0.012736864387989044, + 0.07568749040365219, + 0.07210884988307953, + 0.06526799499988556, + 0.08642823249101639, + 0.05789237469434738, + 0.06470231711864471, + 0.04279465600848198, + 0.05040422081947327, + 0.013947250321507454, + 0.014522477053105831, + 0.008824272081255913, + 0.008640116080641747, + 0.013666429556906223, + 0.011325033381581306, + 0.00839413981884718, + 0.009906680323183537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.014202469028532505 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07788824290037155, + 0.020244238898158073, + 0.01090928167104721, + 0.06059327721595764, + 0.06262456625699997, + 0.034424226731061935, + 0.040253281593322754, + 0.10961896181106567, + 0.48788201808929443, + 0.09556189924478531 + ], + "credit_map": [ + 0.015581133775413036, + 0.004969202447682619, + 0.004366208333522081, + 0.004426640924066305, + 0.008369834162294865, + 0.028424039483070374, + 0.005096352659165859, + 0.006654832046478987, + 0.001941963448189199, + 0.002641557017341256, + 0.003671478945761919, + 0.003523891093209386, + 0.0017932149348780513, + 0.00282731419429183, + 0.0017735380679368973, + 0.002071282360702753, + 0.0009255764307454228, + 0.0011204549809917808, + 0.0011059418320655823, + 0.001733178156428039, + 0.0014105640584602952, + 0.002411104505881667, + 0.0010779480217024684, + 0.0011245140340179205, + 0.005117312539368868, + 0.00977355893701315, + 0.007980819791555405, + 0.005274580325931311, + 0.00859367847442627, + 0.009046015329658985, + 0.008408328518271446, + 0.0063989837653934956, + 0.00672896346077323, + 0.007501240354031324, + 0.007561610080301762, + 0.006743698380887508, + 0.008880066685378551, + 0.012795439921319485, + 0.005147852469235659, + 0.007265692576766014, + 0.0037911490071564913, + 0.005823926534503698, + 0.0037311401683837175, + 0.0035523197147995234, + 0.004020865075290203, + 0.004200704395771027, + 0.003949482925236225, + 0.005354639142751694, + 0.005482957232743502, + 0.005155735183507204, + 0.004586057737469673, + 0.004591097589582205, + 0.005085278768092394, + 0.0041859885677695274, + 0.004231870640069246, + 0.006934294477105141, + 0.01604796200990677, + 0.013759233988821507, + 0.017932211980223656, + 0.020830515772104263, + 0.010007725097239017, + 0.012771558947861195, + 0.008034564554691315, + 0.010235181078314781, + 0.07048561424016953, + 0.06920485943555832, + 0.06206398084759712, + 0.08188574761152267, + 0.05353527143597603, + 0.06044808775186539, + 0.04054926708340645, + 0.04970918223261833, + 0.016141334548592567, + 0.013725322671234608, + 0.012827370315790176, + 0.01022742222994566, + 0.01600971631705761, + 0.00716422637924552, + 0.007599142845720053, + 0.011867356486618519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.016067897900938988 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808984160423279 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.03808984160423279, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808984160423279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09209699928760529, + 0.015758046880364418, + 0.009910681284964085, + 0.02991325594484806, + 0.09267354756593704, + 0.00687599740922451, + 0.04683449864387512, + 0.05195668339729309, + 0.6075612902641296, + 0.04641900956630707 + ], + "credit_map": [ + 0.016866276040673256, + 0.007576183881610632, + 0.004142523743212223, + 0.009364630095660686, + 0.009528213180601597, + 0.033432040363550186, + 0.004261499270796776, + 0.006925634108483791, + 0.0015904005849733949, + 0.0015734049957245588, + 0.0035322452895343304, + 0.0015417831018567085, + 0.0017248226795345545, + 0.002840729895979166, + 0.0011555354576557875, + 0.0017991232452914119, + 0.0014660146553069353, + 0.0010897715110331774, + 0.0009826583554968238, + 0.0026806530077010393, + 0.000682343787048012, + 0.0010112251620739698, + 0.0011106292949989438, + 0.0008873859769664705, + 0.002859557280316949, + 0.005302416626363993, + 0.0037945061922073364, + 0.0031266342848539352, + 0.004011186771094799, + 0.0035384707152843475, + 0.003682349808514118, + 0.0035981356631964445, + 0.009708122350275517, + 0.013171267695724964, + 0.009632867760956287, + 0.009895270690321922, + 0.012044901959598064, + 0.01839577965438366, + 0.009190451353788376, + 0.01063488982617855, + 0.0008140529971569777, + 0.0008292596903629601, + 0.0009426989709027112, + 0.0006450001965276897, + 0.0008939085528254509, + 0.0010390906827524304, + 0.0006341159460134804, + 0.0010778701398521662, + 0.007246522698551416, + 0.005508975591510534, + 0.006154214963316917, + 0.006748992949724197, + 0.004562628921121359, + 0.005878929048776627, + 0.004065813031047583, + 0.006668422371149063, + 0.007264626212418079, + 0.006608530413359404, + 0.005828554276376963, + 0.004368130583316088, + 0.009516178630292416, + 0.009197195060551167, + 0.004611259792000055, + 0.004562209360301495, + 0.08434959501028061, + 0.088304802775383, + 0.07969386130571365, + 0.1117677241563797, + 0.06174735724925995, + 0.07349267601966858, + 0.05224804952740669, + 0.05595721676945686, + 0.007740104105323553, + 0.005518806166946888, + 0.00812515802681446, + 0.004550490528345108, + 0.004778292961418629, + 0.0036169278901070356, + 0.0034762504510581493, + 0.008612980134785175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.03808984160423279 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from an extremely low angle, captures a vast, deep blue wave, as vast as a mountain, surging upward with unstoppable force. The core of this majestic body of water is a deep indigo, while its sharp, crumbling peak, shimmering with sunlight filtering through gaps in the dark clouds, appears a translucent turquoise. Its edges tumble with countless tiny, dazzling white foam and mist. The crest curves inward in an elegant yet fatal arc, forming a vast tunnel that slowly and inexorably closes. Its interior, smooth as glass, reflects the gloomy light of the sky, creating a claustrophobic and majestic visual space. The surrounding sea surges and tumbles, merging with the wave's massive base. The entire scene, illuminated by a dramatic, cold palette of light and shadow, is imbued with an atmosphere of tension and grandeur, perfectly showcasing the awe-inspiring raw power of nature.", + "seed": 20670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20670_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": -0.1020650795660913, + "vlm_logprob": -0.1020650795660913, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.1020650795660913, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17354004085063934 + ], + "opposite_logprob_tokens": [ + -1.9235399961471558 + ], + "logprob_reward": -0.17354004085063934, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17354004085063934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17197664082050323, + 0.10007209330797195, + 0.08708019554615021, + 0.05043727532029152, + 0.07225047796964645, + 0.0368940532207489, + 0.15413376688957214, + 0.04484134912490845, + 0.034123603254556656, + 0.2481905072927475 + ], + "credit_map": [ + 0.040775980800390244, + 0.017126990482211113, + 0.009467204101383686, + 0.006602453533560038, + 0.05702542886137962, + 0.01969134993851185, + 0.011348825879395008, + 0.009938386268913746, + 0.01477888971567154, + 0.012633915059268475, + 0.015808826312422752, + 0.011351325549185276, + 0.008083702996373177, + 0.011391405016183853, + 0.010992884635925293, + 0.01503114402294159, + 0.008338479325175285, + 0.008281500078737736, + 0.007868248037993908, + 0.021134521812200546, + 0.008116812445223331, + 0.009072612971067429, + 0.008253258652985096, + 0.016014765948057175, + 0.005229005124419928, + 0.005553066730499268, + 0.004525178577750921, + 0.010905321687459946, + 0.006231286562979221, + 0.006379140540957451, + 0.004142908379435539, + 0.007471366785466671, + 0.007416039705276489, + 0.007521489635109901, + 0.008100831881165504, + 0.009141613729298115, + 0.01419857144355774, + 0.011499167419970036, + 0.006996438838541508, + 0.0073763239197432995, + 0.004779977723956108, + 0.0042375223711133, + 0.0057811252772808075, + 0.0036171607207506895, + 0.004949284251779318, + 0.004862827714532614, + 0.004021803382784128, + 0.004644350614398718, + 0.019602831453084946, + 0.024706261232495308, + 0.01942804642021656, + 0.016855167225003242, + 0.02238486520946026, + 0.02300814352929592, + 0.015841558575630188, + 0.012306892313063145, + 0.0034173745661973953, + 0.004363017622381449, + 0.005004218779504299, + 0.003045931924134493, + 0.007089377846568823, + 0.009571419097483158, + 0.009277825243771076, + 0.0030721835792064667, + 0.0023817168548703194, + 0.003771305549889803, + 0.004686933476477861, + 0.0054999059066176414, + 0.005277404561638832, + 0.004384145140647888, + 0.0046245018020272255, + 0.0034976876340806484, + 0.016103623434901237, + 0.02147919312119484, + 0.07065702974796295, + 0.01740454137325287, + 0.06235231086611748, + 0.026416925713419914, + 0.011249598115682602, + 0.022527307271957397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.17354004085063934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134293854236603 + ], + "opposite_logprob_tokens": [ + -1.313429355621338 + ], + "logprob_reward": -0.3134293854236603, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134293854236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10032187402248383, + 0.07247849553823471, + 0.20696866512298584, + 0.13039283454418182, + 0.22754676640033722, + 0.0936499610543251, + 0.05628359317779541, + 0.01718193292617798, + 0.026568831875920296, + 0.06860701739788055 + ], + "credit_map": [ + 0.01592114008963108, + 0.016416803002357483, + 0.009413963183760643, + 0.007768625859171152, + 0.023708218708634377, + 0.011282785795629025, + 0.00866421964019537, + 0.007146128453314304, + 0.007864801213145256, + 0.011649055406451225, + 0.008650456555187702, + 0.006653875578194857, + 0.00938450824469328, + 0.009002145379781723, + 0.00944868940860033, + 0.009824970737099648, + 0.025804772973060608, + 0.017325563356280327, + 0.013758393935859203, + 0.062143724411726, + 0.017094271257519722, + 0.016808239743113518, + 0.01840650476515293, + 0.035627204924821854, + 0.02036799117922783, + 0.007696392480283976, + 0.007204336114227772, + 0.04961370676755905, + 0.0064277974888682365, + 0.012516014277935028, + 0.011188754811882973, + 0.015377847477793694, + 0.03457252308726311, + 0.02924576960504055, + 0.042034029960632324, + 0.02556742914021015, + 0.027865229174494743, + 0.032093923538923264, + 0.012745085172355175, + 0.023422785103321075, + 0.01061259862035513, + 0.009249914437532425, + 0.016358867287635803, + 0.01092539168894291, + 0.01687813177704811, + 0.013348255306482315, + 0.0077132186852395535, + 0.00856359489262104, + 0.004846976138651371, + 0.007757681421935558, + 0.006277988664805889, + 0.00591444643214345, + 0.012698519043624401, + 0.009604710154235363, + 0.005483777262270451, + 0.0036994966212660074, + 0.0015127809019759297, + 0.0021384269930422306, + 0.0023986422456800938, + 0.0014149212511256337, + 0.0035602981224656105, + 0.002281768014654517, + 0.0023106327280402184, + 0.001564463833346963, + 0.002268620766699314, + 0.0035883220843970776, + 0.003166855312883854, + 0.004336310084909201, + 0.0033027727622538805, + 0.002943604253232479, + 0.0038414725568145514, + 0.003120878478512168, + 0.0047483788803219795, + 0.007156079169362783, + 0.016646062955260277, + 0.004810903687030077, + 0.016233175992965698, + 0.007525062188506126, + 0.0041283974424004555, + 0.007358970586210489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.3134293854236603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632380351424217 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632380351424217, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632380351424217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16489049792289734, + 0.09940258413553238, + 0.10698030889034271, + 0.10652606189250946, + 0.08963219821453094, + 0.07062574476003647, + 0.0889059528708458, + 0.057409483939409256, + 0.05182892456650734, + 0.16379813849925995 + ], + "credit_map": [ + 0.03188857063651085, + 0.019868019968271255, + 0.01108411606401205, + 0.011314812116324902, + 0.053145747631788254, + 0.019053202122449875, + 0.009673508815467358, + 0.008862525224685669, + 0.01714225858449936, + 0.009127660654485226, + 0.011335314251482487, + 0.0059959557838737965, + 0.00787951610982418, + 0.010357372462749481, + 0.012674368917942047, + 0.024890132248401642, + 0.012454903684556484, + 0.015613095834851265, + 0.012275890447199345, + 0.018146937713027, + 0.011132585816085339, + 0.010092519223690033, + 0.009474662132561207, + 0.01778971031308174, + 0.019704259932041168, + 0.0109330452978611, + 0.009584062732756138, + 0.024219917133450508, + 0.008339290507137775, + 0.010626441799104214, + 0.008696764707565308, + 0.014422285370528698, + 0.010233554057776928, + 0.01219408493489027, + 0.013365487568080425, + 0.009859997779130936, + 0.012834485620260239, + 0.012803434394299984, + 0.0074525210075080395, + 0.010888637974858284, + 0.007035795133560896, + 0.007382409181445837, + 0.00983283668756485, + 0.006702990736812353, + 0.013796279206871986, + 0.00999783631414175, + 0.009552190080285072, + 0.006325406953692436, + 0.006033772137016058, + 0.01153216976672411, + 0.008546659722924232, + 0.013662858866155148, + 0.019163968041539192, + 0.011475257575511932, + 0.012275613844394684, + 0.006215657107532024, + 0.004329899325966835, + 0.005575951188802719, + 0.006419294513761997, + 0.005231156479567289, + 0.008662228472530842, + 0.010957550257444382, + 0.010948127135634422, + 0.005285276100039482, + 0.004871038720011711, + 0.005553874187171459, + 0.006477465387433767, + 0.011218649335205555, + 0.005857309326529503, + 0.006102020852267742, + 0.00673820311203599, + 0.005010363645851612, + 0.017365504056215286, + 0.015953106805682182, + 0.03243165463209152, + 0.011602726764976978, + 0.038531649857759476, + 0.016234958544373512, + 0.01184128038585186, + 0.01983725093305111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.007632380351424217 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08728311210870743, + 0.1554529219865799, + 0.30699020624160767, + 0.046650297939777374, + 0.05724329873919487, + 0.06383629888296127, + 0.028406739234924316, + 0.01921447366476059, + 0.04287094250321388, + 0.19205166399478912 + ], + "credit_map": [ + 0.012097676284611225, + 0.01616968959569931, + 0.00821312703192234, + 0.009111976251006126, + 0.015723448246717453, + 0.010588468983769417, + 0.007911007851362228, + 0.007467718794941902, + 0.024475276470184326, + 0.020835157483816147, + 0.022326229140162468, + 0.02214663289487362, + 0.012507624924182892, + 0.014044967480003834, + 0.016456076875329018, + 0.02266095019876957, + 0.04477904364466667, + 0.04729157313704491, + 0.04518568143248558, + 0.06784700602293015, + 0.025471806526184082, + 0.024103976786136627, + 0.024689288809895515, + 0.02762187272310257, + 0.006331202108412981, + 0.004791074898093939, + 0.006549629848450422, + 0.01105100754648447, + 0.0033976007252931595, + 0.004934750031679869, + 0.004120314959436655, + 0.00547472108155489, + 0.007572353817522526, + 0.005142138339579105, + 0.012524174526333809, + 0.006617673672735691, + 0.00953181553632021, + 0.006890179589390755, + 0.0039766342379152775, + 0.004988333210349083, + 0.0067672873847186565, + 0.006269640754908323, + 0.008546687662601471, + 0.005942226853221655, + 0.015087469480931759, + 0.009848321788012981, + 0.006348372437059879, + 0.0050262948498129845, + 0.0030977013520896435, + 0.0022277606185525656, + 0.003967176657170057, + 0.001699554966762662, + 0.007459981832653284, + 0.004111727233976126, + 0.003776394994929433, + 0.002066443208605051, + 0.0016744607128202915, + 0.002443829318508506, + 0.002725196536630392, + 0.0018293381435796618, + 0.0035586070735007524, + 0.002946909051388502, + 0.0026063276454806328, + 0.001429806463420391, + 0.003377278335392475, + 0.004641619976609945, + 0.006835694890469313, + 0.008070847950875759, + 0.007768786046653986, + 0.00518803671002388, + 0.003790071466937661, + 0.003198611317202449, + 0.012539955787360668, + 0.012884251773357391, + 0.05712763965129852, + 0.01149033848196268, + 0.05818239599466324, + 0.016745230183005333, + 0.006677587982267141, + 0.01640426367521286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.009779990650713444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005943600554019213 + ], + "opposite_logprob_tokens": [ + -5.130943775177002 + ], + "logprob_reward": -0.005943600554019213, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005943600554019213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08271601051092148, + 0.09321850538253784, + 0.09832534939050674, + 0.22579382359981537, + 0.2940952479839325, + 0.0839814841747284, + 0.041950419545173645, + 0.021889738738536835, + 0.025108108296990395, + 0.03292129188776016 + ], + "credit_map": [ + 0.009845737367868423, + 0.01192847453057766, + 0.009003141894936562, + 0.009641507640480995, + 0.01239060889929533, + 0.011645693331956863, + 0.00872871559113264, + 0.009532135911285877, + 0.013738065958023071, + 0.013831275515258312, + 0.01534031331539154, + 0.010051331482827663, + 0.008036036044359207, + 0.010461383499205112, + 0.009337817318737507, + 0.012422275729477406, + 0.011772100813686848, + 0.010178684256970882, + 0.014404206536710262, + 0.01432992983609438, + 0.011912234127521515, + 0.012728977017104626, + 0.008847878314554691, + 0.014151339419186115, + 0.04275738075375557, + 0.013552519492805004, + 0.011911220848560333, + 0.08795220404863358, + 0.005843765567988157, + 0.019028186798095703, + 0.017669932916760445, + 0.02707861177623272, + 0.04170025512576103, + 0.03747228533029556, + 0.05606349557638168, + 0.032120585441589355, + 0.03215447813272476, + 0.04258395731449127, + 0.017249900847673416, + 0.034750279039144516, + 0.011050696484744549, + 0.010536015965044498, + 0.01653069630265236, + 0.009068374522030354, + 0.011226054280996323, + 0.010874688625335693, + 0.0074190422892570496, + 0.007275919895619154, + 0.0035486158449202776, + 0.004999023862183094, + 0.004706213716417551, + 0.004645606502890587, + 0.010467278771102428, + 0.0072511243633925915, + 0.0037404540926218033, + 0.0025921030901372433, + 0.0015089916996657848, + 0.002321796491742134, + 0.0023809478152543306, + 0.0021896595135331154, + 0.005043382290750742, + 0.0036845491267740726, + 0.003373640589416027, + 0.0013867689995095134, + 0.002033565891906619, + 0.0028995422180742025, + 0.004120588302612305, + 0.0034910549875348806, + 0.004292930942028761, + 0.003175679361447692, + 0.002860492095351219, + 0.0022342561278492212, + 0.0031890543177723885, + 0.004125287756323814, + 0.004457634873688221, + 0.003653298830613494, + 0.0049670301377773285, + 0.005279130302369595, + 0.0029687972273677588, + 0.00428105890750885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.005943600554019213 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from an extremely low angle, captures a vast, deep blue wave, as vast as a mountain, surging upward with unstoppable force. The core of this majestic body of water is a deep indigo, while its sharp, crumbling peak, shimmering with sunlight filtering through gaps in the dark clouds, appears a translucent turquoise. Its edges tumble with countless tiny, dazzling white foam and mist. The crest curves inward in an elegant yet fatal arc, forming a vast tunnel that slowly and inexorably closes. Its interior, smooth as glass, reflects the gloomy light of the sky, creating a claustrophobic and majestic visual space. The surrounding sea surges and tumbles, merging with the wave's massive base. The entire scene, illuminated by a dramatic, cold palette of light and shadow, is imbued with an atmosphere of tension and grandeur, perfectly showcasing the awe-inspiring raw power of nature.", + "seed": 20671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20671_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": -0.1038038894534111, + "vlm_logprob": -0.1038038894534111, + "vlm_yes_no_margin": 3.374999988079071, + "vlm_token_credit": -0.1038038894534111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14220665395259857 + ], + "opposite_logprob_tokens": [ + -2.1422066688537598 + ], + "logprob_reward": -0.14220665395259857, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.14220665395259857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10022066533565521, + 0.026439273729920387, + 0.09709787368774414, + 0.10234376043081284, + 0.05317477509379387, + 0.08659534901380539, + 0.04717347398400307, + 0.07054166495800018, + 0.2854556143283844, + 0.1309574842453003 + ], + "credit_map": [ + 0.013346481136977673, + 0.012346120551228523, + 0.010548212565481663, + 0.015273453667759895, + 0.005055974703282118, + 0.009817230515182018, + 0.015310967341065407, + 0.018522225320339203, + 0.0033088966738432646, + 0.005300263874232769, + 0.0019529114942997694, + 0.002755445893853903, + 0.00241414038464427, + 0.004036560654640198, + 0.0030451349448412657, + 0.0036259195767343044, + 0.008171835914254189, + 0.00824158638715744, + 0.01394034456461668, + 0.0072780088521540165, + 0.022212181240320206, + 0.020342525094747543, + 0.008203970268368721, + 0.008707418106496334, + 0.009714907966554165, + 0.009790068492293358, + 0.005390201695263386, + 0.009223328903317451, + 0.026653334498405457, + 0.0089559406042099, + 0.008088074624538422, + 0.02452789805829525, + 0.005212561227381229, + 0.007915015332400799, + 0.006269564386457205, + 0.0062323459424078465, + 0.015059362165629864, + 0.005536600947380066, + 0.002558899810537696, + 0.0043904222548007965, + 0.006073246710002422, + 0.008105254732072353, + 0.004791386891156435, + 0.027642419561743736, + 0.006615881342440844, + 0.021909380331635475, + 0.006152081303298473, + 0.005305703729391098, + 0.0071337283588945866, + 0.0052890232764184475, + 0.006540514063090086, + 0.0053727817721664906, + 0.007542431354522705, + 0.007234974764287472, + 0.00351041741669178, + 0.004549599252641201, + 0.011190622113645077, + 0.007992617785930634, + 0.008733542636036873, + 0.011463645845651627, + 0.010399851016700268, + 0.011229446157813072, + 0.0045136259868741035, + 0.005018315277993679, + 0.03579598292708397, + 0.02460484579205513, + 0.033396873623132706, + 0.04713502898812294, + 0.034578822553157806, + 0.06324975192546844, + 0.01819634437561035, + 0.028497964143753052, + 0.010106386616826057, + 0.010189480148255825, + 0.016478121280670166, + 0.015618780627846718, + 0.01995333842933178, + 0.035615187138319016, + 0.011645282618701458, + 0.011350912973284721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.14220665395259857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34861618280410767 + ], + "opposite_logprob_tokens": [ + -1.223616123199463 + ], + "logprob_reward": -0.34861618280410767, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34861618280410767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1087786927819252, + 0.02651042677462101, + 0.08094083517789841, + 0.10169621556997299, + 0.17419876158237457, + 0.18040892481803894, + 0.08267568051815033, + 0.04698900133371353, + 0.05975055322051048, + 0.13805094361305237 + ], + "credit_map": [ + 0.012121320702135563, + 0.009756145067512989, + 0.01459436770528555, + 0.016555167734622955, + 0.007241683080792427, + 0.013271654024720192, + 0.02051854319870472, + 0.01471981406211853, + 0.0034332280047237873, + 0.002839690074324608, + 0.002957400167360902, + 0.003504370339214802, + 0.0028749648481607437, + 0.003324138233438134, + 0.003891743253916502, + 0.0036848937161266804, + 0.009816211648285389, + 0.013187668286263943, + 0.007486904971301556, + 0.006525994278490543, + 0.011768022552132607, + 0.00808336865156889, + 0.010551809333264828, + 0.013520857319235802, + 0.011173350736498833, + 0.011257304809987545, + 0.01646193116903305, + 0.009740020148456097, + 0.016315916553139687, + 0.01767457090318203, + 0.00878881011158228, + 0.01028431486338377, + 0.01859886385500431, + 0.021387891843914986, + 0.013714400120079517, + 0.028178557753562927, + 0.03223544731736183, + 0.016646932810544968, + 0.009228788316249847, + 0.034207865595817566, + 0.009904948063194752, + 0.011612191796302795, + 0.010380680672824383, + 0.0599651075899601, + 0.010668895207345486, + 0.05688147991895676, + 0.011202570982277393, + 0.009793052449822426, + 0.008266521617770195, + 0.010689752176404, + 0.009715949185192585, + 0.01268357876688242, + 0.012830797582864761, + 0.012321638874709606, + 0.00893241073936224, + 0.007235031109303236, + 0.006629052106291056, + 0.006292629987001419, + 0.0058031403459608555, + 0.005465224385261536, + 0.007743051275610924, + 0.0063996994867920876, + 0.00317066116258502, + 0.005485537461936474, + 0.010715147480368614, + 0.005196205340325832, + 0.008371619507670403, + 0.006304266396909952, + 0.00828526820987463, + 0.007123645395040512, + 0.005006845109164715, + 0.008747554384171963, + 0.02204279415309429, + 0.010249846614897251, + 0.010629797354340553, + 0.012047966942191124, + 0.01723199523985386, + 0.02204074151813984, + 0.013117252849042416, + 0.030690543353557587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.34861618280410767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008641940541565418 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.008641940541565418, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008641940541565418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13551577925682068, + 0.07769644260406494, + 0.08495243638753891, + 0.051996417343616486, + 0.3475368320941925, + 0.07585768401622772, + 0.03230352699756622, + 0.03810751438140869, + 0.05826907604932785, + 0.09776429086923599 + ], + "credit_map": [ + 0.020127050578594208, + 0.013777690008282661, + 0.014408349059522152, + 0.017757494002580643, + 0.01219924446195364, + 0.018323564901947975, + 0.01893959566950798, + 0.01998279057443142, + 0.010868394747376442, + 0.011840521357953548, + 0.006259992253035307, + 0.007763043977320194, + 0.007229079492390156, + 0.007823391817510128, + 0.010259446687996387, + 0.015652574598789215, + 0.010795652866363525, + 0.012486224062740803, + 0.012584485113620758, + 0.00680018262937665, + 0.012928685173392296, + 0.011811157688498497, + 0.009043173864483833, + 0.008502882905304432, + 0.00687829963862896, + 0.007985344156622887, + 0.005891614127904177, + 0.004341618157923222, + 0.00931029673665762, + 0.005032653454691172, + 0.005142257548868656, + 0.0074143363162875175, + 0.027081990614533424, + 0.06221794709563255, + 0.022346559911966324, + 0.04383230581879616, + 0.06805052608251572, + 0.024756671860814095, + 0.01653916947543621, + 0.08271165192127228, + 0.012224584817886353, + 0.012247082777321339, + 0.007065854966640472, + 0.009504993446171284, + 0.009961309842765331, + 0.010857938788831234, + 0.005719507113099098, + 0.008276411332190037, + 0.0038086322601884604, + 0.0031319467816501856, + 0.004066583700478077, + 0.004575768951326609, + 0.004896348807960749, + 0.004822021350264549, + 0.003476586891338229, + 0.0035256370902061462, + 0.004989044740796089, + 0.0034650990273803473, + 0.007918882183730602, + 0.003884949954226613, + 0.0059390682727098465, + 0.005132447928190231, + 0.002791693899780512, + 0.003986329305917025, + 0.006069202907383442, + 0.004600187297910452, + 0.006220206152647734, + 0.009793221019208431, + 0.0073901088908314705, + 0.015212181024253368, + 0.0035446372348815203, + 0.0054393345490098, + 0.010235985741019249, + 0.009195404127240181, + 0.01388632319867611, + 0.01137642003595829, + 0.016778668388724327, + 0.018756220117211342, + 0.007603496313095093, + 0.009931770153343678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.008641940541565418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779401123523712 + ], + "opposite_logprob_tokens": [ + -4.634779453277588 + ], + "logprob_reward": -0.009779401123523712, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779401123523712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07722905278205872, + 0.0640261322259903, + 0.1845751255750656, + 0.05763494595885277, + 0.2593582570552826, + 0.094429612159729, + 0.04627148434519768, + 0.06122569739818573, + 0.09574916213750839, + 0.0595005638897419 + ], + "credit_map": [ + 0.012587880715727806, + 0.006910268682986498, + 0.009123747237026691, + 0.011808225885033607, + 0.0040743350982666016, + 0.00476240087300539, + 0.00727118318900466, + 0.02069101296365261, + 0.009964048862457275, + 0.004818175919353962, + 0.005611343774944544, + 0.0052401358261704445, + 0.0035084402188658714, + 0.004318926949054003, + 0.012770392000675201, + 0.017794670537114143, + 0.015844840556383133, + 0.02154107391834259, + 0.02790124900639057, + 0.013439740985631943, + 0.031679049134254456, + 0.0342535674571991, + 0.022257652133703232, + 0.01765795610845089, + 0.00572575256228447, + 0.0064103263430297375, + 0.007774730678647757, + 0.003958862740546465, + 0.012014041654765606, + 0.009304018691182137, + 0.0057979766279459, + 0.00664923433214426, + 0.017871776595711708, + 0.04776628315448761, + 0.017565960064530373, + 0.014055009000003338, + 0.07650155574083328, + 0.0205320306122303, + 0.014093044213950634, + 0.0509725920855999, + 0.007127601187676191, + 0.008007504045963287, + 0.006033828016370535, + 0.0240750964730978, + 0.007136301137506962, + 0.030538728460669518, + 0.004776837769895792, + 0.006733715068548918, + 0.003853572066873312, + 0.00536576472222805, + 0.0066833533346652985, + 0.0054753683507442474, + 0.008548199199140072, + 0.007649427279829979, + 0.0041015129536390305, + 0.004594289232045412, + 0.0066407956182956696, + 0.006190117448568344, + 0.009171497076749802, + 0.007567232940346003, + 0.009636729024350643, + 0.009327110834419727, + 0.0062080323696136475, + 0.006484178826212883, + 0.0071115754544734955, + 0.0039005086291581392, + 0.005910565610975027, + 0.01633446477353573, + 0.016893181949853897, + 0.03326413407921791, + 0.005331374704837799, + 0.007003362290561199, + 0.009393775835633278, + 0.005778140388429165, + 0.00911148451268673, + 0.007328550796955824, + 0.007466966286301613, + 0.010424734093248844, + 0.004731932654976845, + 0.00526498444378376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.009779401123523712 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10179172456264496, + 0.02848738804459572, + 0.07116580009460449, + 0.046323779970407486, + 0.229237899184227, + 0.43084532022476196, + 0.009757515043020248, + 0.013803165405988693, + 0.04277695342898369, + 0.025810474529862404 + ], + "credit_map": [ + 0.017141465097665787, + 0.007437207270413637, + 0.008548655547201633, + 0.014626326970756054, + 0.00646631233394146, + 0.00863419845700264, + 0.015880603343248367, + 0.02305694855749607, + 0.004309463314712048, + 0.0048422194086015224, + 0.0020193473901599646, + 0.0029616050887852907, + 0.003364223288372159, + 0.004201309755444527, + 0.0038922864478081465, + 0.002896930556744337, + 0.005761394742876291, + 0.007031144108623266, + 0.010607524774968624, + 0.005515639204531908, + 0.013711730018258095, + 0.011624816805124283, + 0.008928526192903519, + 0.007985024712979794, + 0.004376730881631374, + 0.00409532617777586, + 0.0033858169335871935, + 0.004549872130155563, + 0.01112756785005331, + 0.0037655942142009735, + 0.0035900534130632877, + 0.011432819068431854, + 0.02294945903122425, + 0.020477566868066788, + 0.015648245811462402, + 0.047152068465948105, + 0.023051666095852852, + 0.0165119431912899, + 0.011791527271270752, + 0.07165541499853134, + 0.022980745881795883, + 0.029062354937195778, + 0.021786540746688843, + 0.13707095384597778, + 0.02751908265054226, + 0.14740177989006042, + 0.023581523448228836, + 0.02144232578575611, + 0.0009412821964360774, + 0.0010850140824913979, + 0.0011106476886197925, + 0.0018472156953066587, + 0.0014855334302410483, + 0.001528236549347639, + 0.0008920961408875883, + 0.0008674882701598108, + 0.0016236732481047511, + 0.0015554907731711864, + 0.0017268446972593665, + 0.001910052727907896, + 0.0021554387640208006, + 0.0019062710925936699, + 0.001153632765635848, + 0.0017717601731419563, + 0.006348558235913515, + 0.004159246571362019, + 0.00430353544652462, + 0.006193767301738262, + 0.005744781345129013, + 0.008851886726915836, + 0.002746827434748411, + 0.004428350366652012, + 0.0030213233549147844, + 0.0020328988321125507, + 0.0032435297034680843, + 0.0025837530847638845, + 0.0030440837144851685, + 0.004606630653142929, + 0.0018203773070126772, + 0.005457875784486532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.009775268845260143 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, shot from an extremely low angle, captures a vast, deep blue wave, as vast as a mountain, surging upward with unstoppable force. The core of this majestic body of water is a deep indigo, while its sharp, crumbling peak, shimmering with sunlight filtering through gaps in the dark clouds, appears a translucent turquoise. Its edges tumble with countless tiny, dazzling white foam and mist. The crest curves inward in an elegant yet fatal arc, forming a vast tunnel that slowly and inexorably closes. Its interior, smooth as glass, reflects the gloomy light of the sky, creating a claustrophobic and majestic visual space. The surrounding sea surges and tumbles, merging with the wave's massive base. The entire scene, illuminated by a dramatic, cold palette of light and shadow, is imbued with an atmosphere of tension and grandeur, perfectly showcasing the awe-inspiring raw power of nature.", + "seed": 20672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_3/index_283_seed_20672_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": -0.11245971946045756, + "vlm_logprob": -0.11245971946045756, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.11245971946045756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05269403010606766 + ], + "opposite_logprob_tokens": [ + -3.052694082260132 + ], + "logprob_reward": -0.05269403010606766, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05269403010606766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36157578229904175, + 0.05973406881093979, + 0.08030163496732712, + 0.05341723933815956, + 0.07296467572450638, + 0.0626305341720581, + 0.024444984272122383, + 0.09261097759008408, + 0.14045019447803497, + 0.051869895309209824 + ], + "credit_map": [ + 0.034768857061862946, + 0.04422663152217865, + 0.03309249505400658, + 0.057142697274684906, + 0.03508162871003151, + 0.03976038470864296, + 0.049251653254032135, + 0.06825143098831177, + 0.00650013005360961, + 0.007462012115865946, + 0.005659328307956457, + 0.008545381017029285, + 0.007539407350122929, + 0.009314865805208683, + 0.006562388502061367, + 0.008150553330779076, + 0.009185385890305042, + 0.01126745343208313, + 0.008391344919800758, + 0.0076600331813097, + 0.008489890955388546, + 0.012002257630228996, + 0.01377179753035307, + 0.0095334742218256, + 0.005484228488057852, + 0.004284687805920839, + 0.00742866238579154, + 0.004967843648046255, + 0.004183883313089609, + 0.009173374623060226, + 0.01171358022838831, + 0.006180980242788792, + 0.010532352142035961, + 0.005863597616553307, + 0.012014076113700867, + 0.008827272802591324, + 0.010745622217655182, + 0.012022769078612328, + 0.006666616536676884, + 0.006292367819696665, + 0.007750013843178749, + 0.006350715644657612, + 0.006361081730574369, + 0.006157363764941692, + 0.015689745545387268, + 0.008452137000858784, + 0.007282332982867956, + 0.00458713760599494, + 0.003520905738696456, + 0.0026284423656761646, + 0.0037747761234641075, + 0.0025733248330652714, + 0.0041701169684529305, + 0.0036638816818594933, + 0.002060313243418932, + 0.002053223317489028, + 0.008169261738657951, + 0.006221727933734655, + 0.006125657819211483, + 0.007443451788276434, + 0.03286116570234299, + 0.021368524059653282, + 0.0055088503286242485, + 0.004912344738841057, + 0.011946839280426502, + 0.019511759281158447, + 0.017955686897039413, + 0.018694307655096054, + 0.020948702469468117, + 0.02078988589346409, + 0.015555243007838726, + 0.015047774650156498, + 0.004998620133846998, + 0.006679420825093985, + 0.006402905099093914, + 0.006845156662166119, + 0.006160230375826359, + 0.007046328391879797, + 0.007504056207835674, + 0.006233182270079851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.05269403010606766 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47426390647888184 + ], + "opposite_logprob_tokens": [ + -0.9742639064788818 + ], + "logprob_reward": -0.47426390647888184, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47426390647888184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1364542692899704, + 0.05121942237019539, + 0.05229244381189346, + 0.07830977439880371, + 0.0861777737736702, + 0.07168333977460861, + 0.024656571447849274, + 0.3460652530193329, + 0.07911264896392822, + 0.07402845472097397 + ], + "credit_map": [ + 0.015352608636021614, + 0.0104941725730896, + 0.009358162991702557, + 0.017264552414417267, + 0.009999777190387249, + 0.021980803459882736, + 0.02338062971830368, + 0.0286235548555851, + 0.005137283820658922, + 0.004894046578556299, + 0.004500631708651781, + 0.006371321622282267, + 0.007438861299306154, + 0.009600190445780754, + 0.006641468033194542, + 0.006635623052716255, + 0.007848646491765976, + 0.007731845136731863, + 0.005966485477983952, + 0.005282090976834297, + 0.005065553355962038, + 0.008300255052745342, + 0.007394604384899139, + 0.004702962469309568, + 0.008466455154120922, + 0.007042931858450174, + 0.012846731580793858, + 0.007492844946682453, + 0.008508078753948212, + 0.012100725434720516, + 0.011478307656943798, + 0.010373697616159916, + 0.010753510519862175, + 0.00633371202275157, + 0.013572490774095058, + 0.008209954015910625, + 0.012667440809309483, + 0.017166707664728165, + 0.008976326324045658, + 0.008497628383338451, + 0.011833791621029377, + 0.00774871651083231, + 0.009676016867160797, + 0.009014591574668884, + 0.012302099727094173, + 0.008885192684829235, + 0.006905687972903252, + 0.005317248869687319, + 0.0021899405401200056, + 0.002723385812714696, + 0.0021806787699460983, + 0.0021476754918694496, + 0.006863870657980442, + 0.0037905131466686726, + 0.0027298573404550552, + 0.0020306494552642107, + 0.03234538808465004, + 0.013531415723264217, + 0.01402493566274643, + 0.018647637218236923, + 0.13372328877449036, + 0.0898522362112999, + 0.021647414192557335, + 0.022292939946055412, + 0.00491362065076828, + 0.008066240698099136, + 0.004847089760005474, + 0.004735761787742376, + 0.012105600908398628, + 0.02495713345706463, + 0.010129238478839397, + 0.009357962757349014, + 0.005357122514396906, + 0.009951951913535595, + 0.009202919900417328, + 0.010894758626818657, + 0.006170305423438549, + 0.009314139373600483, + 0.01266758143901825, + 0.010469678789377213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.47426390647888184 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0970274806022644, + 0.040271222591400146, + 0.1198488399386406, + 0.12465979158878326, + 0.09294034540653229, + 0.0392373688519001, + 0.022479107603430748, + 0.28455328941345215, + 0.09135136753320694, + 0.0876312255859375 + ], + "credit_map": [ + 0.010133950039744377, + 0.013678734190762043, + 0.008584222756326199, + 0.01211290992796421, + 0.00858389213681221, + 0.014843539334833622, + 0.013158143498003483, + 0.01593208685517311, + 0.0038617763202637434, + 0.004609588999301195, + 0.002949453890323639, + 0.005263934377580881, + 0.006186404265463352, + 0.006645233836025, + 0.005695927422493696, + 0.005058901384472847, + 0.012913492508232594, + 0.016329940408468246, + 0.01267190370708704, + 0.013772240839898586, + 0.013564594089984894, + 0.018087808042764664, + 0.017897313460707664, + 0.014611533842980862, + 0.016105076298117638, + 0.014418038539588451, + 0.017049795016646385, + 0.019366580992937088, + 0.009583622217178345, + 0.016261078417301178, + 0.018690912052989006, + 0.013184679672122002, + 0.015750965103507042, + 0.00847487524151802, + 0.014000552706420422, + 0.008809956721961498, + 0.010757548734545708, + 0.016688600182533264, + 0.010121331550180912, + 0.008336514234542847, + 0.0029233142267912626, + 0.004380153026431799, + 0.0035148216411471367, + 0.004123789723962545, + 0.011212728917598724, + 0.005675208289176226, + 0.0039403121918439865, + 0.0034670403692871332, + 0.002324872650206089, + 0.002181952353566885, + 0.00249821157194674, + 0.001716427505016327, + 0.004191217012703419, + 0.004560877103358507, + 0.0024713154416531324, + 0.0025342328008264303, + 0.028764799237251282, + 0.015298349782824516, + 0.018633509054780006, + 0.03566928952932358, + 0.09885451197624207, + 0.05768659710884094, + 0.017986686900258064, + 0.011659537442028522, + 0.006591069046407938, + 0.005633961874991655, + 0.004138526972383261, + 0.0048668645322322845, + 0.014488711021840572, + 0.033025629818439484, + 0.014098484069108963, + 0.008508125320076942, + 0.009682985953986645, + 0.013063806109130383, + 0.013162883929908276, + 0.012301539070904255, + 0.007634256966412067, + 0.009569166228175163, + 0.014903844334185123, + 0.007312743458896875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 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\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18800243735313416, + 0.05712363123893738, + 0.09826048463582993, + 0.09980104863643646, + 0.11663025617599487, + 0.07545464485883713, + 0.06258167326450348, + 0.0579090416431427, + 0.15637150406837463, + 0.08786525577306747 + ], + "credit_map": [ + 0.017828529700636864, + 0.021210912615060806, + 0.022543108090758324, + 0.028428465127944946, + 0.01589622162282467, + 0.021384961903095245, + 0.02563198283314705, + 0.035078246146440506, + 0.004852178506553173, + 0.006861577276140451, + 0.005201782565563917, + 0.008053058758378029, + 0.00607773382216692, + 0.00866954680532217, + 0.008079592138528824, + 0.009328160434961319, + 0.011784433387219906, + 0.008481866680085659, + 0.00926235318183899, + 0.011305363848805428, + 0.007618146017193794, + 0.016524819657206535, + 0.01842859946191311, + 0.014854899607598782, + 0.013733569532632828, + 0.012026813812553883, + 0.013358443975448608, + 0.01070250105112791, + 0.011416468769311905, + 0.011626199819147587, + 0.01316124014556408, + 0.013775809668004513, + 0.011848784051835537, + 0.008788127452135086, + 0.020015571266412735, + 0.01317628100514412, + 0.019321128726005554, + 0.02185317501425743, + 0.012147939763963223, + 0.009479254484176636, + 0.007271153852343559, + 0.008312640711665154, + 0.008670873939990997, + 0.008904769085347652, + 0.01754719763994217, + 0.012383177876472473, + 0.007557491306215525, + 0.004807342775166035, + 0.007528801914304495, + 0.007458838168531656, + 0.007026939187198877, + 0.005879552103579044, + 0.01209450326859951, + 0.010823900811374187, + 0.006865509785711765, + 0.004903630819171667, + 0.0053415787406265736, + 0.004352755378931761, + 0.0040542748756706715, + 0.005733262747526169, + 0.018423594534397125, + 0.012360906228423119, + 0.004322242457419634, + 0.0033204222563654184, + 0.012694641947746277, + 0.022131578996777534, + 0.02149198204278946, + 0.026336273178458214, + 0.021182086318731308, + 0.017346404492855072, + 0.015207432210445404, + 0.01998109184205532, + 0.010287056677043438, + 0.012627302668988705, + 0.011527973227202892, + 0.00871514156460762, + 0.012903321534395218, + 0.009813591837882996, + 0.012060255743563175, + 0.009930606931447983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.012536413036286831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the wave's core color deep indigo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wave's sharp peak appear translucent turquoise with shimmering sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the wave's crest curve inward to form a tunnel that is slowly closing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a dramatic, cold palette of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063440591096878 + ], + "opposite_logprob_tokens": [ + -4.141063213348389 + ], + "logprob_reward": -0.016063440591096878, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063440591096878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0890333503484726, + 0.09892530739307404, + 0.09112143516540527, + 0.14312133193016052, + 0.06834667176008224, + 0.17995736002922058, + 0.023777194321155548, + 0.16210003197193146, + 0.09569450467824936, + 0.047922760248184204 + ], + "credit_map": [ + 0.008302991278469563, + 0.007931352593004704, + 0.009692559950053692, + 0.014996295794844627, + 0.006907857023179531, + 0.013720636256039143, + 0.013918050564825535, + 0.01356362272053957, + 0.00965970940887928, + 0.008560920134186745, + 0.008483381941914558, + 0.013853237964212894, + 0.013927312567830086, + 0.015792502090334892, + 0.014439959079027176, + 0.014208294451236725, + 0.012467510998249054, + 0.010230154730379581, + 0.00832286849617958, + 0.009683476760983467, + 0.008023696020245552, + 0.015472973696887493, + 0.01780015602707863, + 0.009120606817305088, + 0.023338334634900093, + 0.016114013269543648, + 0.022189809009432793, + 0.012645523995161057, + 0.012387678027153015, + 0.01919713243842125, + 0.020473798736929893, + 0.016775069758296013, + 0.008037487044930458, + 0.0064844293519854546, + 0.010083583183586597, + 0.007540066260844469, + 0.01169949397444725, + 0.010267660953104496, + 0.006867786403745413, + 0.007366173435002565, + 0.02950521931052208, + 0.021373465657234192, + 0.027384812012314796, + 0.021893631666898727, + 0.026145633310079575, + 0.025136694312095642, + 0.015366487205028534, + 0.013151425868272781, + 0.0030730057042092085, + 0.002264216775074601, + 0.0027790628373622894, + 0.0025529153645038605, + 0.0052335248328745365, + 0.003586288308724761, + 0.0025980432983487844, + 0.0016901426715776324, + 0.011268146336078644, + 0.013758552260696888, + 0.015519457869231701, + 0.015475527383387089, + 0.035453785210847855, + 0.035268187522888184, + 0.01573057472705841, + 0.01962583139538765, + 0.00686939712613821, + 0.010053237900137901, + 0.005221085157245398, + 0.007057073060423136, + 0.01863509975373745, + 0.027010712772607803, + 0.013760016299784184, + 0.007087902165949345, + 0.006398690398782492, + 0.0059205335564911366, + 0.008077836595475674, + 0.006826962810009718, + 0.0037863710895180702, + 0.005976590793579817, + 0.0048936763778328896, + 0.006042104680091143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.016063440591096878 + } + ], + "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/0000152/group_4/index_tensor([6625, 6625])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_tensor([6625, 6625])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..04993950852cbb4df98816b007cc2f6b99c2e84a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_tensor([6625, 6625])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "precise, full-body motion capture, captured with a dynamic tracking angle similar to a sports broadcast, follows a heavy transport truck. Set in a sprawling industrial complex, a thick, shrouded industrial fog obscures everything, blurring the distant steel structures and pipelines. In the background, a massive turbine emits a steady, low roar, its massive blades turning slowly and powerfully in the hazy air, creating a breathtaking sense of scale. The multi-axle truck, painted a weathered matte dark gray, its massive tires caked with mud, accelerates from a standstill with a deafening roar of its engine, its tires gripping the slippery concrete and sending up a cloud of spray. The camera captures its rapid acceleration with exceptional steadiness, highlighting the striking contrast of its dynamic mass and size with the vast, relatively static industrial complex in the background. Towering searchlights cast several beams of cold white light from above, penetrating the fog and forming moving spots of light and long shadows on the ground. The entire picture is dominated by cold blue and metallic gray, creating an atmosphere full of power, tension and grand industrial aesthetics.", + "seed": 9454, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_6625_seed_9454_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": -0.1949458561837673, + "vlm_logprob": -0.1949458561837673, + "vlm_yes_no_margin": 2.05, + "vlm_token_credit": -0.1949458561837673, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.043919533491134644 + ], + "opposite_logprob_tokens": [ + -3.543919563293457 + ], + "logprob_reward": -0.043919533491134644, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.043919533491134644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1154153048992157, + 0.07303068786859512, + 0.04067201167345047, + 0.08161109685897827, + 0.10583426058292389, + 0.24511635303497314, + 0.04092679172754288, + 0.1479220688343048, + 0.08449462801218033, + 0.0649767816066742 + ], + "credit_map": [ + 0.00910377036780119, + 0.01405874639749527, + 0.011259162798523903, + 0.012033738195896149, + 0.019809454679489136, + 0.016534989699721336, + 0.016200056299567223, + 0.016415385529398918, + 0.009064296260476112, + 0.007388326805084944, + 0.007887838408350945, + 0.00701726321130991, + 0.009883187711238861, + 0.014653457328677177, + 0.006895160302519798, + 0.010241154581308365, + 0.0051566786132752895, + 0.004115004558116198, + 0.004111018031835556, + 0.0038908813148736954, + 0.004581437446177006, + 0.009762302972376347, + 0.0033994237892329693, + 0.005655262153595686, + 0.010771173983812332, + 0.012479403056204319, + 0.009924703277647495, + 0.008313105441629887, + 0.010530995205044746, + 0.009760838933289051, + 0.007959545589983463, + 0.011871334165334702, + 0.016199970617890358, + 0.009346923790872097, + 0.017732826992869377, + 0.012704840861260891, + 0.008652052842080593, + 0.022862639278173447, + 0.007606291212141514, + 0.010728715918958187, + 0.046198662370443344, + 0.02964923530817032, + 0.02457500994205475, + 0.028399983420968056, + 0.015849506482481956, + 0.05206726863980293, + 0.022922098636627197, + 0.025454597547650337, + 0.004792408086359501, + 0.005505394656211138, + 0.00687568262219429, + 0.005797681864351034, + 0.005590728484094143, + 0.0036952258087694645, + 0.004219638649374247, + 0.004450032953172922, + 0.017783576622605324, + 0.014600238762795925, + 0.02003275230526924, + 0.02323349192738533, + 0.01665034145116806, + 0.015495811589062214, + 0.018461966887116432, + 0.02166389860212803, + 0.011701932176947594, + 0.011076652444899082, + 0.008988536894321442, + 0.00924207828938961, + 0.005559584591537714, + 0.008516483940184116, + 0.011159680783748627, + 0.01824967749416828, + 0.010376354679465294, + 0.008879722096025944, + 0.016729578375816345, + 0.005733588710427284, + 0.006349881179630756, + 0.004406334832310677, + 0.0062357415445148945, + 0.006265580188483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.043919533491134644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.2521366775035858 + ], + "opposite_logprob_tokens": [ + -1.5021367073059082 + ], + "logprob_reward": -0.2521366775035858, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521366775035858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08609554171562195, + 0.047109346836805344, + 0.03406732156872749, + 0.09233740717172623, + 0.13815966248512268, + 0.39126837253570557, + 0.033701397478580475, + 0.06906221807003021, + 0.03638176992535591, + 0.07181695103645325 + ], + "credit_map": [ + 0.006950482726097107, + 0.012781416065990925, + 0.010759971104562283, + 0.00881152506917715, + 0.011918002739548683, + 0.0158147644251585, + 0.00973878987133503, + 0.009320586919784546, + 0.0038313427940011024, + 0.004649169277399778, + 0.004702726379036903, + 0.007743800524622202, + 0.007078066468238831, + 0.006743310019373894, + 0.005870181135833263, + 0.006490751169621944, + 0.004874528851360083, + 0.002931412076577544, + 0.0035218317061662674, + 0.004202122334390879, + 0.004769646562635899, + 0.005882756318897009, + 0.004162385128438473, + 0.0037226364947855473, + 0.013527996838092804, + 0.010772638022899628, + 0.011044641025364399, + 0.013582631945610046, + 0.007880684919655323, + 0.011625362560153008, + 0.010528285056352615, + 0.013375164940953255, + 0.018435142934322357, + 0.021206608042120934, + 0.015408217906951904, + 0.02015509083867073, + 0.013435089029371738, + 0.016419662162661552, + 0.013125075958669186, + 0.019974779337644577, + 0.09308428317308426, + 0.04480428248643875, + 0.03363186866044998, + 0.04340285062789917, + 0.02192101813852787, + 0.09211239218711853, + 0.03177027776837349, + 0.03054139018058777, + 0.0037446958012878895, + 0.0031540989875793457, + 0.007490116637200117, + 0.003800664795562625, + 0.001945939497090876, + 0.003126258496195078, + 0.0032073860056698322, + 0.007232239004224539, + 0.008188622072339058, + 0.0074020130559802055, + 0.009393163956701756, + 0.007591831032186747, + 0.008854382671415806, + 0.008239185437560081, + 0.008232500404119492, + 0.011160524562001228, + 0.004558124579489231, + 0.005130616016685963, + 0.004681331105530262, + 0.00542862294241786, + 0.0026622824370861053, + 0.0030047574546188116, + 0.004218476824462414, + 0.006697560660541058, + 0.012756112031638622, + 0.016097724437713623, + 0.016719145700335503, + 0.006751012988388538, + 0.004026705864816904, + 0.005697995889931917, + 0.004659665748476982, + 0.005108592100441456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.2521366775035858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.05494685098528862 + ], + "opposite_logprob_tokens": [ + -2.9299468994140625 + ], + "logprob_reward": -0.05494685098528862, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494685098528862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17097720503807068, + 0.12716832756996155, + 0.08511794358491898, + 0.18061120808124542, + 0.07237271219491959, + 0.15922604501247406, + 0.04680970683693886, + 0.07403770089149475, + 0.05310414731502533, + 0.03057500720024109 + ], + "credit_map": [ + 0.02001110464334488, + 0.01801597699522972, + 0.013268444687128067, + 0.016526510939002037, + 0.031459372490644455, + 0.03354143351316452, + 0.01792048290371895, + 0.020233871415257454, + 0.011350960470736027, + 0.01331261545419693, + 0.014016557484865189, + 0.012958825565874577, + 0.019224505871534348, + 0.02232849784195423, + 0.015389636158943176, + 0.01858673058450222, + 0.008551876060664654, + 0.008251979947090149, + 0.009791899472475052, + 0.007061479613184929, + 0.009337968192994595, + 0.01460201758891344, + 0.016016760841012, + 0.01150396466255188, + 0.037632815539836884, + 0.025880251079797745, + 0.02289503626525402, + 0.014539087191224098, + 0.018156154081225395, + 0.023377474397420883, + 0.014770922251045704, + 0.023359473794698715, + 0.011548248119652271, + 0.006402806844562292, + 0.010329230688512325, + 0.008982409723103046, + 0.005774174816906452, + 0.01505355816334486, + 0.006247221492230892, + 0.008035063743591309, + 0.03358519822359085, + 0.023661913350224495, + 0.012834956869482994, + 0.0165119469165802, + 0.01131493877619505, + 0.03232855722308159, + 0.014031586237251759, + 0.014956939965486526, + 0.006603407207876444, + 0.0044106184504926205, + 0.009436161257326603, + 0.005981495138257742, + 0.004256700165569782, + 0.004041094798594713, + 0.004141983576118946, + 0.007938246242702007, + 0.007157950196415186, + 0.00971516128629446, + 0.008225510828197002, + 0.008491511456668377, + 0.010874641127884388, + 0.008103953674435616, + 0.00927682500332594, + 0.012192140333354473, + 0.0075256298296153545, + 0.008963266387581825, + 0.007965696044266224, + 0.006758201867341995, + 0.0030072990339249372, + 0.0045193699188530445, + 0.005658411420881748, + 0.008706270717084408, + 0.0037328789476305246, + 0.004987378604710102, + 0.004577456042170525, + 0.003651547245681286, + 0.002336438512429595, + 0.0019305248279124498, + 0.004556390922516584, + 0.004802393727004528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.05494685098528862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.5234111547470093 + ], + "opposite_logprob_tokens": [ + -0.8984111547470093 + ], + "logprob_reward": -0.5234111547470093, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234111547470093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37523072957992554, + 0.11138711869716644, + 0.11209924519062042, + 0.16612674295902252, + 0.0323215126991272, + 0.10257577151060104, + 0.021086402237415314, + 0.031588997691869736, + 0.029147112742066383, + 0.018436389043927193 + ], + "credit_map": [ + 0.03884274885058403, + 0.06692080944776535, + 0.022281978279352188, + 0.024491677060723305, + 0.07837219536304474, + 0.06193076819181442, + 0.04438404738903046, + 0.038006491959095, + 0.012550653889775276, + 0.009070995263755322, + 0.007376973517239094, + 0.007704705931246281, + 0.01918906904757023, + 0.029845349490642548, + 0.011958250775933266, + 0.013691132888197899, + 0.010576083324849606, + 0.008788187056779861, + 0.007664185483008623, + 0.0061731114983558655, + 0.01624957285821438, + 0.03895419090986252, + 0.008238200098276138, + 0.015455706045031548, + 0.03444367274641991, + 0.015250473283231258, + 0.01473233476281166, + 0.01587868109345436, + 0.014860308729112148, + 0.017479395493865013, + 0.028793923556804657, + 0.024687951430678368, + 0.004820257890969515, + 0.0037241182290017605, + 0.004099749028682709, + 0.004011764656752348, + 0.0021733627654612064, + 0.005515860393643379, + 0.0045234886929392815, + 0.0034529122058302164, + 0.01643790490925312, + 0.012180926278233528, + 0.011854296550154686, + 0.01203393004834652, + 0.006970476824790239, + 0.01878182403743267, + 0.010380751453340054, + 0.013935663737356663, + 0.003470125375315547, + 0.002396396594122052, + 0.003672470338642597, + 0.00266314041800797, + 0.0018108736258000135, + 0.0026695849373936653, + 0.0019466077210381627, + 0.002457202645018697, + 0.00296204537153244, + 0.003429945558309555, + 0.004872993566095829, + 0.0036970602814108133, + 0.0045883310958743095, + 0.003812360344454646, + 0.0037512832786887884, + 0.004474979359656572, + 0.003982485271990299, + 0.004348562099039555, + 0.004085569176822901, + 0.003721588524058461, + 0.0022656808141618967, + 0.0025190638843923807, + 0.003534632967785001, + 0.004689529538154602, + 0.002387647284194827, + 0.0025949901901185513, + 0.004124694969505072, + 0.0021931659430265427, + 0.0018968458753079176, + 0.0012585308868438005, + 0.0021181399933993816, + 0.0018623743671923876 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.5234111547470093 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.10031506419181824 + ], + "opposite_logprob_tokens": [ + -2.3503150939941406 + ], + "logprob_reward": -0.10031506419181824, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031506419181824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07770589739084244, + 0.03253021463751793, + 0.026467327028512955, + 0.10260152816772461, + 0.03934910148382187, + 0.5305567979812622, + 0.009954017587006092, + 0.02923819236457348, + 0.04269932210445404, + 0.10889764875173569 + ], + "credit_map": [ + 0.008978192694485188, + 0.008768199011683464, + 0.005327560938894749, + 0.007557934615761042, + 0.011997944675385952, + 0.013517984189093113, + 0.009980806149542332, + 0.011577273719012737, + 0.00512712961062789, + 0.003688712138682604, + 0.0026694361586123705, + 0.003565762424841523, + 0.004630950279533863, + 0.0053507424890995026, + 0.003637548303231597, + 0.0038599330000579357, + 0.0030030300840735435, + 0.002478145994246006, + 0.0018457855330780149, + 0.0022929497063159943, + 0.0034068638924509287, + 0.007900916039943695, + 0.0020907761063426733, + 0.003448860952630639, + 0.01810840703547001, + 0.009901221841573715, + 0.007846235297620296, + 0.007867933250963688, + 0.008467952720820904, + 0.011126515455543995, + 0.020973455160856247, + 0.018309812992811203, + 0.007352717220783234, + 0.004967926535755396, + 0.005150061100721359, + 0.003950407262891531, + 0.002676436211913824, + 0.008191550150513649, + 0.0035231781657785177, + 0.0035368232056498528, + 0.1336982250213623, + 0.0525968074798584, + 0.04272802174091339, + 0.058558981865644455, + 0.027822021394968033, + 0.1337534785270691, + 0.0433150976896286, + 0.038084130734205246, + 0.0013451119884848595, + 0.0012180021731182933, + 0.0017849034629762173, + 0.0014958371175453067, + 0.0008009130833670497, + 0.0009325381251983345, + 0.0010163364931941032, + 0.0013603748520836234, + 0.002722875913605094, + 0.0025189961306750774, + 0.004785534460097551, + 0.003415775950998068, + 0.0027710713911801577, + 0.0031178705394268036, + 0.0035594485234469175, + 0.006346621550619602, + 0.0052770283073186874, + 0.00574862165376544, + 0.005089206155389547, + 0.006208761595189571, + 0.003573421621695161, + 0.0034528423566371202, + 0.006103610619902611, + 0.007245829328894615, + 0.01592981442809105, + 0.020671360194683075, + 0.02546301670372486, + 0.013757115229964256, + 0.007096017710864544, + 0.008859259076416492, + 0.009105542674660683, + 0.008015526458621025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.10031506419181824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "precise, full-body motion capture, captured with a dynamic tracking angle similar to a sports broadcast, follows a heavy transport truck. Set in a sprawling industrial complex, a thick, shrouded industrial fog obscures everything, blurring the distant steel structures and pipelines. In the background, a massive turbine emits a steady, low roar, its massive blades turning slowly and powerfully in the hazy air, creating a breathtaking sense of scale. The multi-axle truck, painted a weathered matte dark gray, its massive tires caked with mud, accelerates from a standstill with a deafening roar of its engine, its tires gripping the slippery concrete and sending up a cloud of spray. The camera captures its rapid acceleration with exceptional steadiness, highlighting the striking contrast of its dynamic mass and size with the vast, relatively static industrial complex in the background. Towering searchlights cast several beams of cold white light from above, penetrating the fog and forming moving spots of light and long shadows on the ground. The entire picture is dominated by cold blue and metallic gray, creating an atmosphere full of power, tension and grand industrial aesthetics.", + "seed": 9455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_6625_seed_9455_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": -0.25235852859914304, + "vlm_logprob": -0.25235852859914304, + "vlm_yes_no_margin": 2.300000011920929, + "vlm_token_credit": -0.25235852859914304, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.027328232303261757 + ], + "opposite_logprob_tokens": [ + -3.9023282527923584 + ], + "logprob_reward": -0.027328232303261757, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.027328232303261757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027519315481185913, + 0.02917328290641308, + 0.06548549234867096, + 0.14437736570835114, + 0.05551616847515106, + 0.08574409782886505, + 0.5004217028617859, + 0.02938666008412838, + 0.038137730211019516, + 0.024238213896751404 + ], + "credit_map": [ + 0.0028907989617437124, + 0.0034147812984883785, + 0.003062367206439376, + 0.003178950399160385, + 0.005024961661547422, + 0.002657524775713682, + 0.0039858161471784115, + 0.003304116427898407, + 0.002476390451192856, + 0.0030222167260944843, + 0.003351400839164853, + 0.002827410353347659, + 0.006142971571534872, + 0.005768198985606432, + 0.0028675287030637264, + 0.002717164810746908, + 0.008150480687618256, + 0.012289149686694145, + 0.008020284585654736, + 0.0061527034267783165, + 0.008884167298674583, + 0.005787867121398449, + 0.008957672864198685, + 0.00724316481500864, + 0.019269123673439026, + 0.025100698694586754, + 0.014220952056348324, + 0.010691807605326176, + 0.01546464953571558, + 0.03326127678155899, + 0.012418915517628193, + 0.013949951156973839, + 0.007505905348807573, + 0.005881924647837877, + 0.00794101133942604, + 0.004399820230901241, + 0.008068488910794258, + 0.0064027621410787106, + 0.008264176547527313, + 0.0070520807057619095, + 0.011769902892410755, + 0.008958514779806137, + 0.016384458169341087, + 0.007820185273885727, + 0.008864057250320911, + 0.012524460442364216, + 0.007206270936876535, + 0.01221624854952097, + 0.1661391258239746, + 0.17647087574005127, + 0.018043996766209602, + 0.02001512236893177, + 0.031101172789931297, + 0.02443000301718712, + 0.03583195060491562, + 0.028389455750584602, + 0.0038815983571112156, + 0.003202445339411497, + 0.001599038252606988, + 0.0019234138308092952, + 0.0027433009818196297, + 0.007967901416122913, + 0.0030168648809194565, + 0.005052096210420132, + 0.008085047826170921, + 0.0035416344180703163, + 0.009185567498207092, + 0.00212086527608335, + 0.005713194143027067, + 0.004096202086657286, + 0.002946362132206559, + 0.002448857529088855, + 0.003602411365136504, + 0.003330519422888756, + 0.0023061935789883137, + 0.004311508499085903, + 0.0026705635245889425, + 0.002515398897230625, + 0.002521730726584792, + 0.002979888580739498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.027328232303261757 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.34866851568222046 + ], + "opposite_logprob_tokens": [ + -1.2236685752868652 + ], + "logprob_reward": -0.34866851568222046, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34866851568222046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05231103673577309, + 0.06776459515094757, + 0.10028534382581711, + 0.180145725607872, + 0.0372258685529232, + 0.05272061005234718, + 0.297852098941803, + 0.0794595405459404, + 0.08191168308258057, + 0.050323475152254105 + ], + "credit_map": [ + 0.005930828396230936, + 0.008415821939706802, + 0.006463041994720697, + 0.00785604864358902, + 0.00737549876794219, + 0.004426673986017704, + 0.00739920511841774, + 0.004443923011422157, + 0.007830612361431122, + 0.007853862829506397, + 0.00738762179389596, + 0.010826028883457184, + 0.011620710603892803, + 0.0072693852707743645, + 0.009092009626328945, + 0.005884375888854265, + 0.011952687054872513, + 0.02038709819316864, + 0.016420073807239532, + 0.006914460100233555, + 0.011353596113622189, + 0.009868660941720009, + 0.012904694303870201, + 0.010484078899025917, + 0.026593545451760292, + 0.03183240815997124, + 0.017676204442977905, + 0.012995685450732708, + 0.01724356785416603, + 0.04284375533461571, + 0.013619821518659592, + 0.017340756952762604, + 0.004672544542700052, + 0.004427582025527954, + 0.008268172852694988, + 0.0030730010475963354, + 0.004820636007934809, + 0.004071895033121109, + 0.004113772884011269, + 0.003778268815949559, + 0.007393623236566782, + 0.005825709085911512, + 0.0085360212251544, + 0.004402007907629013, + 0.006375058554112911, + 0.007948800921440125, + 0.0044395229779183865, + 0.007799872197210789, + 0.10741395503282547, + 0.1144518256187439, + 0.009928504936397076, + 0.010503501631319523, + 0.018631374463438988, + 0.012326315976679325, + 0.011985468678176403, + 0.012611192651093006, + 0.00967654213309288, + 0.011418421752750874, + 0.00805730652064085, + 0.0076156482100486755, + 0.012289538979530334, + 0.011424899101257324, + 0.00832610297948122, + 0.010651091113686562, + 0.016247572377324104, + 0.007695831824094057, + 0.01887073740363121, + 0.006804470904171467, + 0.011001656763255596, + 0.009443122893571854, + 0.006208659615367651, + 0.0056396410800516605, + 0.006030282936990261, + 0.00716770812869072, + 0.007327926345169544, + 0.009249215945601463, + 0.00604241993278265, + 0.004362940322607756, + 0.004744629841297865, + 0.005398356821388006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.34866851568222046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.029796933755278587 + ], + "opposite_logprob_tokens": [ + -3.529796838760376 + ], + "logprob_reward": -0.029796933755278587, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796933755278587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04506755620241165, + 0.052800945937633514, + 0.07446521520614624, + 0.17010772228240967, + 0.05061326175928116, + 0.07829292118549347, + 0.45943984389305115, + 0.020037271082401276, + 0.013698551803827286, + 0.03547675162553787 + ], + "credit_map": [ + 0.004872743505984545, + 0.004637544509023428, + 0.005590333137661219, + 0.005680215544998646, + 0.005575424525886774, + 0.005136494524776936, + 0.007976359687745571, + 0.005598440300673246, + 0.004851056728512049, + 0.006542665883898735, + 0.00544758653268218, + 0.007273640017956495, + 0.008945190347731113, + 0.008567072451114655, + 0.006582815200090408, + 0.004590919706970453, + 0.011284262873232365, + 0.015330611728131771, + 0.009567514061927795, + 0.005955375265330076, + 0.00990446750074625, + 0.0064325579442083836, + 0.008930766955018044, + 0.007059658411890268, + 0.021816084161400795, + 0.03149181231856346, + 0.016573891043663025, + 0.012718613259494305, + 0.016118546947836876, + 0.041728366166353226, + 0.013829778879880905, + 0.01583063416182995, + 0.006526031531393528, + 0.005628692917525768, + 0.00930837169289589, + 0.003963721450418234, + 0.007036652881652117, + 0.0054006678983569145, + 0.007213866338133812, + 0.005535257514566183, + 0.010943476110696793, + 0.00947599671781063, + 0.014904157258570194, + 0.0066751232370734215, + 0.00819604005664587, + 0.011316874995827675, + 0.00579410744830966, + 0.01098714955151081, + 0.15348516404628754, + 0.16344016790390015, + 0.01612534187734127, + 0.01768294721841812, + 0.028229061514139175, + 0.022060830146074295, + 0.03216264769434929, + 0.026253676041960716, + 0.0030539771541953087, + 0.0019424722995609045, + 0.0009285116102546453, + 0.0010417008306831121, + 0.0017314512515440583, + 0.0066060153767466545, + 0.0014868879225105047, + 0.0032462545204907656, + 0.003728393232449889, + 0.0010679636616259813, + 0.0035919479560106993, + 0.0008201151504181325, + 0.0011443498078733683, + 0.001222318154759705, + 0.001139876781962812, + 0.00098358653485775, + 0.005170160438865423, + 0.0069855451583862305, + 0.004820581991225481, + 0.007535506039857864, + 0.002938369521871209, + 0.0021854680962860584, + 0.0032877298071980476, + 0.002553388709202409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.029796933755278587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.8261958956718445 + ], + "opposite_logprob_tokens": [ + -0.5761958956718445 + ], + "logprob_reward": -0.8261958956718445, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261958956718445, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15676265954971313, + 0.2627127766609192, + 0.09934619069099426, + 0.03669920191168785, + 0.023346025496721268, + 0.026580261066555977, + 0.14424271881580353, + 0.11793653666973114, + 0.07761494070291519, + 0.05475869029760361 + ], + "credit_map": [ + 0.017443237826228142, + 0.025948459282517433, + 0.013920261524617672, + 0.02014649286866188, + 0.031802497804164886, + 0.0184283796697855, + 0.014766409993171692, + 0.014306922443211079, + 0.028753839433193207, + 0.03429514169692993, + 0.024577192962169647, + 0.03899114206433296, + 0.041100844740867615, + 0.028018398210406303, + 0.0407043918967247, + 0.02627183124423027, + 0.01871616393327713, + 0.01897202804684639, + 0.020215285941958427, + 0.008380931802093983, + 0.008539775386452675, + 0.010430397465825081, + 0.006532630883157253, + 0.007558983284980059, + 0.005562170408666134, + 0.006302400026470423, + 0.004309041891247034, + 0.00311443954706192, + 0.0037720687687397003, + 0.004863239359110594, + 0.004227892495691776, + 0.004547948949038982, + 0.0032180685084313154, + 0.0036851672921329737, + 0.0049920580349862576, + 0.002622304018586874, + 0.0022772105876356363, + 0.002325938083231449, + 0.0018515909323468804, + 0.0023736876901239157, + 0.003482936415821314, + 0.0029143723659217358, + 0.004903237335383892, + 0.0022348843049257994, + 0.0029196275863796473, + 0.0031585965771228075, + 0.004082379397004843, + 0.0028842275496572256, + 0.03729294240474701, + 0.041202299296855927, + 0.007149231620132923, + 0.007638338953256607, + 0.011085838079452515, + 0.009412485174834728, + 0.015242808498442173, + 0.015218782238662243, + 0.013589639216661453, + 0.020563412457704544, + 0.011026902124285698, + 0.010401499457657337, + 0.0199008546769619, + 0.01635698787868023, + 0.013695396482944489, + 0.012401843443512917, + 0.013426924124360085, + 0.00684477761387825, + 0.01625910960137844, + 0.006859994027763605, + 0.010856729932129383, + 0.010214582085609436, + 0.0074584647081792355, + 0.005694364197552204, + 0.009639366529881954, + 0.009031539782881737, + 0.013012698851525784, + 0.00773583585396409, + 0.004706121049821377, + 0.0034471964463591576, + 0.0025554148014634848, + 0.004630517214536667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.8261958956718445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.029803065583109856 + ], + "opposite_logprob_tokens": [ + -3.5298030376434326 + ], + "logprob_reward": -0.029803065583109856, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803065583109856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05275772139430046, + 0.160261332988739, + 0.10923822224140167, + 0.08170641958713531, + 0.03207226097583771, + 0.043304648250341415, + 0.3559954762458801, + 0.07185224443674088, + 0.0632326602935791, + 0.029579007998108864 + ], + "credit_map": [ + 0.00554167665541172, + 0.006987835746258497, + 0.00516124302521348, + 0.0061622606590390205, + 0.009473002515733242, + 0.0052469223737716675, + 0.008423984050750732, + 0.005760792177170515, + 0.016742028295993805, + 0.022008731961250305, + 0.01452120952308178, + 0.02383904904127121, + 0.025846658274531364, + 0.015942201018333435, + 0.02509603463113308, + 0.016265425831079483, + 0.018638648092746735, + 0.02034238539636135, + 0.0102348318323493, + 0.007978423498570919, + 0.01497082319110632, + 0.01828598417341709, + 0.010265371762216091, + 0.00852175522595644, + 0.010516996495425701, + 0.01585962250828743, + 0.007715046871453524, + 0.0063391695730388165, + 0.008631463162600994, + 0.018292279914021492, + 0.007092804182320833, + 0.007259033620357513, + 0.00424595270305872, + 0.003275538794696331, + 0.005057742819190025, + 0.0028624169062823057, + 0.0050807190127670765, + 0.0036057610996067524, + 0.004358845762908459, + 0.0035852829460054636, + 0.005795667879283428, + 0.00552211282774806, + 0.009619955904781818, + 0.0037686722353100777, + 0.003931622486561537, + 0.005551141686737537, + 0.003561275778338313, + 0.0055541954934597015, + 0.1184878945350647, + 0.12565995752811432, + 0.012237679213285446, + 0.013535448350012302, + 0.021705515682697296, + 0.017130373045802116, + 0.02706781029701233, + 0.020170822739601135, + 0.00830857828259468, + 0.01070944219827652, + 0.006089097820222378, + 0.006186353042721748, + 0.012880117632448673, + 0.011209146119654179, + 0.008051162585616112, + 0.008418337441980839, + 0.006933107972145081, + 0.0058789425529539585, + 0.006650850176811218, + 0.00566462567076087, + 0.012250362895429134, + 0.011084135621786118, + 0.008608978241682053, + 0.0061616599559783936, + 0.0050116837956011295, + 0.006200406234711409, + 0.004675562027841806, + 0.00507556926459074, + 0.002750915242359042, + 0.0022434061393141747, + 0.0017833234742283821, + 0.0018381410045549273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.029803065583109856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "precise, full-body motion capture, captured with a dynamic tracking angle similar to a sports broadcast, follows a heavy transport truck. Set in a sprawling industrial complex, a thick, shrouded industrial fog obscures everything, blurring the distant steel structures and pipelines. In the background, a massive turbine emits a steady, low roar, its massive blades turning slowly and powerfully in the hazy air, creating a breathtaking sense of scale. The multi-axle truck, painted a weathered matte dark gray, its massive tires caked with mud, accelerates from a standstill with a deafening roar of its engine, its tires gripping the slippery concrete and sending up a cloud of spray. The camera captures its rapid acceleration with exceptional steadiness, highlighting the striking contrast of its dynamic mass and size with the vast, relatively static industrial complex in the background. Towering searchlights cast several beams of cold white light from above, penetrating the fog and forming moving spots of light and long shadows on the ground. The entire picture is dominated by cold blue and metallic gray, creating an atmosphere full of power, tension and grand industrial aesthetics.", + "seed": 9456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_6625_seed_9456_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": -0.3329491727054119, + "vlm_logprob": -0.3329491727054119, + "vlm_yes_no_margin": 1.425, + "vlm_token_credit": -0.3329491727054119, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.05829932913184166 + ], + "opposite_logprob_tokens": [ + -3.0582993030548096 + ], + "logprob_reward": -0.05829932913184166, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05829932913184166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16841094195842743, + 0.06473030894994736, + 0.03673958778381348, + 0.0756475031375885, + 0.14186248183250427, + 0.05450724810361862, + 0.07089803367853165, + 0.156859889626503, + 0.12201346457004547, + 0.10833049565553665 + ], + "credit_map": [ + 0.01576162502169609, + 0.018697133287787437, + 0.029350515455007553, + 0.016548054292798042, + 0.018502116203308105, + 0.018885795027017593, + 0.024675775319337845, + 0.025989947840571404, + 0.009183375164866447, + 0.008029353804886341, + 0.008160474710166454, + 0.00565811013802886, + 0.00839922484010458, + 0.008656143210828304, + 0.007134885992854834, + 0.00950874388217926, + 0.0042380597442388535, + 0.003103826427832246, + 0.004931291565299034, + 0.007107085082679987, + 0.003035797970369458, + 0.0037492532283067703, + 0.005668542813509703, + 0.0049057346768677235, + 0.00982483010739088, + 0.00949885230511427, + 0.010033066384494305, + 0.007622304372489452, + 0.006966680288314819, + 0.012945812195539474, + 0.006570202298462391, + 0.012185757048428059, + 0.0181629229336977, + 0.013293316587805748, + 0.04091215133666992, + 0.013209239579737186, + 0.009968195110559464, + 0.014775380492210388, + 0.014034037478268147, + 0.01750725321471691, + 0.006621648091822863, + 0.006485437974333763, + 0.007439168635755777, + 0.004908385686576366, + 0.008482796140015125, + 0.0066854036413133144, + 0.007743935100734234, + 0.0061404770240187645, + 0.00613377895206213, + 0.005973923485726118, + 0.009332170709967613, + 0.00989831704646349, + 0.0063795181922614574, + 0.007016516290605068, + 0.01886584237217903, + 0.007297975476831198, + 0.01726342737674713, + 0.019453950226306915, + 0.016594508662819862, + 0.025970282033085823, + 0.020265977829694748, + 0.0211995430290699, + 0.016181237995624542, + 0.019930969923734665, + 0.011690668761730194, + 0.012730911374092102, + 0.011751698330044746, + 0.013189796358346939, + 0.015551451593637466, + 0.016628913581371307, + 0.02074464224278927, + 0.019725395366549492, + 0.011996963992714882, + 0.013243340887129307, + 0.020571863278746605, + 0.008628159761428833, + 0.011195504106581211, + 0.015039927326142788, + 0.013108478859066963, + 0.01454626303166151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.05829932913184166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.4289025068283081 + ], + "opposite_logprob_tokens": [ + -1.053902506828308 + ], + "logprob_reward": -0.4289025068283081, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289025068283081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11480700969696045, + 0.0828949436545372, + 0.04339331388473511, + 0.11733648180961609, + 0.15138103067874908, + 0.04303192347288132, + 0.12253300100564957, + 0.06123926490545273, + 0.0829038918018341, + 0.18047912418842316 + ], + "credit_map": [ + 0.01279251929372549, + 0.01707807555794716, + 0.018242906779050827, + 0.012762023136019707, + 0.011849779635667801, + 0.01285862922668457, + 0.017399882897734642, + 0.011823196895420551, + 0.007362036500126123, + 0.010729367844760418, + 0.00821698922663927, + 0.00803831685334444, + 0.007061085198074579, + 0.011372880078852177, + 0.008099595084786415, + 0.022014671936631203, + 0.0067207543179392815, + 0.0059672449715435505, + 0.005839143879711628, + 0.004477246664464474, + 0.005731435026973486, + 0.0052892654202878475, + 0.005727642681449652, + 0.00364057719707489, + 0.01735500432550907, + 0.009191893972456455, + 0.02355208620429039, + 0.009512966498732567, + 0.009823152795433998, + 0.018877381458878517, + 0.00892555434256792, + 0.02009844407439232, + 0.02088407799601555, + 0.011857770383358002, + 0.03220090642571449, + 0.010789359919726849, + 0.009813645854592323, + 0.022840255871415138, + 0.016149263828992844, + 0.02684575878083706, + 0.005463569425046444, + 0.0037654319312423468, + 0.009403125382959843, + 0.004443823359906673, + 0.005676406901329756, + 0.005005887243896723, + 0.005325546022504568, + 0.003948129713535309, + 0.009271076880395412, + 0.007065040059387684, + 0.015711817890405655, + 0.016137951985001564, + 0.009706302545964718, + 0.00881544966250658, + 0.04012903571128845, + 0.01569632813334465, + 0.006483661010861397, + 0.006860498338937759, + 0.010208127088844776, + 0.010320642963051796, + 0.008274544961750507, + 0.0068292636424303055, + 0.006655337288975716, + 0.00560718821361661, + 0.00892560463398695, + 0.011042780242860317, + 0.014884336851537228, + 0.007944085635244846, + 0.00837567076086998, + 0.008790056221187115, + 0.009464472532272339, + 0.013476891443133354, + 0.02150026522576809, + 0.024595890194177628, + 0.036495305597782135, + 0.016856135800480843, + 0.025065762922167778, + 0.017443310469388962, + 0.021055204793810844, + 0.017467252910137177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 6 + ], + "selected_reward": -0.4289025068283081 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.2813262939453125 + ], + "opposite_logprob_tokens": [ + -1.4063262939453125 + ], + "logprob_reward": -0.2813262939453125, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813262939453125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21058854460716248, + 0.14915220439434052, + 0.1464862823486328, + 0.07448779791593552, + 0.04011170193552971, + 0.05093176290392876, + 0.14194908738136292, + 0.04102388769388199, + 0.052205439656972885, + 0.09306329488754272 + ], + "credit_map": [ + 0.028349079191684723, + 0.027139637619256973, + 0.035109251737594604, + 0.0221843421459198, + 0.02128567174077034, + 0.023780707269906998, + 0.03131081908941269, + 0.021429020911455154, + 0.020205434411764145, + 0.021820958703756332, + 0.015059144236147404, + 0.01821175217628479, + 0.014585520140826702, + 0.018761998042464256, + 0.019054915755987167, + 0.021452464163303375, + 0.017827218398451805, + 0.012961296364665031, + 0.025680361315608025, + 0.01496189646422863, + 0.015189399011433125, + 0.017252566292881966, + 0.0141723258420825, + 0.028441207483410835, + 0.00971448514610529, + 0.006309394259005785, + 0.008731905370950699, + 0.005742157343775034, + 0.007142127025872469, + 0.011452795937657356, + 0.006213486660271883, + 0.01918143592774868, + 0.005974898114800453, + 0.0030290032736957073, + 0.0062434556894004345, + 0.002954237861558795, + 0.0037983509246259928, + 0.006093481555581093, + 0.0035347098018974066, + 0.008483555167913437, + 0.006199945695698261, + 0.004515101667493582, + 0.009815414436161518, + 0.00416490389034152, + 0.005827067419886589, + 0.004067299421876669, + 0.007650468032807112, + 0.008691557683050632, + 0.01112543884664774, + 0.010160165838897228, + 0.020133310928940773, + 0.018918955698609352, + 0.012194795534014702, + 0.013126790523529053, + 0.043128110468387604, + 0.013161508366465569, + 0.0045003630220890045, + 0.006079852115362883, + 0.003846867009997368, + 0.008313765749335289, + 0.005148854106664658, + 0.003766813315451145, + 0.004189580678939819, + 0.0051777851767838, + 0.0068465410731732845, + 0.011193948797881603, + 0.006896475795656443, + 0.005635891575366259, + 0.00482829287648201, + 0.006293945014476776, + 0.0050169373862445354, + 0.005493399687111378, + 0.015509353019297123, + 0.010816738940775394, + 0.03285480663180351, + 0.00829107966274023, + 0.005259777419269085, + 0.007156017702072859, + 0.007097650319337845, + 0.006077863741666079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2813262939453125 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.8261888027191162 + ], + "opposite_logprob_tokens": [ + -0.5761888027191162 + ], + "logprob_reward": -0.8261888027191162, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261888027191162, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1904091238975525, + 0.05719552934169769, + 0.18385836482048035, + 0.14484508335590363, + 0.07453619688749313, + 0.04779989272356033, + 0.11779134720563889, + 0.04203150421380997, + 0.0779682919383049, + 0.06356458365917206 + ], + "credit_map": [ + 0.01883246749639511, + 0.020326612517237663, + 0.036647163331508636, + 0.02155575528740883, + 0.022023925557732582, + 0.019367102533578873, + 0.026885060593485832, + 0.02477106638252735, + 0.005962113384157419, + 0.006634640973061323, + 0.00964168831706047, + 0.006389359012246132, + 0.006275626830756664, + 0.006843647453933954, + 0.00793374888598919, + 0.007514709606766701, + 0.020847991108894348, + 0.022099457681179047, + 0.027946602553129196, + 0.019643563777208328, + 0.022081108763813972, + 0.020285753533244133, + 0.022170668467879295, + 0.028783243149518967, + 0.017207615077495575, + 0.011368965730071068, + 0.02658694237470627, + 0.011488813906908035, + 0.010414871387183666, + 0.029251782223582268, + 0.01423357892781496, + 0.024292536079883575, + 0.00998452864587307, + 0.0058461204171180725, + 0.01722589321434498, + 0.005278896540403366, + 0.005276031792163849, + 0.009547317400574684, + 0.006853051017969847, + 0.014524365775287151, + 0.005604117643088102, + 0.004692637361586094, + 0.009735429659485817, + 0.004674756899476051, + 0.006221620365977287, + 0.004507900215685368, + 0.007429690100252628, + 0.004933751653879881, + 0.009341695345938206, + 0.008539513684809208, + 0.015028157271444798, + 0.015273600816726685, + 0.009503571316599846, + 0.01066532451659441, + 0.03631824254989624, + 0.013121264986693859, + 0.004247779026627541, + 0.005541085731238127, + 0.004919723141938448, + 0.0066147190518677235, + 0.005255379714071751, + 0.004872522782534361, + 0.0033915748354047537, + 0.007188725285232067, + 0.007927455008029938, + 0.009974376298487186, + 0.01258857361972332, + 0.00874567124992609, + 0.007699891924858093, + 0.00901617668569088, + 0.009105429984629154, + 0.012910731136798859, + 0.008259748108685017, + 0.009549121372401714, + 0.012525391764938831, + 0.00634729815647006, + 0.005568893160670996, + 0.007422586902976036, + 0.007344333920627832, + 0.0065472181886434555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.8261888027191162 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.07002893090248108 + ], + "opposite_logprob_tokens": [ + -2.6950290203094482 + ], + "logprob_reward": -0.07002893090248108, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002893090248108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25269076228141785, + 0.1322353333234787, + 0.030599458143115044, + 0.22948472201824188, + 0.05261780321598053, + 0.03904538229107857, + 0.06460999697446823, + 0.04530661925673485, + 0.07420367747545242, + 0.07920617610216141 + ], + "credit_map": [ + 0.029827887192368507, + 0.028954321518540382, + 0.039850521832704544, + 0.028557948768138885, + 0.025517860427498817, + 0.02382153458893299, + 0.038795970380306244, + 0.037364743649959564, + 0.01988399215042591, + 0.01851555146276951, + 0.018405625596642494, + 0.015954818576574326, + 0.013227262534201145, + 0.01689867116510868, + 0.015062985941767693, + 0.014286438934504986, + 0.0031453014817088842, + 0.0038053286261856556, + 0.00394239230081439, + 0.004739025607705116, + 0.0032442850060760975, + 0.0036608113441616297, + 0.0035851825959980488, + 0.00447713490575552, + 0.03161557391285896, + 0.015582695603370667, + 0.046852994710206985, + 0.013131526298820972, + 0.017365751788020134, + 0.048998743295669556, + 0.01870538853108883, + 0.03723207861185074, + 0.006090410053730011, + 0.00463342247530818, + 0.01122016180306673, + 0.00409650569781661, + 0.004985255189239979, + 0.006648805923759937, + 0.0067664687521755695, + 0.008176775649189949, + 0.003910496365278959, + 0.0032201381400227547, + 0.009198488667607307, + 0.004359447862952948, + 0.003907086327672005, + 0.003961560782045126, + 0.006203318014740944, + 0.004284848924726248, + 0.005369671620428562, + 0.0049987719394266605, + 0.006921491585671902, + 0.008302266709506512, + 0.005782355088740587, + 0.005319961812347174, + 0.020373400300741196, + 0.0075420839712023735, + 0.004534791223704815, + 0.0049805245362222195, + 0.005621578078716993, + 0.009604123421013355, + 0.0037036784924566746, + 0.005812491290271282, + 0.00503317266702652, + 0.006016266066581011, + 0.007368401624262333, + 0.008434710092842579, + 0.01004955917596817, + 0.006090093404054642, + 0.006276396103203297, + 0.010935214348137379, + 0.011652054265141487, + 0.013397255912423134, + 0.013953790068626404, + 0.014570437371730804, + 0.017323724925518036, + 0.008513221517205238, + 0.004985692910850048, + 0.0072722467593848705, + 0.0056985365226864815, + 0.006888531614094973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07002893090248108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "precise, full-body motion capture, captured with a dynamic tracking angle similar to a sports broadcast, follows a heavy transport truck. Set in a sprawling industrial complex, a thick, shrouded industrial fog obscures everything, blurring the distant steel structures and pipelines. In the background, a massive turbine emits a steady, low roar, its massive blades turning slowly and powerfully in the hazy air, creating a breathtaking sense of scale. The multi-axle truck, painted a weathered matte dark gray, its massive tires caked with mud, accelerates from a standstill with a deafening roar of its engine, its tires gripping the slippery concrete and sending up a cloud of spray. The camera captures its rapid acceleration with exceptional steadiness, highlighting the striking contrast of its dynamic mass and size with the vast, relatively static industrial complex in the background. Towering searchlights cast several beams of cold white light from above, penetrating the fog and forming moving spots of light and long shadows on the ground. The entire picture is dominated by cold blue and metallic gray, creating an atmosphere full of power, tension and grand industrial aesthetics.", + "seed": 9457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_4/index_6625_seed_9457_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": -0.20941957095637917, + "vlm_logprob": -0.20941957095637917, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.20941957095637917, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.020141830667853355 + ], + "opposite_logprob_tokens": [ + -4.2701416015625 + ], + "logprob_reward": -0.020141830667853355, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020141830667853355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1461161971092224, + 0.023531779646873474, + 0.04509541019797325, + 0.11125506460666656, + 0.05510832741856575, + 0.22648915648460388, + 0.02546466328203678, + 0.024510908871889114, + 0.3035154342651367, + 0.03891308978199959 + ], + "credit_map": [ + 0.021650496870279312, + 0.015430991537868977, + 0.01682281121611595, + 0.016646381467580795, + 0.022051891312003136, + 0.015369867905974388, + 0.012693153694272041, + 0.025450609624385834, + 0.0029705320484936237, + 0.0021073543466627598, + 0.0018526605563238263, + 0.006287549622356892, + 0.0024489935021847486, + 0.0027791548054665327, + 0.0025917834136635065, + 0.0024937510024756193, + 0.002979966113343835, + 0.006789229344576597, + 0.0029055362101644278, + 0.007955489680171013, + 0.00869661569595337, + 0.006079067941755056, + 0.005197550635784864, + 0.0044919527135789394, + 0.008766697719693184, + 0.02221638523042202, + 0.01896917261183262, + 0.012058251537382603, + 0.011595597490668297, + 0.019661996513605118, + 0.008279956877231598, + 0.009707006625831127, + 0.0077530560083687305, + 0.006911600474268198, + 0.0037873825058341026, + 0.006376414559781551, + 0.010921935550868511, + 0.006658902857452631, + 0.0071251848712563515, + 0.0055738515220582485, + 0.08373676985502243, + 0.021983861923217773, + 0.055081598460674286, + 0.015116351656615734, + 0.016678376123309135, + 0.015873659402132034, + 0.007508024573326111, + 0.010510504245758057, + 0.004368064925074577, + 0.00278916722163558, + 0.004479049239307642, + 0.0037771768402308226, + 0.002496515866369009, + 0.002701446181163192, + 0.0024046776816248894, + 0.0024485657922923565, + 0.005181610584259033, + 0.0032300224993377924, + 0.004304386209696531, + 0.0035304701887071133, + 0.0019703656435012817, + 0.0020289162639528513, + 0.0019796874839812517, + 0.002285450929775834, + 0.09652242064476013, + 0.0147434426471591, + 0.10845674574375153, + 0.018148787319660187, + 0.015298595651984215, + 0.009858395904302597, + 0.014548812992870808, + 0.025938214734196663, + 0.005650978535413742, + 0.004590809345245361, + 0.005130809266120195, + 0.006579129956662655, + 0.004280746914446354, + 0.004006014671176672, + 0.0041168853640556335, + 0.0045577147975564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.020141830667853355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.5761520862579346 + ], + "opposite_logprob_tokens": [ + -0.8261520862579346 + ], + "logprob_reward": -0.5761520862579346, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761520862579346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10774382203817368, + 0.01789812743663788, + 0.04566924646496773, + 0.11495114862918854, + 0.028112804517149925, + 0.3091728389263153, + 0.009200850501656532, + 0.022819261997938156, + 0.3186880648136139, + 0.025743817910552025 + ], + "credit_map": [ + 0.02402806095778942, + 0.012615892104804516, + 0.010201279073953629, + 0.01718870922923088, + 0.014403563924133778, + 0.011027839034795761, + 0.009247532114386559, + 0.00903094932436943, + 0.0022330526262521744, + 0.002015406498685479, + 0.0013237923849374056, + 0.0038759384769946337, + 0.0017565703019499779, + 0.0023972359485924244, + 0.0026335706934332848, + 0.0016625610878691077, + 0.0039155916310846806, + 0.00672982819378376, + 0.00422885175794363, + 0.010747100226581097, + 0.004471499472856522, + 0.005182249005883932, + 0.006135975476354361, + 0.004258152097463608, + 0.008105059154331684, + 0.018813403323292732, + 0.017472892999649048, + 0.011847889982163906, + 0.011419006623327732, + 0.02081224136054516, + 0.010289760306477547, + 0.016190892085433006, + 0.003512926399707794, + 0.0026918649673461914, + 0.0014575410168617964, + 0.0040818131528794765, + 0.004579515662044287, + 0.0029666514601558447, + 0.004675358533859253, + 0.00414713378995657, + 0.1197500079870224, + 0.029654698446393013, + 0.07625064253807068, + 0.020916469395160675, + 0.022033214569091797, + 0.021286142989993095, + 0.008419271558523178, + 0.01086239144206047, + 0.001133710378780961, + 0.0011916628573089838, + 0.0016820915043354034, + 0.0017470621969550848, + 0.0009949234081432223, + 0.0007425848161801696, + 0.0009270673617720604, + 0.0007817476289346814, + 0.004331548232585192, + 0.0021783995907753706, + 0.0042498973198235035, + 0.003068806603550911, + 0.0018492121016606688, + 0.0022025650832802057, + 0.002652318449690938, + 0.002286514500156045, + 0.10450773686170578, + 0.01488830428570509, + 0.11743586510419846, + 0.018800610676407814, + 0.016240036115050316, + 0.009859380312263966, + 0.012443586252629757, + 0.024512553587555885, + 0.003132519545033574, + 0.002641452942043543, + 0.003925923723727465, + 0.0058799865655601025, + 0.0022326514590531588, + 0.0016578944632783532, + 0.003802310908213258, + 0.0024710800498723984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.5761520862579346 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.007627647835761309 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.007627647835761309, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627647835761309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09946934133768082, + 0.040798332542181015, + 0.14634016156196594, + 0.21223679184913635, + 0.03890025615692139, + 0.04965771362185478, + 0.024378051981329918, + 0.02255377732217312, + 0.30418726801872253, + 0.06147828698158264 + ], + "credit_map": [ + 0.02497658133506775, + 0.010897024534642696, + 0.010976724326610565, + 0.007336171809583902, + 0.013160375878214836, + 0.01291544921696186, + 0.00858400110155344, + 0.010623016394674778, + 0.0053333197720348835, + 0.004803275223821402, + 0.0033228914253413677, + 0.008512003347277641, + 0.004328311886638403, + 0.006233793683350086, + 0.003913629334419966, + 0.004351108800619841, + 0.011866867542266846, + 0.021992528811097145, + 0.008756817318499088, + 0.03307320922613144, + 0.02203981950879097, + 0.01665118709206581, + 0.017325876280665398, + 0.014633851125836372, + 0.01613110490143299, + 0.02155803143978119, + 0.017416121438145638, + 0.020742610096931458, + 0.023843979462981224, + 0.04000726342201233, + 0.023667695000767708, + 0.04887000098824501, + 0.00632719649001956, + 0.006705397740006447, + 0.004699450917541981, + 0.004776290617883205, + 0.006642728578299284, + 0.003080665832385421, + 0.003663800423964858, + 0.0030047248583287, + 0.015467451885342598, + 0.005627613980323076, + 0.010210467502474785, + 0.0034020731691271067, + 0.004450319334864616, + 0.005785392597317696, + 0.0018488822970539331, + 0.0028655168134719133, + 0.003772990545257926, + 0.0023365174420177937, + 0.003392141778022051, + 0.0030917166732251644, + 0.0036864904686808586, + 0.003108050674200058, + 0.002583672059699893, + 0.0024064749013632536, + 0.003776004072278738, + 0.0023190330248326063, + 0.002735544927418232, + 0.003682821523398161, + 0.0022561694495379925, + 0.0019733295775949955, + 0.00323248072527349, + 0.0025783947203308344, + 0.09846488386392593, + 0.014441811479628086, + 0.11435936391353607, + 0.018226895481348038, + 0.015586402267217636, + 0.009646362625062466, + 0.012853967025876045, + 0.020607583224773407, + 0.006266070529818535, + 0.00574432173743844, + 0.008826375007629395, + 0.010524582117795944, + 0.005401552189141512, + 0.005575255956500769, + 0.010200455784797668, + 0.00893967505544424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.007627647835761309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.4289819300174713 + ], + "opposite_logprob_tokens": [ + -1.053981900215149 + ], + "logprob_reward": -0.4289819300174713, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289819300174713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13139960169792175, + 0.030343079939484596, + 0.12733009457588196, + 0.1323775202035904, + 0.022165171802043915, + 0.09071916341781616, + 0.007439898326992989, + 0.022970663383603096, + 0.4173443913459778, + 0.017910398542881012 + ], + "credit_map": [ + 0.018107593059539795, + 0.012576485052704811, + 0.012169435620307922, + 0.03086111880838871, + 0.015325638465583324, + 0.01165778934955597, + 0.01135788019746542, + 0.019343668594956398, + 0.005124265793710947, + 0.00323179317638278, + 0.003769730217754841, + 0.00460100406780839, + 0.004198177717626095, + 0.0030466392636299133, + 0.0035870380233973265, + 0.0027844354044646025, + 0.009489263407886028, + 0.020051583647727966, + 0.0075925011187791824, + 0.027293197810649872, + 0.02136574313044548, + 0.01751171611249447, + 0.013660857453942299, + 0.010365243069827557, + 0.01057899184525013, + 0.013410697691142559, + 0.0122905932366848, + 0.01216681394726038, + 0.014044350944459438, + 0.026051845401525497, + 0.014156912453472614, + 0.02967733144760132, + 0.0033511300571262836, + 0.0020607009064406157, + 0.001433535828255117, + 0.002588920760899782, + 0.0034663486294448376, + 0.002662014914676547, + 0.002857037354260683, + 0.0037454853300005198, + 0.031615518033504486, + 0.00921439379453659, + 0.02041231282055378, + 0.008078089915215969, + 0.006946089211851358, + 0.0060945660807192326, + 0.0033181689213961363, + 0.0050400313921272755, + 0.0018881664145737886, + 0.0008942829445004463, + 0.0011670748936012387, + 0.0008003361290320754, + 0.0008556467946618795, + 0.0005882474360987544, + 0.0006527675432153046, + 0.0005933764041401446, + 0.0038543404079973698, + 0.0021946467459201813, + 0.0037483926862478256, + 0.003588504157960415, + 0.0019847233779728413, + 0.00193948729429394, + 0.0025380428414791822, + 0.003122529946267605, + 0.13806404173374176, + 0.01931861788034439, + 0.15874560177326202, + 0.024318624287843704, + 0.02169857732951641, + 0.012623877264559269, + 0.014292856678366661, + 0.02828221581876278, + 0.0030029285699129105, + 0.0019929951522499323, + 0.002611378440633416, + 0.0033972663804888725, + 0.0016812344547361135, + 0.0016029924154281616, + 0.0016591158928349614, + 0.001962487818673253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.4289819300174713 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy multi-axle transport truck painted a weathered matte dark gray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the truck accelerating from a standstill with its tires caked in mud and sending up a cloud of spray?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sprawling industrial complex with thick fog obscuring distant steel structures and pipelines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive turbine visible in the background with slowly turning blades emitting a low roar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature towering searchlights casting cold white beams that penetrate the fog and create moving spots 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": [ + -0.014194360002875328 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194360002875328, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194360002875328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24996523559093475, + 0.1898079514503479, + 0.04092962667346001, + 0.1513567417860031, + 0.022462982684373856, + 0.05332764983177185, + 0.024678459390997887, + 0.023380229249596596, + 0.2096295803785324, + 0.034461572766304016 + ], + "credit_map": [ + 0.044966112822294235, + 0.03168759122490883, + 0.024765552952885628, + 0.01793958619236946, + 0.03971582278609276, + 0.035320159047842026, + 0.02623935416340828, + 0.029331043362617493, + 0.017137371003627777, + 0.034821778535842896, + 0.014603104442358017, + 0.01517870370298624, + 0.036362044513225555, + 0.03971046209335327, + 0.013757351785898209, + 0.018237130716443062, + 0.0033154732082039118, + 0.00591587508097291, + 0.0036311536096036434, + 0.00882499385625124, + 0.005578024312853813, + 0.005254374351352453, + 0.005190310068428516, + 0.00321942032314837, + 0.012885498814284801, + 0.01561000943183899, + 0.011444083414971828, + 0.013759392313659191, + 0.016875097528100014, + 0.03108658641576767, + 0.016254663467407227, + 0.03344140574336052, + 0.0027546396013349295, + 0.0028285167645663023, + 0.002080135513097048, + 0.002144615398719907, + 0.0047479975037276745, + 0.0027070147916674614, + 0.00226790108717978, + 0.0029321610927581787, + 0.014767755754292011, + 0.007032323163002729, + 0.009207071736454964, + 0.004469671286642551, + 0.005161413922905922, + 0.006448192987591028, + 0.002546753268688917, + 0.0036944665480405092, + 0.004841439891606569, + 0.0022566565312445164, + 0.004078550264239311, + 0.002967010485008359, + 0.002990873297676444, + 0.0026015974581241608, + 0.0025912460405379534, + 0.0023510842584073544, + 0.00462664058431983, + 0.003138086060062051, + 0.0039183348417282104, + 0.0024387482553720474, + 0.002052083844318986, + 0.0025545216631144285, + 0.0023131759371608496, + 0.0023386350367218256, + 0.05962705984711647, + 0.013956798240542412, + 0.0715671107172966, + 0.01722090132534504, + 0.010991081595420837, + 0.00713180098682642, + 0.01454066950827837, + 0.014594131149351597, + 0.005591377150267363, + 0.004921140614897013, + 0.004793745931237936, + 0.005329973064363003, + 0.0034142022486776114, + 0.0024903113953769207, + 0.0038188747130334377, + 0.004101946018636227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.014194360002875328 + } + ], + "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/0000152/group_5/index_tensor([3719, 3719])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_tensor([3719, 3719])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..af6360578b5e621f77e91277b156771058d0ae5b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_tensor([3719, 3719])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, captured from a slightly upward angle, captured in slow motion and smoothly tilted downward to follow the subject's movements, focuses on a young boy, perhaps five or six, with short, shaggy brown hair. He wears a one-piece flannel pajamas printed with blue and green cartoon dinosaurs, his face radiating pure, unrestrained joy. Soft morning sunlight slants through a side window, framing him in a soft golden glow. Standing on a large bed covered in a thick, rumpled white duvet, he suddenly bends his knees to gather momentum, then leaps upward with all his might, his body unfurling in the air, arms outstretched for balance, like a bird poised for flight. As the camera slowly pans downward, his feet plant themselves firmly and precisely into the thick, beige, shag carpet below. The soft fibers instantly spread out at the impact, perfectly cushioning the force of his fall. The whole scene is painted in warm tones with soft light and shadow, creating a morning atmosphere full of childlike fun, warmth and carefreeness.", + "seed": 30284, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_3719_seed_30284_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": -0.013302308972924948, + "vlm_logprob": -0.013302308972924948, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.013302308972924948, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012257955968379974 + ], + "opposite_logprob_tokens": [ + -4.762258052825928 + ], + "logprob_reward": -0.012257955968379974, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012257955968379974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0641416385769844, + 0.01916695013642311, + 0.1411222368478775, + 0.033384598791599274, + 0.26919811964035034, + 0.18189524114131927, + 0.0421927310526371, + 0.07848668098449707, + 0.06553302705287933, + 0.10487872362136841 + ], + "credit_map": [ + 0.005010646767914295, + 0.007509025279432535, + 0.008212633430957794, + 0.0054795281030237675, + 0.011180736124515533, + 0.009548556059598923, + 0.007465757895261049, + 0.009734753519296646, + 0.0018684094538912177, + 0.003005936508998275, + 0.0027620939072221518, + 0.0019291085191071033, + 0.0025131951551884413, + 0.0025663047563284636, + 0.0021875458769500256, + 0.002334356075152755, + 0.009216979146003723, + 0.012289910577237606, + 0.011233718134462833, + 0.020740173757076263, + 0.020792676135897636, + 0.022646917030215263, + 0.029289595782756805, + 0.014912263490259647, + 0.003293541492894292, + 0.005103605799376965, + 0.005484842229634523, + 0.003737173741683364, + 0.004007941577583551, + 0.004657138139009476, + 0.003111408092081547, + 0.003988945856690407, + 0.008369358256459236, + 0.009315789677202702, + 0.10734311491250992, + 0.11804462969303131, + 0.00407031225040555, + 0.003903180593624711, + 0.00846719741821289, + 0.009684554301202297, + 0.01684466004371643, + 0.009408174082636833, + 0.06371171772480011, + 0.021379807963967323, + 0.0058786156587302685, + 0.006594108417630196, + 0.009547881782054901, + 0.04853029176592827, + 0.0066481358371675014, + 0.004901806358247995, + 0.00784347951412201, + 0.007224132772535086, + 0.003213103162124753, + 0.004075421951711178, + 0.002743176417425275, + 0.005543475039303303, + 0.01400550827383995, + 0.009465147741138935, + 0.01093219593167305, + 0.010702710598707199, + 0.009312689304351807, + 0.005864819977432489, + 0.008192173205316067, + 0.010011437349021435, + 0.006072794087231159, + 0.007122554816305637, + 0.014744178391993046, + 0.006268918979912996, + 0.007322157267481089, + 0.006476923357695341, + 0.009482093155384064, + 0.008043404668569565, + 0.006998295895755291, + 0.014114159159362316, + 0.018396291881799698, + 0.013151144608855247, + 0.012730863876640797, + 0.010074908845126629, + 0.01541866920888424, + 0.013994386419653893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.012257955968379974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018196435645222664 + ], + "opposite_logprob_tokens": [ + -4.0181965827941895 + ], + "logprob_reward": -0.018196435645222664, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018196435645222664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025442147627472878, + 0.019114162772893906, + 0.02816617675125599, + 0.04647012799978256, + 0.6203109622001648, + 0.13289743661880493, + 0.03377735987305641, + 0.024383939802646637, + 0.04162467643618584, + 0.027813050895929337 + ], + "credit_map": [ + 0.0025451702531427145, + 0.003560513723641634, + 0.002955663949251175, + 0.002673139562830329, + 0.0039353021420538425, + 0.003369057783856988, + 0.002434770343825221, + 0.003968528006225824, + 0.0014774914598092437, + 0.002692553913220763, + 0.002385804196819663, + 0.0020795357413589954, + 0.0036793805193156004, + 0.0025850823149085045, + 0.002160200383514166, + 0.0020541127305477858, + 0.0016695205122232437, + 0.003988613374531269, + 0.003023018129169941, + 0.003779240418225527, + 0.005086751189082861, + 0.003747560316696763, + 0.004906255751848221, + 0.001965214731171727, + 0.005057212430983782, + 0.004041725769639015, + 0.008898373693227768, + 0.005507655441761017, + 0.005598137620836496, + 0.006441074889153242, + 0.0053305551409721375, + 0.005595390219241381, + 0.017175329849123955, + 0.023282084614038467, + 0.23925870656967163, + 0.2793727219104767, + 0.008547878824174404, + 0.009243222884833813, + 0.020667359232902527, + 0.02276363968849182, + 0.016360512003302574, + 0.007396036293357611, + 0.055927351117134094, + 0.021923484280705452, + 0.0043759471736848354, + 0.005392910912632942, + 0.008672448806464672, + 0.012848739512264729, + 0.0033393031917512417, + 0.0033494564704596996, + 0.005737010855227709, + 0.0064799971878528595, + 0.003418495412915945, + 0.0031346359755843878, + 0.0037886512000113726, + 0.004529805853962898, + 0.004754018969833851, + 0.003173659322783351, + 0.003410256002098322, + 0.003093636827543378, + 0.002342858584597707, + 0.0019628105219453573, + 0.0021634025033563375, + 0.003483296139165759, + 0.003744810586795211, + 0.005004952661693096, + 0.006119654513895512, + 0.006234307307749987, + 0.005630346946418285, + 0.0038849744014441967, + 0.006237637717276812, + 0.004767992999404669, + 0.002129698172211647, + 0.004132050555199385, + 0.003251612652093172, + 0.003700074041262269, + 0.003245876170694828, + 0.0033225705847144127, + 0.0035575826186686754, + 0.004473584238439798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.018196435645222664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017120479606091976 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017120479606091976, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017120479606091976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028326351195573807, + 0.017883554100990295, + 0.013534332625567913, + 0.043364863842725754, + 0.07023618370294571, + 0.6144969463348389, + 0.09837725013494492, + 0.022915704175829887, + 0.03773212432861328, + 0.05313273146748543 + ], + "credit_map": [ + 0.0029084349516779184, + 0.0027038471307605505, + 0.004961916711181402, + 0.002753674052655697, + 0.00244935997761786, + 0.004683939274400473, + 0.0024779941886663437, + 0.005387185141444206, + 0.0022108363918960094, + 0.002157103968784213, + 0.0039010262116789818, + 0.0018495656549930573, + 0.0020988467149436474, + 0.0021942758467048407, + 0.0015101097524166107, + 0.001961786998435855, + 0.001836080919019878, + 0.0014266272773966193, + 0.002907051704823971, + 0.0013177454238757491, + 0.0018625096417963505, + 0.0016627871664240956, + 0.0015302174724638462, + 0.0009913123212754726, + 0.005572914611548185, + 0.004731709137558937, + 0.009897105395793915, + 0.004389290232211351, + 0.0032953794579952955, + 0.006531269755214453, + 0.003906344063580036, + 0.005040844902396202, + 0.0043703410774469376, + 0.005254014395177364, + 0.02142471633851528, + 0.022317437455058098, + 0.0038546198047697544, + 0.002563371090218425, + 0.004555979277938604, + 0.005895697977393866, + 0.09042692929506302, + 0.0525728277862072, + 0.14252348244190216, + 0.05399666354060173, + 0.026730064302682877, + 0.0412168875336647, + 0.049890290945768356, + 0.15713977813720703, + 0.021413031965494156, + 0.008985752239823341, + 0.014478951692581177, + 0.008655606769025326, + 0.005650573410093784, + 0.007940802723169327, + 0.006411449518054724, + 0.024841079488396645, + 0.0028518023900687695, + 0.004013547208160162, + 0.002919287420809269, + 0.0023998022079467773, + 0.0029853808227926493, + 0.0015806048177182674, + 0.003568536601960659, + 0.0025967408437281847, + 0.0033966549672186375, + 0.005292891059070826, + 0.006057019345462322, + 0.004681482911109924, + 0.0053110672160983086, + 0.0035423962399363518, + 0.006161765661090612, + 0.003288842970505357, + 0.004669709596782923, + 0.008310053497552872, + 0.007103299256414175, + 0.006655630189925432, + 0.007617872674018145, + 0.0056834290735423565, + 0.0071718706749379635, + 0.005920860916376114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0017120479606091976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07790009677410126, + 0.06369840353727341, + 0.06004217267036438, + 0.08018054813146591, + 0.41237515211105347, + 0.09175930172204971, + 0.10355789214372635, + 0.023414164781570435, + 0.03191635012626648, + 0.0551559142768383 + ], + "credit_map": [ + 0.007913492619991302, + 0.007187874987721443, + 0.018789364024996758, + 0.005784083157777786, + 0.007132087368518114, + 0.020823761820793152, + 0.004144309088587761, + 0.006125126499682665, + 0.008966499008238316, + 0.012023736722767353, + 0.010928834788501263, + 0.0064704567193984985, + 0.007187596056610346, + 0.006694345269352198, + 0.00573849817737937, + 0.0056884377263486385, + 0.005474317353218794, + 0.006660715211182833, + 0.008969706483185291, + 0.006844511721283197, + 0.00798763707280159, + 0.008176183328032494, + 0.010733983479440212, + 0.005195118486881256, + 0.014316248707473278, + 0.010826404206454754, + 0.012507005594670773, + 0.005794861353933811, + 0.006638654973357916, + 0.016102580353617668, + 0.004379788413643837, + 0.009615003131330013, + 0.011893424205482006, + 0.015325792133808136, + 0.1599004566669464, + 0.18441209197044373, + 0.005919344257563353, + 0.005582731682807207, + 0.013566331937909126, + 0.01577499322593212, + 0.012771629728376865, + 0.008330428041517735, + 0.02053181640803814, + 0.008024334907531738, + 0.005373572930693626, + 0.006771383341401815, + 0.007564701605588198, + 0.022391436621546745, + 0.025161484256386757, + 0.007212400436401367, + 0.016644150018692017, + 0.006765617989003658, + 0.004704720340669155, + 0.007700130809098482, + 0.005473756697028875, + 0.029895637184381485, + 0.0031505655497312546, + 0.0028127431869506836, + 0.004361388273537159, + 0.002208285266533494, + 0.0024664283264428377, + 0.0024074516259133816, + 0.002681462559849024, + 0.0033258413895964622, + 0.0035386630333960056, + 0.004258837550878525, + 0.007053246721625328, + 0.0030551038216799498, + 0.003886562306433916, + 0.003473730059340596, + 0.0032821500208228827, + 0.003368058940395713, + 0.004546206910163164, + 0.006353908218443394, + 0.011497762985527515, + 0.005805786233395338, + 0.006827327888458967, + 0.004302431829273701, + 0.009292751550674438, + 0.00652974471449852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.011074034497141838 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023271070793271065 + ], + "opposite_logprob_tokens": [ + -3.773271083831787 + ], + "logprob_reward": -0.023271070793271065, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023271070793271065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01671295240521431, + 0.009311942383646965, + 0.008104872889816761, + 0.024122506380081177, + 0.5545042753219604, + 0.23863255977630615, + 0.06934521347284317, + 0.03564174473285675, + 0.029940331354737282, + 0.013683576136827469 + ], + "credit_map": [ + 0.0014318968169391155, + 0.0016256276285275817, + 0.002755595138296485, + 0.0017727246740832925, + 0.0014881556853652, + 0.0030889627523720264, + 0.0016025553923100233, + 0.002947433851659298, + 0.0009314876515418291, + 0.0008680915925651789, + 0.0015259889187291265, + 0.001425537164323032, + 0.0012272641761228442, + 0.0010090244468301535, + 0.0013718055561184883, + 0.0009527432848699391, + 0.0006794890505261719, + 0.0009056298877112567, + 0.0012018483830615878, + 0.0009664244716987014, + 0.0015562811167910695, + 0.0012366373557597399, + 0.0009735789499245584, + 0.0005849833833053708, + 0.002988832537084818, + 0.002821072470396757, + 0.005575790535658598, + 0.002560307504609227, + 0.0017454092158004642, + 0.003604003693908453, + 0.0019655840005725622, + 0.002861505141481757, + 0.01556257251650095, + 0.020089970901608467, + 0.21595650911331177, + 0.24932464957237244, + 0.007619284093379974, + 0.007881307043135166, + 0.017564231529831886, + 0.02050570771098137, + 0.017202235758304596, + 0.009413173422217369, + 0.1043342798948288, + 0.03826521709561348, + 0.004035074729472399, + 0.008203241974115372, + 0.012129529379308224, + 0.045049816370010376, + 0.015320910140872002, + 0.005913691129535437, + 0.009950238279998302, + 0.007188253570348024, + 0.004658402409404516, + 0.004140432458370924, + 0.004492584150284529, + 0.017680708318948746, + 0.007122997660189867, + 0.004174967296421528, + 0.005423848517239094, + 0.0032873277086764574, + 0.002601762069389224, + 0.0032044872641563416, + 0.003953264560550451, + 0.005873090587556362, + 0.002817864064127207, + 0.0035339451860636473, + 0.004548168275505304, + 0.003172591095790267, + 0.0034383274614810944, + 0.0031479208264499903, + 0.006296321749687195, + 0.002985192695632577, + 0.0013190199388191104, + 0.0015682225348427892, + 0.0026136739179491997, + 0.0017975070513784885, + 0.0016118900384753942, + 0.0011186543852090836, + 0.0020193567033857107, + 0.0016352524980902672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.023271070793271065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from a slightly upward angle, captured in slow motion and smoothly tilted downward to follow the subject's movements, focuses on a young boy, perhaps five or six, with short, shaggy brown hair. He wears a one-piece flannel pajamas printed with blue and green cartoon dinosaurs, his face radiating pure, unrestrained joy. Soft morning sunlight slants through a side window, framing him in a soft golden glow. Standing on a large bed covered in a thick, rumpled white duvet, he suddenly bends his knees to gather momentum, then leaps upward with all his might, his body unfurling in the air, arms outstretched for balance, like a bird poised for flight. As the camera slowly pans downward, his feet plant themselves firmly and precisely into the thick, beige, shag carpet below. The soft fibers instantly spread out at the impact, perfectly cushioning the force of his fall. The whole scene is painted in warm tones with soft light and shadow, creating a morning atmosphere full of childlike fun, warmth and carefreeness.", + "seed": 30285, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_3719_seed_30285_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": -0.018691836670041085, + "vlm_logprob": -0.018691836670041085, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.018691836670041085, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024200046434998512 + ], + "opposite_logprob_tokens": [ + -3.899199962615967 + ], + "logprob_reward": -0.024200046434998512, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024200046434998512, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12066599726676941, + 0.045999981462955475, + 0.04189443215727806, + 0.3895522952079773, + 0.043804991990327835, + 0.12284407019615173, + 0.04663373529911041, + 0.04938804730772972, + 0.05391557142138481, + 0.08530087023973465 + ], + "credit_map": [ + 0.009479538537561893, + 0.014782102778553963, + 0.010729429312050343, + 0.01158104557543993, + 0.024668235331773758, + 0.019438927993178368, + 0.01219105813652277, + 0.017795654013752937, + 0.0037040533497929573, + 0.005783435422927141, + 0.00520358607172966, + 0.0050141941756010056, + 0.008151990361511707, + 0.008257156237959862, + 0.005113173741847277, + 0.00477239117026329, + 0.003811000846326351, + 0.004203577060252428, + 0.0047756279818713665, + 0.004543708171695471, + 0.0054406370036304, + 0.008738894946873188, + 0.005461980123072863, + 0.004919008817523718, + 0.050348758697509766, + 0.017356514930725098, + 0.17035596072673798, + 0.026010597124695778, + 0.04892003536224365, + 0.04826425015926361, + 0.013896829448640347, + 0.014399346895515919, + 0.008737877942621708, + 0.004865881986916065, + 0.0049819061532616615, + 0.007772916927933693, + 0.0036909643094986677, + 0.0038076015189290047, + 0.0046080658212304115, + 0.005339775234460831, + 0.012364685535430908, + 0.014184906147420406, + 0.022397683933377266, + 0.015862835571169853, + 0.012771861627697945, + 0.010076544247567654, + 0.015950651839375496, + 0.019234903156757355, + 0.0066998968832194805, + 0.00592719204723835, + 0.005431932397186756, + 0.006515674293041229, + 0.005709651857614517, + 0.004520002752542496, + 0.006250304635614157, + 0.005579075310379267, + 0.004879343789070845, + 0.006459394935518503, + 0.007818003185093403, + 0.005532766226679087, + 0.0056104715913534164, + 0.003906979691237211, + 0.006102450657635927, + 0.009078635834157467, + 0.004611853510141373, + 0.007630554959177971, + 0.0065779732540249825, + 0.007704183924943209, + 0.0058352770283818245, + 0.006184215657413006, + 0.007969476282596588, + 0.007402034942060709, + 0.007387200836092234, + 0.010742436163127422, + 0.01098704244941473, + 0.01732086017727852, + 0.01075894758105278, + 0.007460119668394327, + 0.011339953169226646, + 0.009304314851760864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.024200046434998512 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193624913692474 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193624913692474, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193624913692474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06309376657009125, + 0.05318009480834007, + 0.10561467707157135, + 0.14503826200962067, + 0.1118381917476654, + 0.0934348776936531, + 0.1314171850681305, + 0.13407817482948303, + 0.06781906634569168, + 0.09448563307523727 + ], + "credit_map": [ + 0.006019435822963715, + 0.007024712394922972, + 0.00888644065707922, + 0.010128242895007133, + 0.006364210043102503, + 0.01200444158166647, + 0.006319814827293158, + 0.006346466951072216, + 0.0053373402915894985, + 0.007069992832839489, + 0.005618534050881863, + 0.004751645028591156, + 0.009284976869821548, + 0.007191624958068132, + 0.005608167964965105, + 0.00831780955195427, + 0.00920853391289711, + 0.011130271479487419, + 0.020317304879426956, + 0.015905719250440598, + 0.013004191219806671, + 0.013239903375506401, + 0.009374921210110188, + 0.013433828949928284, + 0.011826497502624989, + 0.011095704510807991, + 0.04020630195736885, + 0.020304547622799873, + 0.013976014219224453, + 0.019701741635799408, + 0.014209871180355549, + 0.013717589899897575, + 0.0188453309237957, + 0.014516503550112247, + 0.015970448032021523, + 0.017999017611145973, + 0.007752616889774799, + 0.009842544794082642, + 0.013282235711812973, + 0.013629493303596973, + 0.009524286724627018, + 0.0068832822144031525, + 0.016793707385659218, + 0.011310664936900139, + 0.012247825041413307, + 0.006361578591167927, + 0.021916940808296204, + 0.008396594785153866, + 0.012865805998444557, + 0.014439355581998825, + 0.025185661390423775, + 0.012388186529278755, + 0.022486615926027298, + 0.008080701343715191, + 0.02072826959192753, + 0.015242595225572586, + 0.01482480950653553, + 0.012845713645219803, + 0.020530525594949722, + 0.014356628991663456, + 0.01797359809279442, + 0.012355448678135872, + 0.013810100965201855, + 0.027381349354982376, + 0.006085368804633617, + 0.01106185745447874, + 0.006613245699554682, + 0.0071710254997015, + 0.007324202451854944, + 0.011349077336490154, + 0.006121039856225252, + 0.0120932562276721, + 0.008801487274467945, + 0.011423102580010891, + 0.01217424962669611, + 0.015015168115496635, + 0.010187103413045406, + 0.009620518423616886, + 0.013498491607606411, + 0.013765506446361542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.018193624913692474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189953975379467 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003189953975379467, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189953975379467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06057899072766304, + 0.02865522913634777, + 0.0353592224419117, + 0.13631616532802582, + 0.06098293885588646, + 0.10938415676355362, + 0.11410273611545563, + 0.17413435876369476, + 0.07969848811626434, + 0.20078766345977783 + ], + "credit_map": [ + 0.0058735441416502, + 0.007816525176167488, + 0.009538562037050724, + 0.010693504475057125, + 0.006691462825983763, + 0.0075748455710709095, + 0.006777594331651926, + 0.005612957291305065, + 0.0026269410736858845, + 0.0031068725511431694, + 0.004333376418799162, + 0.0029358090832829475, + 0.00333555880934, + 0.0048005166463553905, + 0.0032706812489777803, + 0.004245472140610218, + 0.0038133251946419477, + 0.003753078868612647, + 0.00443683285266161, + 0.006096456665545702, + 0.004393923562020063, + 0.00490586506202817, + 0.0037764778826385736, + 0.004183266311883926, + 0.010072366334497929, + 0.00841879565268755, + 0.04069999232888222, + 0.012823606841266155, + 0.020421696826815605, + 0.02002418413758278, + 0.01455159205943346, + 0.00930392649024725, + 0.014647225849330425, + 0.0059515247121453285, + 0.009272131137549877, + 0.012456338852643967, + 0.003485872643068433, + 0.003386235097423196, + 0.00455790339037776, + 0.007225706707686186, + 0.010715224780142307, + 0.01389214489609003, + 0.01981176808476448, + 0.01959497667849064, + 0.011783372610807419, + 0.007989898324012756, + 0.01699632592499256, + 0.008600450120866299, + 0.018186388537287712, + 0.01359330303966999, + 0.020394539460539818, + 0.01663026586174965, + 0.01693202182650566, + 0.008386963978409767, + 0.009769003838300705, + 0.010210239328444004, + 0.023545175790786743, + 0.01675238087773323, + 0.025693396106362343, + 0.027030091732740402, + 0.016642989590764046, + 0.016455689445137978, + 0.02161887101829052, + 0.0263957716524601, + 0.007304369006305933, + 0.011862200684845448, + 0.009492396377027035, + 0.009151868522167206, + 0.011205436661839485, + 0.008061220869421959, + 0.013138774782419205, + 0.009482219815254211, + 0.018034329637885094, + 0.01882736012339592, + 0.027274657040834427, + 0.02538440190255642, + 0.022648505866527557, + 0.027459898963570595, + 0.025585023686289787, + 0.03557348623871803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.003189953975379467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369107097387314 + ], + "opposite_logprob_tokens": [ + -3.408691167831421 + ], + "logprob_reward": -0.03369107097387314, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369107097387314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0768590122461319, + 0.03268914669752121, + 0.030242254957556725, + 0.5142019987106323, + 0.05080648511648178, + 0.12835349142551422, + 0.10399900376796722, + 0.013491387479007244, + 0.023090220987796783, + 0.02626699022948742 + ], + "credit_map": [ + 0.010411455295979977, + 0.00720489164814353, + 0.00968772079795599, + 0.01653113402426243, + 0.01064819935709238, + 0.008950954303145409, + 0.007773678749799728, + 0.005650974344462156, + 0.0027671395801007748, + 0.003763972083106637, + 0.0035464426036924124, + 0.0037268998567014933, + 0.0062741548754274845, + 0.004545300267636776, + 0.004270296078175306, + 0.0037949413526803255, + 0.0025242844130843878, + 0.0029971504118293524, + 0.005077660083770752, + 0.0052275750786066055, + 0.0030451184138655663, + 0.0043099443428218365, + 0.003100716508924961, + 0.003959805238991976, + 0.06634710729122162, + 0.023371582850813866, + 0.22367222607135773, + 0.03580779582262039, + 0.0650680661201477, + 0.06273595988750458, + 0.018769582733511925, + 0.01842965930700302, + 0.005948812235146761, + 0.005539011210203171, + 0.009691070765256882, + 0.011230274103581905, + 0.0037244046106934547, + 0.0039619943127036095, + 0.004607159178704023, + 0.006103755906224251, + 0.014335229992866516, + 0.015210283920168877, + 0.02135983109474182, + 0.01631130278110504, + 0.015185799449682236, + 0.011287592351436615, + 0.02452266588807106, + 0.010140801779925823, + 0.014672601595520973, + 0.018956003710627556, + 0.016307026147842407, + 0.013517619110643864, + 0.013172762468457222, + 0.010504858568310738, + 0.007708832621574402, + 0.009159298613667488, + 0.0015987077495083213, + 0.001138464780524373, + 0.001974984770640731, + 0.0017053158953785896, + 0.0010567576391622424, + 0.0010760521981865168, + 0.001227635657414794, + 0.0037134692538529634, + 0.0016762823797762394, + 0.003700581844896078, + 0.0021514450199902058, + 0.0024743475951254368, + 0.002644711174070835, + 0.0036342148669064045, + 0.003329235129058361, + 0.003479400184005499, + 0.0031410837545990944, + 0.0029657070990651846, + 0.0032904690597206354, + 0.0031253863126039505, + 0.0029253524262458086, + 0.0031643465626984835, + 0.003418570850044489, + 0.004236073233187199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.03369107097387314 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418448705226183 + ], + "opposite_logprob_tokens": [ + -4.264184474945068 + ], + "logprob_reward": -0.01418448705226183, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418448705226183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039042890071868896, + 0.026226487010717392, + 0.039378128945827484, + 0.27793964743614197, + 0.07121194899082184, + 0.1891980618238449, + 0.07717479020357132, + 0.0423390194773674, + 0.18569156527519226, + 0.05179739370942116 + ], + "credit_map": [ + 0.0069479309022426605, + 0.004084267187863588, + 0.0061485557816922665, + 0.006133818533271551, + 0.004442192148417234, + 0.004205118399113417, + 0.0035599092952907085, + 0.003521102713420987, + 0.002637835219502449, + 0.0036073338706046343, + 0.004801495466381311, + 0.0032501090317964554, + 0.003342628013342619, + 0.0033072081860154867, + 0.0027910068165510893, + 0.00248887506313622, + 0.003597449976950884, + 0.004407270811498165, + 0.008293245919048786, + 0.006848211400210857, + 0.004934932570904493, + 0.004434745758771896, + 0.0032271856907755136, + 0.003635096363723278, + 0.03580824285745621, + 0.012943319976329803, + 0.12117183208465576, + 0.018335139378905296, + 0.03529801219701767, + 0.03393922746181488, + 0.010195272974669933, + 0.010248641483485699, + 0.009616291150450706, + 0.007271090988069773, + 0.008711322210729122, + 0.018899215385317802, + 0.006005495321005583, + 0.004565883427858353, + 0.006594967562705278, + 0.009547694586217403, + 0.02073265053331852, + 0.024328140541911125, + 0.030688654631376266, + 0.02240411937236786, + 0.019534636288881302, + 0.014542479068040848, + 0.03404508903622627, + 0.022922316566109657, + 0.00877845473587513, + 0.011156417429447174, + 0.00961293000727892, + 0.010053176432847977, + 0.00924639217555523, + 0.009724454022943974, + 0.010505360551178455, + 0.008097618818283081, + 0.004500983748584986, + 0.0055693225003778934, + 0.004374229349195957, + 0.00683447765186429, + 0.004826266318559647, + 0.00395984435454011, + 0.005872706416994333, + 0.006401198450475931, + 0.0147852199152112, + 0.029430828988552094, + 0.03047029674053192, + 0.019603142514824867, + 0.019041065126657486, + 0.016009733080863953, + 0.03298424929380417, + 0.023367058485746384, + 0.004915022756904364, + 0.007362079806625843, + 0.005303980316966772, + 0.0060594105161726475, + 0.01002214103937149, + 0.0037644864059984684, + 0.006794124376028776, + 0.007576152216643095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.01418448705226183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from a slightly upward angle, captured in slow motion and smoothly tilted downward to follow the subject's movements, focuses on a young boy, perhaps five or six, with short, shaggy brown hair. He wears a one-piece flannel pajamas printed with blue and green cartoon dinosaurs, his face radiating pure, unrestrained joy. Soft morning sunlight slants through a side window, framing him in a soft golden glow. Standing on a large bed covered in a thick, rumpled white duvet, he suddenly bends his knees to gather momentum, then leaps upward with all his might, his body unfurling in the air, arms outstretched for balance, like a bird poised for flight. As the camera slowly pans downward, his feet plant themselves firmly and precisely into the thick, beige, shag carpet below. The soft fibers instantly spread out at the impact, perfectly cushioning the force of his fall. The whole scene is painted in warm tones with soft light and shadow, creating a morning atmosphere full of childlike fun, warmth and carefreeness.", + "seed": 30286, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_3719_seed_30286_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": -0.014978359732776881, + "vlm_logprob": -0.014978359732776881, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.014978359732776881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011724492534995079 + ], + "opposite_logprob_tokens": [ + -4.886724472045898 + ], + "logprob_reward": -0.011724492534995079, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011724492534995079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23686738312244415, + 0.09888369590044022, + 0.048004116863012314, + 0.038105081766843796, + 0.07102105021476746, + 0.10762396454811096, + 0.21113640069961548, + 0.08368321508169174, + 0.05093960091471672, + 0.05373554304242134 + ], + "credit_map": [ + 0.012706768698990345, + 0.014214974828064442, + 0.045274656265974045, + 0.014488769695162773, + 0.04356642812490463, + 0.06148718670010567, + 0.021266592666506767, + 0.02386201173067093, + 0.00663557555526495, + 0.004189025145024061, + 0.023366032168269157, + 0.005828621797263622, + 0.02367771975696087, + 0.02101496420800686, + 0.007971220649778843, + 0.006200536619871855, + 0.006645782385021448, + 0.005291182082146406, + 0.007587339263409376, + 0.005544336512684822, + 0.008844826370477676, + 0.006354312878102064, + 0.004354397766292095, + 0.003381941467523575, + 0.006076735910028219, + 0.004157896153628826, + 0.0059236325323581696, + 0.0029660044237971306, + 0.0064916955307126045, + 0.0037996540777385235, + 0.003977725747972727, + 0.004711735527962446, + 0.013677941635251045, + 0.005862149875611067, + 0.011912780813872814, + 0.008369804359972477, + 0.00795933697372675, + 0.007651293650269508, + 0.009470371529459953, + 0.006117370445281267, + 0.01640532910823822, + 0.014711215160787106, + 0.017730850726366043, + 0.012085488066077232, + 0.010643786750733852, + 0.008985121734440327, + 0.01107574999332428, + 0.015986420214176178, + 0.029881082475185394, + 0.035220734775066376, + 0.026971185579895973, + 0.03025498427450657, + 0.02717369981110096, + 0.018779462203383446, + 0.02623406983911991, + 0.016621174290776253, + 0.011078729294240475, + 0.011830977164208889, + 0.009782318957149982, + 0.015455737709999084, + 0.00944463163614273, + 0.006087334360927343, + 0.011188028380274773, + 0.008815453387796879, + 0.006850828882306814, + 0.006393864285200834, + 0.007244849111884832, + 0.005411455407738686, + 0.011153312399983406, + 0.0034505499061197042, + 0.00432573864236474, + 0.0061090001836419106, + 0.0064840251579880714, + 0.005660162772983313, + 0.010698282159864902, + 0.005884561687707901, + 0.007525533437728882, + 0.005500289611518383, + 0.004400913137942553, + 0.007581775542348623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.011724492534995079 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345614343881607 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.026345614343881607, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345614343881607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09056881070137024, + 0.145143061876297, + 0.08135873079299927, + 0.06521616131067276, + 0.06172109395265579, + 0.10360266268253326, + 0.22434981167316437, + 0.10607325285673141, + 0.052167393267154694, + 0.0697990134358406 + ], + "credit_map": [ + 0.0042050667107105255, + 0.006768195424228907, + 0.018253915011882782, + 0.011280438862740993, + 0.014889712445437908, + 0.016209082677960396, + 0.01023055799305439, + 0.008731843903660774, + 0.009941181167960167, + 0.006125079933553934, + 0.03387495502829552, + 0.008496277034282684, + 0.033488672226667404, + 0.03121505305171013, + 0.011944456957280636, + 0.010057386010885239, + 0.0053901467472314835, + 0.006952742114663124, + 0.015516706742346287, + 0.0073745474219322205, + 0.021425005048513412, + 0.011850444599986076, + 0.008193650282919407, + 0.004655493423342705, + 0.009331484325230122, + 0.0061556631699204445, + 0.009327540174126625, + 0.0072077009826898575, + 0.010595536790788174, + 0.007656127214431763, + 0.007403277326375246, + 0.0075388322584331036, + 0.013724276795983315, + 0.005685156676918268, + 0.010532758198678493, + 0.008740173652768135, + 0.006418108008801937, + 0.0045411475002765656, + 0.006393732037395239, + 0.005685742478817701, + 0.018703995272517204, + 0.012507508508861065, + 0.01584651693701744, + 0.011614208109676838, + 0.010522279888391495, + 0.011458499357104301, + 0.008496616967022419, + 0.014453033916652203, + 0.02941564843058586, + 0.037504587322473526, + 0.028596188873052597, + 0.03249729797244072, + 0.028422242030501366, + 0.02119583450257778, + 0.02768636867403984, + 0.019031651318073273, + 0.014370488002896309, + 0.018584055826067924, + 0.013780864886939526, + 0.014658436179161072, + 0.010222884826362133, + 0.008219328708946705, + 0.01436245534569025, + 0.011874736286699772, + 0.0072737038135528564, + 0.006933110300451517, + 0.006422697100788355, + 0.004998999182134867, + 0.010320399887859821, + 0.003946386277675629, + 0.007906123995780945, + 0.00436597503721714, + 0.007298128213733435, + 0.009057366289198399, + 0.009671514853835106, + 0.0075410413555800915, + 0.010838542133569717, + 0.007204283494502306, + 0.006031673867255449, + 0.012156466953456402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.026345614343881607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076214964501559734 + ], + "opposite_logprob_tokens": [ + -4.8826212882995605 + ], + "logprob_reward": -0.0076214964501559734, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076214964501559734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0931583121418953, + 0.05466435104608536, + 0.1363724023103714, + 0.0554356686770916, + 0.06724151223897934, + 0.10182510316371918, + 0.17059388756752014, + 0.07592815905809402, + 0.14158815145492554, + 0.10319250077009201 + ], + "credit_map": [ + 0.007220392115414143, + 0.009859594516456127, + 0.015783173963427544, + 0.010642874985933304, + 0.012609361670911312, + 0.018019087612628937, + 0.0065150754526257515, + 0.012508736923336983, + 0.0043485104106366634, + 0.006307776086032391, + 0.009784948080778122, + 0.00699602859094739, + 0.00626908615231514, + 0.010059458203613758, + 0.0058573330752551556, + 0.005041202995926142, + 0.010440412908792496, + 0.007937425747513771, + 0.02370629459619522, + 0.015062050893902779, + 0.029901092872023582, + 0.01949012465775013, + 0.017698097974061966, + 0.012136894278228283, + 0.008814462460577488, + 0.0067840926349163055, + 0.008165238425135612, + 0.004830401390790939, + 0.005735740531235933, + 0.00784968864172697, + 0.005624115001410246, + 0.007631927262991667, + 0.0136009082198143, + 0.006274614483118057, + 0.010784642770886421, + 0.006970077753067017, + 0.007454116363078356, + 0.007186736911535263, + 0.007583021651953459, + 0.007387394551187754, + 0.017981501296162605, + 0.012459521181881428, + 0.016943147405982018, + 0.011439822614192963, + 0.009001905098557472, + 0.009350921027362347, + 0.00877510104328394, + 0.015873176977038383, + 0.02585388533771038, + 0.027270283550024033, + 0.02167069911956787, + 0.022691166028380394, + 0.018434729427099228, + 0.019699210301041603, + 0.017686497420072556, + 0.01728740707039833, + 0.009924756363034248, + 0.010364309884607792, + 0.009457957930862904, + 0.014767291955649853, + 0.006611302960664034, + 0.005975541193038225, + 0.008281808346509933, + 0.010545190423727036, + 0.03334686532616615, + 0.01402800902724266, + 0.02205418050289154, + 0.015255318023264408, + 0.013081008568406105, + 0.013171568512916565, + 0.011748773977160454, + 0.018902428448200226, + 0.013331763446331024, + 0.010800962336361408, + 0.01774631440639496, + 0.011146791279315948, + 0.015273910015821457, + 0.012380831874907017, + 0.008297102525830269, + 0.014214812777936459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.0076214964501559734 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008632485754787922 + ], + "opposite_logprob_tokens": [ + -4.758632659912109 + ], + "logprob_reward": -0.008632485754787922, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008632485754787922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21869617700576782, + 0.05353648215532303, + 0.21875572204589844, + 0.06482339650392532, + 0.03839946165680885, + 0.09143730998039246, + 0.17472419142723083, + 0.03975048288702965, + 0.04660366475582123, + 0.05327310785651207 + ], + "credit_map": [ + 0.01880318485200405, + 0.02396857738494873, + 0.03727467358112335, + 0.025735555216670036, + 0.026050249114632607, + 0.030862240120768547, + 0.017784416675567627, + 0.038217298686504364, + 0.003916723653674126, + 0.004527289420366287, + 0.010547596961259842, + 0.005211707204580307, + 0.009168523363769054, + 0.010166619904339314, + 0.004761622752994299, + 0.0052363988943398, + 0.022593168541789055, + 0.024755651131272316, + 0.039299167692661285, + 0.019366981461644173, + 0.05424555763602257, + 0.02880174107849598, + 0.016550051048398018, + 0.01314341276884079, + 0.010399973951280117, + 0.007652717642486095, + 0.008883505128324032, + 0.006168489344418049, + 0.006519924849271774, + 0.008012055419385433, + 0.008271320722997189, + 0.008915414102375507, + 0.009581120684742928, + 0.003438126528635621, + 0.006428774446249008, + 0.0035612713545560837, + 0.0043391636572778225, + 0.0030065360479056835, + 0.004779292270541191, + 0.0032651808578521013, + 0.013786949217319489, + 0.012911807745695114, + 0.015118500217795372, + 0.011610512621700764, + 0.009444286115467548, + 0.00750085199251771, + 0.009096933528780937, + 0.011967476457357407, + 0.02506544627249241, + 0.028570998460054398, + 0.023605260998010635, + 0.02444741316139698, + 0.022293848916888237, + 0.013978534378111362, + 0.02258405275642872, + 0.014178654178977013, + 0.005210528615862131, + 0.007792946882545948, + 0.0038960252422839403, + 0.004651645664125681, + 0.007926316000521183, + 0.0019105912651866674, + 0.004752467852085829, + 0.0036099606659263372, + 0.006269768346101046, + 0.00463595287874341, + 0.00794086791574955, + 0.005838797893375158, + 0.006960359402000904, + 0.003915435168892145, + 0.005312028340995312, + 0.005730461794883013, + 0.00515005923807621, + 0.003490029601380229, + 0.013941850513219833, + 0.004479753784835339, + 0.003909293562173843, + 0.005880679003894329, + 0.00409006979316473, + 0.012331374920904636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.008632485754787922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056770958006382 + ], + "opposite_logprob_tokens": [ + -3.8955676555633545 + ], + "logprob_reward": -0.02056770958006382, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056770958006382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09730277210474014, + 0.07439298927783966, + 0.08220914006233215, + 0.07884042710065842, + 0.14480474591255188, + 0.0707237720489502, + 0.1384134441614151, + 0.1277136504650116, + 0.06287693977355957, + 0.12272211164236069 + ], + "credit_map": [ + 0.006328203249722719, + 0.007104961667209864, + 0.016940034925937653, + 0.009448603726923466, + 0.01744362711906433, + 0.018213210627436638, + 0.011468109674751759, + 0.010356036014854908, + 0.00473671592772007, + 0.004691512323915958, + 0.015067831613123417, + 0.0061010317876935005, + 0.014075488783419132, + 0.015493687242269516, + 0.007506339345127344, + 0.006720387376844883, + 0.005919741932302713, + 0.004667357075959444, + 0.014212729409337044, + 0.008690432645380497, + 0.018261777237057686, + 0.010926405899226665, + 0.011520735919475555, + 0.008009969256818295, + 0.009720643982291222, + 0.006200023926794529, + 0.013229165226221085, + 0.006426055449992418, + 0.011550010181963444, + 0.011222921311855316, + 0.008900762535631657, + 0.011590848676860332, + 0.042661748826503754, + 0.012800208292901516, + 0.024186965078115463, + 0.01664552465081215, + 0.012730607762932777, + 0.013001459650695324, + 0.009361459873616695, + 0.013416796922683716, + 0.01152880396693945, + 0.00870444718748331, + 0.011776290833950043, + 0.008480418473482132, + 0.007132259197533131, + 0.005711019039154053, + 0.007070687599480152, + 0.010319853201508522, + 0.02088230475783348, + 0.023499146103858948, + 0.019143402576446533, + 0.018809933215379715, + 0.01739628054201603, + 0.008715806528925896, + 0.017958668991923332, + 0.012007907032966614, + 0.016038596630096436, + 0.02172497659921646, + 0.014345197938382626, + 0.019183797761797905, + 0.025814030319452286, + 0.005904189310967922, + 0.015474908985197544, + 0.009227961301803589, + 0.014972815290093422, + 0.006841233000159264, + 0.009012862108647823, + 0.005639714188873768, + 0.006472451612353325, + 0.007507931441068649, + 0.005159902852028608, + 0.007270037196576595, + 0.014614355750381947, + 0.01029905304312706, + 0.026470860466361046, + 0.011332870461046696, + 0.010810527950525284, + 0.017582818865776062, + 0.010719026438891888, + 0.02089259773492813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.02056770958006382 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, captured from a slightly upward angle, captured in slow motion and smoothly tilted downward to follow the subject's movements, focuses on a young boy, perhaps five or six, with short, shaggy brown hair. He wears a one-piece flannel pajamas printed with blue and green cartoon dinosaurs, his face radiating pure, unrestrained joy. Soft morning sunlight slants through a side window, framing him in a soft golden glow. Standing on a large bed covered in a thick, rumpled white duvet, he suddenly bends his knees to gather momentum, then leaps upward with all his might, his body unfurling in the air, arms outstretched for balance, like a bird poised for flight. As the camera slowly pans downward, his feet plant themselves firmly and precisely into the thick, beige, shag carpet below. The soft fibers instantly spread out at the impact, perfectly cushioning the force of his fall. The whole scene is painted in warm tones with soft light and shadow, creating a morning atmosphere full of childlike fun, warmth and carefreeness.", + "seed": 30287, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_5/index_3719_seed_30287_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": -0.01126892191823572, + "vlm_logprob": -0.01126892191823572, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.01126892191823572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012272086925804615 + ], + "opposite_logprob_tokens": [ + -4.762271881103516 + ], + "logprob_reward": -0.012272086925804615, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.012272086925804615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2752074599266052, + 0.03734768554568291, + 0.0776505395770073, + 0.09343985468149185, + 0.056314803659915924, + 0.04928191006183624, + 0.08913446217775345, + 0.053757473826408386, + 0.20914961397647858, + 0.058716244995594025 + ], + "credit_map": [ + 0.052186768501996994, + 0.07087018340826035, + 0.023619797080755234, + 0.022869940847158432, + 0.028752077370882034, + 0.02516707219183445, + 0.028985049575567245, + 0.02275654301047325, + 0.004018559120595455, + 0.004573349375277758, + 0.003614183049649, + 0.010457130149006844, + 0.004409676883369684, + 0.0025693955831229687, + 0.004990075249224901, + 0.0027153107803314924, + 0.01011915598064661, + 0.016010191291570663, + 0.011779717169702053, + 0.009943393059074879, + 0.008075721561908722, + 0.007754527032375336, + 0.006585114169865847, + 0.007382706273347139, + 0.008561994880437851, + 0.009932243265211582, + 0.009304055012762547, + 0.01360731478780508, + 0.023692317306995392, + 0.011931788176298141, + 0.00986352376639843, + 0.0065466067753732204, + 0.00380751327611506, + 0.005323308520019054, + 0.006535778287798166, + 0.01054973155260086, + 0.008618033491075039, + 0.005130094941705465, + 0.006473354063928127, + 0.009876981377601624, + 0.00634772190824151, + 0.004701443947851658, + 0.007798765320330858, + 0.008188673295080662, + 0.0033464431762695312, + 0.0062636807560920715, + 0.00627499632537365, + 0.006360183469951153, + 0.007460706401616335, + 0.010730687528848648, + 0.013568010181188583, + 0.01780419982969761, + 0.007903743535280228, + 0.008313995786011219, + 0.009804251603782177, + 0.013548857532441616, + 0.006480294745415449, + 0.004380662925541401, + 0.005805023945868015, + 0.008138167671859264, + 0.01064604427665472, + 0.004720099736005068, + 0.006689718924462795, + 0.006897452753037214, + 0.015245524235069752, + 0.01468921359628439, + 0.060152750462293625, + 0.06945744901895523, + 0.016281412914395332, + 0.011255836114287376, + 0.01352656353265047, + 0.008540843613445759, + 0.00649956613779068, + 0.0077613829635083675, + 0.008119138889014721, + 0.011439287103712559, + 0.007627767976373434, + 0.003886271733790636, + 0.00769934244453907, + 0.005683484021574259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.012272086925804615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550893239676952 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550893239676952, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550893239676952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0944581925868988, + 0.05536746606230736, + 0.06845841556787491, + 0.04304957389831543, + 0.22002924978733063, + 0.06609240919351578, + 0.19460804760456085, + 0.12225445359945297, + 0.09253867715597153, + 0.04314344376325607 + ], + "credit_map": [ + 0.007989025674760342, + 0.011570489965379238, + 0.012639558874070644, + 0.012809302657842636, + 0.013162669725716114, + 0.009494772180914879, + 0.01599535532295704, + 0.010797020979225636, + 0.005686338525265455, + 0.0078088585287332535, + 0.009196695871651173, + 0.011937413364648819, + 0.004926879424601793, + 0.0041610910557210445, + 0.006296920124441385, + 0.005353280808776617, + 0.00637675728648901, + 0.0114869000390172, + 0.008832577615976334, + 0.008279192261397839, + 0.007688314653933048, + 0.012332488782703876, + 0.0059722852893173695, + 0.007489911280572414, + 0.004110011737793684, + 0.008308108896017075, + 0.005866244900971651, + 0.005404547788202763, + 0.005511363036930561, + 0.006468259729444981, + 0.004112190566956997, + 0.003268853761255741, + 0.014311530627310276, + 0.02067447453737259, + 0.024553174152970314, + 0.03187577798962593, + 0.03529072180390358, + 0.03483704477548599, + 0.023933954536914825, + 0.03455260768532753, + 0.006037525832653046, + 0.00568707101047039, + 0.00949792005121708, + 0.011834132485091686, + 0.006798735819756985, + 0.007110861130058765, + 0.0093552740290761, + 0.009770899079740047, + 0.017555350437760353, + 0.026133382692933083, + 0.02864609658718109, + 0.04417833313345909, + 0.01728019118309021, + 0.014958877116441727, + 0.01914268173277378, + 0.026713166385889053, + 0.014041575603187084, + 0.01673012413084507, + 0.016589581966400146, + 0.021739041432738304, + 0.01464988011866808, + 0.011576896533370018, + 0.014404483139514923, + 0.01252288930118084, + 0.009535153396427631, + 0.0063890451565384865, + 0.024141093716025352, + 0.020612213760614395, + 0.007805653382092714, + 0.007971462793648243, + 0.0071848356164991856, + 0.008899222128093243, + 0.0053893690928816795, + 0.006970839574933052, + 0.007166050374507904, + 0.006346899550408125, + 0.006590342614799738, + 0.0031377223785966635, + 0.0035534799098968506, + 0.003988746553659439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.012550893239676952 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008120815036818385 + ], + "opposite_logprob_tokens": [ + -7.12581205368042 + ], + "logprob_reward": -0.0008120815036818385, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008120815036818385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10251913219690323, + 0.034069061279296875, + 0.10952868312597275, + 0.040248654782772064, + 0.13481682538986206, + 0.11536022275686264, + 0.0812777504324913, + 0.19510754942893982, + 0.1292489916086197, + 0.05782308056950569 + ], + "credit_map": [ + 0.01815754361450672, + 0.026217158883810043, + 0.010274249128997326, + 0.012573783285915852, + 0.0082057761028409, + 0.010438963770866394, + 0.007210397627204657, + 0.009441276080906391, + 0.003575933864340186, + 0.005246338900178671, + 0.0040191467851400375, + 0.007621261291205883, + 0.004134565591812134, + 0.0021179798059165478, + 0.0044313110411167145, + 0.0029225265607237816, + 0.00812085997313261, + 0.015088184736669064, + 0.016793573275208473, + 0.013176330365240574, + 0.0143725061789155, + 0.016293294727802277, + 0.01334968488663435, + 0.01233425922691822, + 0.005141906905919313, + 0.003779397113248706, + 0.0051546646282076836, + 0.008315932005643845, + 0.006085286382585764, + 0.004076735582202673, + 0.004595379810780287, + 0.003099355148151517, + 0.010546396486461163, + 0.013267111033201218, + 0.01658117026090622, + 0.02630789205431938, + 0.011769945733249187, + 0.016627805307507515, + 0.0178325604647398, + 0.021883947774767876, + 0.016729017719626427, + 0.009300004690885544, + 0.016210336238145828, + 0.02158006653189659, + 0.006680512800812721, + 0.011917466297745705, + 0.013825762085616589, + 0.019117066636681557, + 0.00890564639121294, + 0.00901995413005352, + 0.010358783416450024, + 0.01095583289861679, + 0.01091672107577324, + 0.011983036063611507, + 0.010380305349826813, + 0.008757478557527065, + 0.03789142146706581, + 0.03235664591193199, + 0.026838721707463264, + 0.02443688176572323, + 0.014320648275315762, + 0.02635456807911396, + 0.01909482665359974, + 0.013813858851790428, + 0.012636476196348667, + 0.010762174613773823, + 0.03532455489039421, + 0.03913888707756996, + 0.007790958508849144, + 0.009505863301455975, + 0.007815477438271046, + 0.006274609360843897, + 0.007931221276521683, + 0.007483608555048704, + 0.009587865322828293, + 0.011026425287127495, + 0.006154032424092293, + 0.003493987489491701, + 0.005421459209173918, + 0.006724483799189329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.0008120815036818385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535588815808296 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535588815808296, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535588815808296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10147707909345627, + 0.118629090487957, + 0.35893723368644714, + 0.1134076938033104, + 0.10045555979013443, + 0.036408234387636185, + 0.034046608954668045, + 0.03974492475390434, + 0.07691607624292374, + 0.019977519288659096 + ], + "credit_map": [ + 0.010471581481397152, + 0.011488609947264194, + 0.017007632181048393, + 0.018819380551576614, + 0.010406538844108582, + 0.011983450502157211, + 0.008173827081918716, + 0.013126054778695107, + 0.014900625683367252, + 0.028842639178037643, + 0.017668679356575012, + 0.025861840695142746, + 0.007834751158952713, + 0.009406221099197865, + 0.007144531235098839, + 0.006969792768359184, + 0.02641124278306961, + 0.050428666174411774, + 0.04419367387890816, + 0.04484245181083679, + 0.045140210539102554, + 0.059377651661634445, + 0.045931000262498856, + 0.04261229932308197, + 0.021809808909893036, + 0.012836105190217495, + 0.019564419984817505, + 0.0151442252099514, + 0.017619213089346886, + 0.011780617758631706, + 0.0065907686948776245, + 0.00806252658367157, + 0.008043593727052212, + 0.009532849304378033, + 0.011658262461423874, + 0.01542187575250864, + 0.008709832094609737, + 0.01781558431684971, + 0.011827344074845314, + 0.017446214333176613, + 0.003986984025686979, + 0.0034241604153066874, + 0.006191490683704615, + 0.007239319384098053, + 0.002179342322051525, + 0.004168421030044556, + 0.0031197944190353155, + 0.006098717451095581, + 0.004156431183218956, + 0.003913133405148983, + 0.005991724319756031, + 0.005365133751183748, + 0.003762083128094673, + 0.003018575254827738, + 0.004454927984625101, + 0.003384598297998309, + 0.006615058984607458, + 0.00351517996750772, + 0.005425561685115099, + 0.007068724371492863, + 0.0034225189592689276, + 0.004015959333628416, + 0.0036129974760115147, + 0.006068918853998184, + 0.005983424838632345, + 0.0057752542197704315, + 0.023271437734365463, + 0.026946064084768295, + 0.0027192500419914722, + 0.004127849824726582, + 0.004455313552170992, + 0.0036374791525304317, + 0.0030262181535363197, + 0.003021997632458806, + 0.002700203098356724, + 0.002972713904455304, + 0.0026989562902599573, + 0.0013720807619392872, + 0.0020968595054000616, + 0.002088487846776843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.012535588815808296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 young boy with short, shaggy brown hair wearing one-piece flannel pajamas with blue and green cartoon dinosaurs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the boy start by bending his knees to gather momentum and then leap upward with arms outstretched?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the boy land on a thick, beige, shag carpet where the fibers spread out upon impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by soft morning sunlight slanting through a side window, creating a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a slightly upward angle and smoothly tilt downward to follow the boy's movements?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173959106206894 + ], + "opposite_logprob_tokens": [ + -4.018174171447754 + ], + "logprob_reward": -0.018173959106206894, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018173959106206894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09048863500356674, + 0.03601071611046791, + 0.027842842042446136, + 0.07312927395105362, + 0.05101962015032768, + 0.10360962152481079, + 0.2048605978488922, + 0.1911180317401886, + 0.16745126247406006, + 0.05446944013237953 + ], + "credit_map": [ + 0.020962705835700035, + 0.028233295306563377, + 0.009217261336743832, + 0.007965095341205597, + 0.007798223756253719, + 0.0058038244023919106, + 0.005282332189381123, + 0.005225885659456253, + 0.0037195647601038218, + 0.004738334566354752, + 0.004337006248533726, + 0.007843047380447388, + 0.005437497980892658, + 0.0036846427246928215, + 0.0032253216486424208, + 0.0030252982396632433, + 0.003034968627616763, + 0.0046334038488566875, + 0.003499936079606414, + 0.003082436043769121, + 0.004945943132042885, + 0.0029335403814911842, + 0.003251616843044758, + 0.0024609919637441635, + 0.008120077662169933, + 0.009849203750491142, + 0.008422723039984703, + 0.013366211205720901, + 0.009692089632153511, + 0.011108306236565113, + 0.006388029083609581, + 0.00618262542411685, + 0.0034533794969320297, + 0.005925700068473816, + 0.005274731200188398, + 0.00835544615983963, + 0.006682019215077162, + 0.005568378604948521, + 0.008190124295651913, + 0.007569840177893639, + 0.01301049068570137, + 0.007970063015818596, + 0.015348769724369049, + 0.019976399838924408, + 0.009540573693811893, + 0.011419590562582016, + 0.011915525421500206, + 0.014428192749619484, + 0.02715541422367096, + 0.04382462427020073, + 0.02517707832157612, + 0.03405966982245445, + 0.018613776192069054, + 0.01956191286444664, + 0.01604577712714672, + 0.020422326400876045, + 0.026410559192299843, + 0.02712048962712288, + 0.021039968356490135, + 0.03093699924647808, + 0.016072910279035568, + 0.02115781232714653, + 0.02529851719737053, + 0.02308075875043869, + 0.019038546830415726, + 0.015341369435191154, + 0.04285906255245209, + 0.036912184208631516, + 0.01644374057650566, + 0.010049729607999325, + 0.014759448356926441, + 0.01204716693609953, + 0.009926104918122292, + 0.006300094071775675, + 0.00856635719537735, + 0.010258176364004612, + 0.004932214971631765, + 0.004276443738490343, + 0.004092914052307606, + 0.006117130629718304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.018173959106206894 + } + ], + "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/0000152/group_6/index_tensor([899, 899])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_tensor([899, 899])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ed8ffe2e224dc537462e6fd521429cd53dab57a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_tensor([899, 899])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic shot, rendered with Octane for stunning physical realism, begins from a very low angle. A Subaru WRX STI ARX, covered in thick, wet mud, completes a breathtaking leap before crashing down onto the uneven gravel track. At the moment of impact, its specialized suspension system compresses to its limits, sending the car's wide body plummeting, nearly contacting its tires. Simultaneously, a shockwave of mud, water, and gravel erupts from the wheels, each particle visible in breathtaking detail thanks to physically accurate ray tracing. The car's classic blue livery and geometric decals appear faintly beneath the mud, the wet mud reflecting an oily texture in the bright, diffused light beneath the overcast sky. Then, the camera quickly and steadily tilts upwards and pans along with the car, capturing it as it instantly recovers its body posture. With the huge roar of the engine, the four-wheel drive system frantically digs up the ground, kicking up a cloud of dust, and its figure disappears into the dust and mist in the distance of the track. The whole picture is full of adrenaline-pumping tension and mechanical violent aesthetics, showing the moments of power and speed in racing to the fullest.", + "seed": 8607, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8607_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": -0.05801931200549006, + "vlm_logprob": -0.05801931200549006, + "vlm_yes_no_margin": 3.3499999761581423, + "vlm_token_credit": -0.05801931200549006, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10275542736053467 + ], + "opposite_logprob_tokens": [ + -2.477755546569824 + ], + "logprob_reward": -0.10275542736053467, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10275542736053467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13503995537757874, + 0.0641595870256424, + 0.038145989179611206, + 0.06714694947004318, + 0.042149756103754044, + 0.09398242086172104, + 0.038747675716876984, + 0.13780321180820465, + 0.14555270969867706, + 0.2372717410326004 + ], + "credit_map": [ + 0.019085897132754326, + 0.012417338788509369, + 0.01321394182741642, + 0.015720603987574577, + 0.015891795977950096, + 0.010484020225703716, + 0.026547979563474655, + 0.021678373217582703, + 0.00878148153424263, + 0.006623121909797192, + 0.010484005324542522, + 0.005384875927120447, + 0.007981585338711739, + 0.01025413628667593, + 0.0056070988066494465, + 0.009043279103934765, + 0.004633900243788958, + 0.005127195734530687, + 0.004054406192153692, + 0.004179511219263077, + 0.005046892445534468, + 0.007187813986092806, + 0.003927407320588827, + 0.003988862503319979, + 0.00708814337849617, + 0.00746864452958107, + 0.012032667174935341, + 0.006758962757885456, + 0.009699095971882343, + 0.00851126667112112, + 0.008684410713613033, + 0.006903760600835085, + 0.004995005205273628, + 0.004274298436939716, + 0.0037689879536628723, + 0.004102538339793682, + 0.004110239911824465, + 0.0068286750465631485, + 0.0088950265198946, + 0.005174986552447081, + 0.01631169393658638, + 0.024192197248339653, + 0.006107262801378965, + 0.00834526214748621, + 0.013593222945928574, + 0.008788908831775188, + 0.008795778267085552, + 0.007848089560866356, + 0.003742678789421916, + 0.0024718239437788725, + 0.006481684744358063, + 0.008933464996516705, + 0.0025527768302708864, + 0.0038048799615353346, + 0.005601478740572929, + 0.005158889107406139, + 0.014221306890249252, + 0.010837518610060215, + 0.032028790563344955, + 0.023215236142277718, + 0.012450577691197395, + 0.009657232090830803, + 0.018099255859851837, + 0.0172932930290699, + 0.018525440245866776, + 0.010494411922991276, + 0.024339884519577026, + 0.010179240256547928, + 0.013673866167664528, + 0.018918728455901146, + 0.026216380298137665, + 0.023204756900668144, + 0.015366114675998688, + 0.019009817391633987, + 0.023184018209576607, + 0.02034672722220421, + 0.04654799774289131, + 0.029852965846657753, + 0.0338343009352684, + 0.049129802733659744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.10275542736053467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708647549152374 + ], + "opposite_logprob_tokens": [ + -2.1270864009857178 + ], + "logprob_reward": -0.12708647549152374, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708647549152374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0838095024228096, + 0.08175093680620193, + 0.0710478276014328, + 0.07680758833885193, + 0.04789823293685913, + 0.0493096262216568, + 0.02984355390071869, + 0.14291410148143768, + 0.24688096344470978, + 0.16973768174648285 + ], + "credit_map": [ + 0.012735833413898945, + 0.009853648953139782, + 0.007580661214888096, + 0.00834935437887907, + 0.013775634579360485, + 0.009918327443301678, + 0.01305394060909748, + 0.008542098104953766, + 0.01190186943858862, + 0.009396885521709919, + 0.017530033364892006, + 0.007632907945662737, + 0.006772838532924652, + 0.013545839115977287, + 0.006443611811846495, + 0.008526958525180817, + 0.010250717401504517, + 0.013791903853416443, + 0.01056725438684225, + 0.0070547619834542274, + 0.008416008204221725, + 0.007508935406804085, + 0.005302035715430975, + 0.00815621204674244, + 0.0073118675500154495, + 0.008223817683756351, + 0.012416491284966469, + 0.011031423695385456, + 0.009229248389601707, + 0.009165490977466106, + 0.01142624206840992, + 0.008003005757927895, + 0.005856265313923359, + 0.004239924717694521, + 0.006235243286937475, + 0.005079395603388548, + 0.0037854937836527824, + 0.005894115660339594, + 0.010451164096593857, + 0.006356624886393547, + 0.005313483532518148, + 0.006543762516230345, + 0.003825465915724635, + 0.007241080980747938, + 0.005366673693060875, + 0.005787860602140427, + 0.005338018760085106, + 0.009893279522657394, + 0.0030208625830709934, + 0.002033510711044073, + 0.003604200202971697, + 0.005179763305932283, + 0.0020065077114850283, + 0.00312604708597064, + 0.005064156837761402, + 0.005808505229651928, + 0.009930456057190895, + 0.014400874264538288, + 0.01844688504934311, + 0.025145329535007477, + 0.01757613755762577, + 0.015012834221124649, + 0.02301120199263096, + 0.019390376284718513, + 0.03915773332118988, + 0.022697433829307556, + 0.036597929894924164, + 0.018451400101184845, + 0.026403576135635376, + 0.0421629399061203, + 0.029943788424134254, + 0.031466156244277954, + 0.015107541345059872, + 0.016285300254821777, + 0.013724501244723797, + 0.0157148614525795, + 0.03859918937087059, + 0.019601551815867424, + 0.022771798074245453, + 0.027932927012443542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.12708647549152374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608583465218544 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608583465218544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608583465218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11433093249797821, + 0.055919550359249115, + 0.12487407773733139, + 0.12133844196796417, + 0.056205447763204575, + 0.08424608409404755, + 0.0479072630405426, + 0.16019271314144135, + 0.097649946808815, + 0.13733549416065216 + ], + "credit_map": [ + 0.016790660098195076, + 0.010854795575141907, + 0.017865119501948357, + 0.016048254445195198, + 0.009446974843740463, + 0.007924527861177921, + 0.015500023029744625, + 0.019900590181350708, + 0.00793928001075983, + 0.007039833348244429, + 0.005691253114491701, + 0.006328468210995197, + 0.005870393477380276, + 0.011240949854254723, + 0.004775622859597206, + 0.00703375693410635, + 0.01453989464789629, + 0.015408555045723915, + 0.013972160406410694, + 0.01879032887518406, + 0.015782706439495087, + 0.01520504616200924, + 0.012261142022907734, + 0.018914256244897842, + 0.012404714711010456, + 0.011848930269479752, + 0.014909833669662476, + 0.015778830274939537, + 0.013842478394508362, + 0.011320789344608784, + 0.026392115280032158, + 0.014840761199593544, + 0.009052572771906853, + 0.00703387102112174, + 0.008724006824195385, + 0.005032572895288467, + 0.004153452813625336, + 0.0052391039207577705, + 0.009285598993301392, + 0.007684278767555952, + 0.012993118725717068, + 0.018211141228675842, + 0.010636525228619576, + 0.008564606308937073, + 0.008218291215598583, + 0.011312322691082954, + 0.007069332525134087, + 0.007240749429911375, + 0.004410585388541222, + 0.003165425267070532, + 0.0064229462295770645, + 0.011016600765287876, + 0.002762795891612768, + 0.004396100528538227, + 0.008192496374249458, + 0.0075403195805847645, + 0.015434576198458672, + 0.013551861047744751, + 0.03742450103163719, + 0.029136458411812782, + 0.014827095903456211, + 0.011108212172985077, + 0.02008550800383091, + 0.018624527379870415, + 0.014007989317178726, + 0.007025425788015127, + 0.015747593715786934, + 0.0068944236263632774, + 0.009220185689628124, + 0.012372978962957859, + 0.015979032963514328, + 0.01640232466161251, + 0.012937542982399464, + 0.013382764533162117, + 0.01750384271144867, + 0.017351508140563965, + 0.02794541046023369, + 0.017908526584506035, + 0.012619473040103912, + 0.01768643967807293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.020608583465218544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698562532663345 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.033698562532663345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698562532663345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11267028748989105, + 0.09383264183998108, + 0.12142068892717361, + 0.13309435546398163, + 0.07486969977617264, + 0.12996797263622284, + 0.027724379673600197, + 0.07288938015699387, + 0.12057777494192123, + 0.1129528135061264 + ], + "credit_map": [ + 0.016630027443170547, + 0.016339248046278954, + 0.011585458181798458, + 0.016633398830890656, + 0.015211695805191994, + 0.013873206451535225, + 0.01232134085148573, + 0.010075908154249191, + 0.013965529389679432, + 0.010400932282209396, + 0.008731178008019924, + 0.008648106828331947, + 0.013075394555926323, + 0.021491890773177147, + 0.008009593933820724, + 0.00951001513749361, + 0.01372104324400425, + 0.01664811186492443, + 0.010553884319961071, + 0.0106405233964324, + 0.017581971362233162, + 0.020953603088855743, + 0.012719855643808842, + 0.018601693212985992, + 0.01525851245969534, + 0.014788802713155746, + 0.017552878707647324, + 0.01902513951063156, + 0.010784933343529701, + 0.011757449246942997, + 0.018056072294712067, + 0.02587057091295719, + 0.011681433767080307, + 0.00952460803091526, + 0.007397981360554695, + 0.006137980613857508, + 0.004636115860193968, + 0.007509730290621519, + 0.017707476392388344, + 0.010274373926222324, + 0.023696469143033028, + 0.03791952505707741, + 0.0094639603048563, + 0.0119153568521142, + 0.014610201120376587, + 0.015390711836516857, + 0.008672733791172504, + 0.008299005217850208, + 0.0026017064228653908, + 0.002193605760112405, + 0.0037936540320515633, + 0.00574436504393816, + 0.0019449638202786446, + 0.003004078520461917, + 0.004586956463754177, + 0.003855049842968583, + 0.010655844584107399, + 0.006138266995549202, + 0.012557419016957283, + 0.012126807123422623, + 0.005454840138554573, + 0.004741949029266834, + 0.010803421959280968, + 0.01041083037853241, + 0.020430592820048332, + 0.008883854374289513, + 0.01920888014137745, + 0.008280343376100063, + 0.012053798884153366, + 0.020645322278141975, + 0.014287365600466728, + 0.01678762212395668, + 0.014531990513205528, + 0.02107127569615841, + 0.011237921193242073, + 0.015059941448271275, + 0.019010048359632492, + 0.013088848441839218, + 0.009232752025127411, + 0.009720039553940296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.033698562532663345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947511177510023 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947511177510023, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947511177510023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05964742600917816, + 0.021397342905402184, + 0.029794985428452492, + 0.039647094905376434, + 0.026067329570651054, + 0.05951172858476639, + 0.019984496757388115, + 0.04906940832734108, + 0.1439153403043747, + 0.550964891910553 + ], + "credit_map": [ + 0.009465279057621956, + 0.008307981304824352, + 0.007737878244370222, + 0.009022682905197144, + 0.0062908511608839035, + 0.005896549206227064, + 0.00663431640714407, + 0.0062918891198933125, + 0.002735877875238657, + 0.0016583703691139817, + 0.0034726527519524097, + 0.0023604868911206722, + 0.00244529964402318, + 0.00322659476660192, + 0.0019682252313941717, + 0.0035298371221870184, + 0.003917938098311424, + 0.004246511030942202, + 0.003464263863861561, + 0.0038173075299710035, + 0.003548063337802887, + 0.004517373628914356, + 0.002640877617523074, + 0.003642651718109846, + 0.005965183023363352, + 0.0042178998701274395, + 0.008618662133812904, + 0.003652155166491866, + 0.0026591233909130096, + 0.0032230450306087732, + 0.003910360857844353, + 0.007400667294859886, + 0.0028703073039650917, + 0.0022564001847058535, + 0.0019707968458533287, + 0.0024312175810337067, + 0.0034470336977392435, + 0.0044920239597558975, + 0.005240114871412516, + 0.003359433263540268, + 0.011163213290274143, + 0.017159411683678627, + 0.005464138463139534, + 0.004711625631898642, + 0.005789386108517647, + 0.005412361118942499, + 0.003871392225846648, + 0.005940199829638004, + 0.0019593986216932535, + 0.0010234859073534608, + 0.003757541999220848, + 0.0038928163703531027, + 0.0013594187330454588, + 0.0019684575963765383, + 0.0025243146810680628, + 0.0034990631975233555, + 0.0035372611600905657, + 0.005268783308565617, + 0.007194042671471834, + 0.011836913414299488, + 0.004386033862829208, + 0.004072591662406921, + 0.006906265392899513, + 0.005867514759302139, + 0.02253945916891098, + 0.010637711733579636, + 0.017734220251441002, + 0.012003490701317787, + 0.009627272374927998, + 0.018335416913032532, + 0.029445940628647804, + 0.023591825738549232, + 0.05109753459692001, + 0.05813205987215042, + 0.04438577592372894, + 0.03627154603600502, + 0.1854565292596817, + 0.08767854422330856, + 0.049586910754442215, + 0.03835595026612282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.005947511177510023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, rendered with Octane for stunning physical realism, begins from a very low angle. A Subaru WRX STI ARX, covered in thick, wet mud, completes a breathtaking leap before crashing down onto the uneven gravel track. At the moment of impact, its specialized suspension system compresses to its limits, sending the car's wide body plummeting, nearly contacting its tires. Simultaneously, a shockwave of mud, water, and gravel erupts from the wheels, each particle visible in breathtaking detail thanks to physically accurate ray tracing. The car's classic blue livery and geometric decals appear faintly beneath the mud, the wet mud reflecting an oily texture in the bright, diffused light beneath the overcast sky. Then, the camera quickly and steadily tilts upwards and pans along with the car, capturing it as it instantly recovers its body posture. With the huge roar of the engine, the four-wheel drive system frantically digs up the ground, kicking up a cloud of dust, and its figure disappears into the dust and mist in the distance of the track. The whole picture is full of adrenaline-pumping tension and mechanical violent aesthetics, showing the moments of power and speed in racing to the fullest.", + "seed": 8608, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8608_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": -0.06750176791101695, + "vlm_logprob": -0.06750176791101695, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.06750176791101695, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025198258459568024 + ], + "opposite_logprob_tokens": [ + -5.400198459625244 + ], + "logprob_reward": -0.025198258459568024, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.025198258459568024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20198187232017517, + 0.07432793825864792, + 0.04998830333352089, + 0.04447284713387489, + 0.11811845004558563, + 0.2055162936449051, + 0.044376641511917114, + 0.048106711357831955, + 0.12810148298740387, + 0.08500943332910538 + ], + "credit_map": [ + 0.023913264274597168, + 0.02165490947663784, + 0.039308395236730576, + 0.0384843610227108, + 0.015649741515517235, + 0.011958344839513302, + 0.027442574501037598, + 0.023570286110043526, + 0.01153588853776455, + 0.00747821107506752, + 0.008261383511126041, + 0.011793495155870914, + 0.005687524098902941, + 0.005985807627439499, + 0.010866970755159855, + 0.012718657031655312, + 0.005781783722341061, + 0.005030767992138863, + 0.005505947396159172, + 0.00803497526794672, + 0.003524885280057788, + 0.004883725196123123, + 0.008442847989499569, + 0.00878337025642395, + 0.004890571814030409, + 0.006394789554178715, + 0.006006449926644564, + 0.004881834611296654, + 0.005457752384245396, + 0.0050524757243692875, + 0.004456991329789162, + 0.0073319789953529835, + 0.006727748550474644, + 0.0063996990211308, + 0.007063260301947594, + 0.00554290134459734, + 0.00697306590154767, + 0.033950626850128174, + 0.013380315154790878, + 0.03808082267642021, + 0.009877587668597698, + 0.007274592760950327, + 0.011305056512355804, + 0.014637351967394352, + 0.007311827968806028, + 0.024499528110027313, + 0.04426414147019386, + 0.08634620904922485, + 0.005059251561760902, + 0.0037428021896630526, + 0.007958072237670422, + 0.010364638641476631, + 0.003309955820441246, + 0.0029246562626212835, + 0.005100039765238762, + 0.00591722596436739, + 0.003064665710553527, + 0.003978201188147068, + 0.005432297009974718, + 0.007317693904042244, + 0.005111376289278269, + 0.010090917348861694, + 0.006935594603419304, + 0.006175965536385775, + 0.01004040241241455, + 0.01797293685376644, + 0.009639753960072994, + 0.01211343239992857, + 0.023030323907732964, + 0.01939271204173565, + 0.022515133023262024, + 0.013396785594522953, + 0.006574534811079502, + 0.0145899448543787, + 0.010938589461147785, + 0.010915504768490791, + 0.010491752997040749, + 0.00691362377256155, + 0.01265841256827116, + 0.011927069164812565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.025198258459568024 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282825589179993 + ], + "opposite_logprob_tokens": [ + -2.0178282260894775 + ], + "logprob_reward": -0.14282825589179993, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282825589179993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10404223203659058, + 0.08520831167697906, + 0.03642736002802849, + 0.04753512516617775, + 0.04906340688467026, + 0.07551423460245132, + 0.07334250956773758, + 0.07374506443738937, + 0.1705119013786316, + 0.2846098840236664 + ], + "credit_map": [ + 0.012806622311472893, + 0.011829568073153496, + 0.0182278323918581, + 0.01853054016828537, + 0.009816640987992287, + 0.0072427066043019295, + 0.013082746416330338, + 0.012505565769970417, + 0.007537599187344313, + 0.004520545247942209, + 0.010116241872310638, + 0.009351585991680622, + 0.006379214581102133, + 0.005061653908342123, + 0.019187504425644875, + 0.0230539683252573, + 0.004374452866613865, + 0.0037025471683591604, + 0.006638493854552507, + 0.008100887760519981, + 0.002469204366207123, + 0.0030045330058783293, + 0.004325383342802525, + 0.003811856731772423, + 0.0037338463589549065, + 0.006975299213081598, + 0.008256706409156322, + 0.007282419130206108, + 0.004657419864088297, + 0.0057321819476783276, + 0.0045678638853132725, + 0.006329390220344067, + 0.0046414341777563095, + 0.006333943456411362, + 0.004010329023003578, + 0.007587357424199581, + 0.006476802751421928, + 0.006812504027038813, + 0.006518339738249779, + 0.006682696286588907, + 0.004365311004221439, + 0.004108788911253214, + 0.003711557714268565, + 0.008346742950379848, + 0.004448567051440477, + 0.007708004675805569, + 0.015691781416535378, + 0.027133487164974213, + 0.009785120375454426, + 0.008284954354166985, + 0.010965226218104362, + 0.016539210453629494, + 0.005110983271151781, + 0.00608859583735466, + 0.008128651417791843, + 0.008439767174422741, + 0.0045379349030554295, + 0.007813164964318275, + 0.010286716744303703, + 0.011091182939708233, + 0.007660015486180782, + 0.014096737839281559, + 0.009842714294791222, + 0.008416599594056606, + 0.011758087202906609, + 0.018970871344208717, + 0.013561046682298183, + 0.021863095462322235, + 0.026113856583833694, + 0.017677705734968185, + 0.038557734340429306, + 0.02200949564576149, + 0.020377708598971367, + 0.061916954815387726, + 0.03291629999876022, + 0.03989700227975845, + 0.03603960946202278, + 0.0269723329693079, + 0.03673696890473366, + 0.029752997681498528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.14282825589179993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01422515232115984 + ], + "opposite_logprob_tokens": [ + -4.264225006103516 + ], + "logprob_reward": -0.01422515232115984, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01422515232115984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08470621705055237, + 0.060606520622968674, + 0.07404706627130508, + 0.03806672245264053, + 0.0726398229598999, + 0.07363926619291306, + 0.2526184320449829, + 0.05657367408275604, + 0.1463620364665985, + 0.1407402902841568 + ], + "credit_map": [ + 0.012172402814030647, + 0.010684395208954811, + 0.01328289695084095, + 0.013951980508863926, + 0.00786078255623579, + 0.006200909148901701, + 0.010539386421442032, + 0.010013459250330925, + 0.008655172772705555, + 0.006447755265980959, + 0.01404537633061409, + 0.008697034791111946, + 0.004218710120767355, + 0.004599340260028839, + 0.00838602613657713, + 0.005557098891586065, + 0.0072722844779491425, + 0.006566150113940239, + 0.010774624533951283, + 0.014436074532568455, + 0.008073079399764538, + 0.0068226647563278675, + 0.009124688804149628, + 0.010977492667734623, + 0.0038112362381070852, + 0.004495964851230383, + 0.004483395721763372, + 0.005525241140276194, + 0.004052194766700268, + 0.0046752807684242725, + 0.003908275626599789, + 0.007115129381418228, + 0.003986665979027748, + 0.0053200372494757175, + 0.004650644958019257, + 0.006108302157372236, + 0.005312510300427675, + 0.014264854602515697, + 0.01030373852699995, + 0.022693071514368057, + 0.006402029190212488, + 0.003851252608001232, + 0.006827253848314285, + 0.008945532143115997, + 0.0051379031501710415, + 0.012638242915272713, + 0.008393174968659878, + 0.021443868055939674, + 0.02196854166686535, + 0.024895859882235527, + 0.06535577028989792, + 0.040803395211696625, + 0.021982576698064804, + 0.014139877632260323, + 0.022418085485696793, + 0.04105432331562042, + 0.004244806244969368, + 0.0030741586815565825, + 0.014064575545489788, + 0.007470821496099234, + 0.004034435376524925, + 0.007879954762756824, + 0.008683751337230206, + 0.007121165748685598, + 0.009641250595450401, + 0.017033742740750313, + 0.01092221774160862, + 0.015825174748897552, + 0.036479219794273376, + 0.015590358525514603, + 0.026109006255865097, + 0.01476105023175478, + 0.0107582937926054, + 0.023829149082303047, + 0.016810107976198196, + 0.013252872042357922, + 0.014096577651798725, + 0.010807909071445465, + 0.031528688967227936, + 0.019656682386994362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.01422515232115984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495903640985489 + ], + "opposite_logprob_tokens": [ + -2.9299590587615967 + ], + "logprob_reward": -0.05495903640985489, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495903640985489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10785023868083954, + 0.13773201406002045, + 0.02970636822283268, + 0.0921914353966713, + 0.03244997188448906, + 0.15482014417648315, + 0.17424018681049347, + 0.046020377427339554, + 0.08560109883546829, + 0.13938811421394348 + ], + "credit_map": [ + 0.010382862761616707, + 0.008749039843678474, + 0.017647404223680496, + 0.02091495878994465, + 0.011863506399095058, + 0.006694596726447344, + 0.018702838569879532, + 0.012895035557448864, + 0.01142796128988266, + 0.010039971210062504, + 0.0408625453710556, + 0.013735446147620678, + 0.009013320319354534, + 0.007001469377428293, + 0.035570453852415085, + 0.010080844163894653, + 0.0035078490618616343, + 0.002465195721015334, + 0.005427805706858635, + 0.0065573276951909065, + 0.0023409121204167604, + 0.0022322216536849737, + 0.003315620357170701, + 0.0038594368379563093, + 0.007243919651955366, + 0.015887297689914703, + 0.015453051775693893, + 0.01290605403482914, + 0.009573821909725666, + 0.010325541719794273, + 0.009739482775330544, + 0.01106227282434702, + 0.002500511473044753, + 0.0024930823128670454, + 0.004490779247134924, + 0.0041611953638494015, + 0.00313912658020854, + 0.00320704304613173, + 0.00640763808041811, + 0.006050598807632923, + 0.010401720181107521, + 0.0057007670402526855, + 0.011485590599477291, + 0.00895291194319725, + 0.0070403399877250195, + 0.030233072116971016, + 0.02212470956146717, + 0.05888104811310768, + 0.016053881496191025, + 0.017416896298527718, + 0.047209884971380234, + 0.03924275189638138, + 0.017322387546300888, + 0.01039043813943863, + 0.01231591310352087, + 0.01428805198520422, + 0.0036565365735441446, + 0.004285774659365416, + 0.006059932056814432, + 0.007617484778165817, + 0.0041318428702652454, + 0.007755420636385679, + 0.007204802241176367, + 0.005308589432388544, + 0.006197362672537565, + 0.007736064959317446, + 0.006674407981336117, + 0.01065686997026205, + 0.010611837729811668, + 0.009704405441880226, + 0.02347918227314949, + 0.010540974326431751, + 0.009738797321915627, + 0.026781564578413963, + 0.018915385007858276, + 0.013534863479435444, + 0.017573876306414604, + 0.010862104594707489, + 0.019048526883125305, + 0.022933004423975945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.05495903640985489 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029813647270203 + ], + "opposite_logprob_tokens": [ + -2.3502981662750244 + ], + "logprob_reward": -0.10029813647270203, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029813647270203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07778766751289368, + 0.08955947309732437, + 0.03265771642327309, + 0.025247670710086823, + 0.041975442320108414, + 0.03969033807516098, + 0.07581619173288345, + 0.0700339525938034, + 0.12257856130599976, + 0.42465293407440186 + ], + "credit_map": [ + 0.00871264562010765, + 0.007668034173548222, + 0.01289286557585001, + 0.012488091364502907, + 0.008984830230474472, + 0.006201518699526787, + 0.012356612831354141, + 0.008483070880174637, + 0.009778222069144249, + 0.006399018224328756, + 0.01783742941915989, + 0.010446797125041485, + 0.00803721696138382, + 0.005484554450958967, + 0.020217513665556908, + 0.011358722113072872, + 0.0042870668694376945, + 0.0033950740471482277, + 0.0038906570989638567, + 0.0055341958068311214, + 0.003241131780669093, + 0.0035667498596012592, + 0.004295002203434706, + 0.0044478378258645535, + 0.00277534662745893, + 0.0031899802852422, + 0.0033450631890445948, + 0.003835023380815983, + 0.0024989591911435127, + 0.002652184572070837, + 0.0023393721785396338, + 0.0046117412857711315, + 0.0021455134265124798, + 0.0029895531479269266, + 0.0030080946162343025, + 0.0048810383304953575, + 0.003354901447892189, + 0.009947295300662518, + 0.005873308051377535, + 0.009775742888450623, + 0.0021263472735881805, + 0.002130883513018489, + 0.001906092744320631, + 0.0032340020406991243, + 0.0029519174713641405, + 0.005052578169852495, + 0.008062699809670448, + 0.014225817285478115, + 0.006644492968916893, + 0.005941302515566349, + 0.023213956505060196, + 0.017063621431589127, + 0.00434173783287406, + 0.003195937955752015, + 0.006026065442711115, + 0.009389076381921768, + 0.004647433292120695, + 0.004276781342923641, + 0.010233946144580841, + 0.011081643402576447, + 0.006838268134742975, + 0.012639548629522324, + 0.011254335753619671, + 0.009061990305781364, + 0.008311904035508633, + 0.010798749513924122, + 0.011229763738811016, + 0.015834137797355652, + 0.016972560435533524, + 0.015354440547525883, + 0.030321985483169556, + 0.013755018822848797, + 0.031149184331297874, + 0.06915739923715591, + 0.042329512536525726, + 0.05503105744719505, + 0.04185618832707405, + 0.026236433535814285, + 0.09398432075977325, + 0.06490883231163025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.10029813647270203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, rendered with Octane for stunning physical realism, begins from a very low angle. A Subaru WRX STI ARX, covered in thick, wet mud, completes a breathtaking leap before crashing down onto the uneven gravel track. At the moment of impact, its specialized suspension system compresses to its limits, sending the car's wide body plummeting, nearly contacting its tires. Simultaneously, a shockwave of mud, water, and gravel erupts from the wheels, each particle visible in breathtaking detail thanks to physically accurate ray tracing. The car's classic blue livery and geometric decals appear faintly beneath the mud, the wet mud reflecting an oily texture in the bright, diffused light beneath the overcast sky. Then, the camera quickly and steadily tilts upwards and pans along with the car, capturing it as it instantly recovers its body posture. With the huge roar of the engine, the four-wheel drive system frantically digs up the ground, kicking up a cloud of dust, and its figure disappears into the dust and mist in the distance of the track. The whole picture is full of adrenaline-pumping tension and mechanical violent aesthetics, showing the moments of power and speed in racing to the fullest.", + "seed": 8609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8609_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": -0.04778744913637638, + "vlm_logprob": -0.04778744913637638, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.04778744913637638, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030677789822220802 + ], + "opposite_logprob_tokens": [ + -4.405677795410156 + ], + "logprob_reward": -0.030677789822220802, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.030677789822220802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07271991670131683, + 0.03556354343891144, + 0.08086986839771271, + 0.034879036247730255, + 0.031892456114292145, + 0.15142489969730377, + 0.21853716671466827, + 0.1388009935617447, + 0.09394344687461853, + 0.14136861264705658 + ], + "credit_map": [ + 0.008351092226803303, + 0.010853941552340984, + 0.008934104815125465, + 0.008698152378201485, + 0.010537009686231613, + 0.009047538042068481, + 0.007653443142771721, + 0.00864463858306408, + 0.0031883062329143286, + 0.004431346897035837, + 0.0038128443993628025, + 0.004830569494515657, + 0.005219189450144768, + 0.004439104814082384, + 0.005935055669397116, + 0.003707129042595625, + 0.006469772197306156, + 0.009562510065734386, + 0.008427082560956478, + 0.011520619504153728, + 0.011832191608846188, + 0.00784547533839941, + 0.01483296137303114, + 0.010379256680607796, + 0.0030257466714829206, + 0.0031104362569749355, + 0.0028736938256770372, + 0.004948478192090988, + 0.003052215091884136, + 0.0039756870828568935, + 0.008862419985234737, + 0.005030359141528606, + 0.0041016447357833385, + 0.0039026830345392227, + 0.0033742443192750216, + 0.007225657347589731, + 0.0030663572251796722, + 0.002649191999807954, + 0.003900995710864663, + 0.003671683371067047, + 0.015244686044752598, + 0.01701250486075878, + 0.04008282348513603, + 0.05183609947562218, + 0.007956167683005333, + 0.004836223088204861, + 0.00858659390360117, + 0.005869801621884108, + 0.023391801863908768, + 0.02355245314538479, + 0.038246672600507736, + 0.05566820502281189, + 0.019774287939071655, + 0.014345696195960045, + 0.029099073261022568, + 0.014458993449807167, + 0.008727502077817917, + 0.01167570985853672, + 0.04135496914386749, + 0.037749532610177994, + 0.011351966299116611, + 0.00752190500497818, + 0.013588152825832367, + 0.0068312743678689, + 0.015365437604486942, + 0.008480757474899292, + 0.015385186299681664, + 0.026140758767724037, + 0.006018887273967266, + 0.004852532409131527, + 0.009873026981949806, + 0.007826861925423145, + 0.010849393904209137, + 0.010439636185765266, + 0.01627068966627121, + 0.019237954169511795, + 0.013263280503451824, + 0.012775569222867489, + 0.04212670400738716, + 0.01640539988875389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.030677789822220802 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295779794454575 + ], + "opposite_logprob_tokens": [ + -2.237957715988159 + ], + "logprob_reward": -0.11295779794454575, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295779794454575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0399366095662117, + 0.03125912696123123, + 0.07014833390712738, + 0.05601629614830017, + 0.03977876901626587, + 0.2954845428466797, + 0.2572946548461914, + 0.07638763636350632, + 0.07809410989284515, + 0.055599868297576904 + ], + "credit_map": [ + 0.005826180335134268, + 0.004977895878255367, + 0.00474119745194912, + 0.004109312314540148, + 0.0065129767172038555, + 0.0048737917095422745, + 0.004228808451443911, + 0.004666448570787907, + 0.003231285372748971, + 0.0036372013855725527, + 0.0036527919583022594, + 0.004006365314126015, + 0.0044662440195679665, + 0.003414938924834132, + 0.004183984827250242, + 0.004666313994675875, + 0.006830381229519844, + 0.006890018004924059, + 0.009628907777369022, + 0.007965104654431343, + 0.006812025792896748, + 0.008755475282669067, + 0.013126860372722149, + 0.010139563120901585, + 0.0049872249364852905, + 0.007432566024363041, + 0.0059857843443751335, + 0.006614523474127054, + 0.007098677568137646, + 0.0060744634829461575, + 0.009735067375004292, + 0.008087990805506706, + 0.004631553310900927, + 0.004480654839426279, + 0.004018814768642187, + 0.005939201917499304, + 0.004369413945823908, + 0.004251237493008375, + 0.00544271944090724, + 0.006645173765718937, + 0.020260965451598167, + 0.029902776703238487, + 0.07552209496498108, + 0.1025889590382576, + 0.02871069125831127, + 0.01130086462944746, + 0.016618169844150543, + 0.010580038651823997, + 0.025619719177484512, + 0.029201332479715347, + 0.03784145787358284, + 0.053407732397317886, + 0.03146199882030487, + 0.02702578902244568, + 0.028964539989829063, + 0.023772070184350014, + 0.007577040232717991, + 0.009211540222167969, + 0.010861692018806934, + 0.013142352923750877, + 0.012011798098683357, + 0.007148401811718941, + 0.008180352859199047, + 0.0082544581964612, + 0.014139262028038502, + 0.008905375376343727, + 0.011282014660537243, + 0.015830950811505318, + 0.004853444639593363, + 0.00719232764095068, + 0.008057545870542526, + 0.007833191193640232, + 0.010648022405803204, + 0.008662646636366844, + 0.005402817856520414, + 0.007037049159407616, + 0.006941491272300482, + 0.006504523102194071, + 0.006329824682325125, + 0.004073493182659149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.11295779794454575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221626333892345 + ], + "opposite_logprob_tokens": [ + -4.264221668243408 + ], + "logprob_reward": -0.014221626333892345, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221626333892345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09413208067417145, + 0.27256500720977783, + 0.07219481468200684, + 0.04297535493969917, + 0.03619258850812912, + 0.07672710716724396, + 0.14370888471603394, + 0.13041479885578156, + 0.037921298295259476, + 0.09316804260015488 + ], + "credit_map": [ + 0.00935025792568922, + 0.01491923164576292, + 0.01304082665592432, + 0.011042465455830097, + 0.013814673759043217, + 0.010379718616604805, + 0.00925834383815527, + 0.012326564639806747, + 0.016627062112092972, + 0.026008769869804382, + 0.0566118098795414, + 0.029575824737548828, + 0.03284331038594246, + 0.08025026321411133, + 0.011867607943713665, + 0.018780386075377464, + 0.006130015943199396, + 0.00986333005130291, + 0.008098156191408634, + 0.010769068263471127, + 0.008634878322482109, + 0.007213226053863764, + 0.011380374431610107, + 0.010105764493346214, + 0.003609019098803401, + 0.0025952099822461605, + 0.004365826956927776, + 0.0053949058055877686, + 0.003386386437341571, + 0.005026704166084528, + 0.010653553530573845, + 0.00794374942779541, + 0.003189626382663846, + 0.003985076677054167, + 0.005503133404999971, + 0.005912603810429573, + 0.004095182288438082, + 0.005551215261220932, + 0.003900067415088415, + 0.004055684898048639, + 0.007221940439194441, + 0.01883986033499241, + 0.010471864603459835, + 0.022036932408809662, + 0.005165396258234978, + 0.004329541232436895, + 0.004350015893578529, + 0.004311562981456518, + 0.0165289044380188, + 0.020072640851140022, + 0.017436200752854347, + 0.02205665037035942, + 0.018848994746804237, + 0.017673373222351074, + 0.013679342344403267, + 0.017412777990102768, + 0.008195213042199612, + 0.011708651669323444, + 0.02369959093630314, + 0.028280436992645264, + 0.01431242935359478, + 0.009290376678109169, + 0.024516720324754715, + 0.010411384515464306, + 0.00524043757468462, + 0.005536079406738281, + 0.004534044302999973, + 0.006608164869248867, + 0.003447003662586212, + 0.0033950686920434237, + 0.0051929340697824955, + 0.003967565950006247, + 0.00995014701038599, + 0.009272973984479904, + 0.009525241330265999, + 0.012271284125745296, + 0.010430382564663887, + 0.011439104564487934, + 0.02103312499821186, + 0.009245775640010834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.014221626333892345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000859081745148 + ], + "opposite_logprob_tokens": [ + -2.6950085163116455 + ], + "logprob_reward": -0.07000859081745148, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000859081745148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0649627223610878, + 0.08976233750581741, + 0.09956557303667068, + 0.06334695965051651, + 0.0869283527135849, + 0.16324707865715027, + 0.06570456176996231, + 0.21059761941432953, + 0.11240578442811966, + 0.043479036539793015 + ], + "credit_map": [ + 0.007030178792774677, + 0.010044327937066555, + 0.011901085264980793, + 0.008671538904309273, + 0.007056671194732189, + 0.0074310521595180035, + 0.006395244039595127, + 0.006432624533772469, + 0.0052896346896886826, + 0.010491535998880863, + 0.014760167337954044, + 0.010168863460421562, + 0.014198641292750835, + 0.021192731335759163, + 0.007669900078326464, + 0.00599086657166481, + 0.011569787748157978, + 0.010248040780425072, + 0.012867970392107964, + 0.01146398764103651, + 0.012349316850304604, + 0.00999528169631958, + 0.017633114010095596, + 0.01343807578086853, + 0.006020879838615656, + 0.005829797126352787, + 0.006912715267390013, + 0.007945619523525238, + 0.006436321418732405, + 0.00881683174520731, + 0.007946550846099854, + 0.013438238762319088, + 0.005283810663968325, + 0.005390767939388752, + 0.010796326212584972, + 0.019768599420785904, + 0.006933476310223341, + 0.008890584111213684, + 0.010155364871025085, + 0.019709428772330284, + 0.015204059891402721, + 0.01672041416168213, + 0.0439210869371891, + 0.053191687911748886, + 0.01115185022354126, + 0.005781367886811495, + 0.010678408667445183, + 0.006598180625587702, + 0.012281189672648907, + 0.010911596938967705, + 0.008908922784030437, + 0.007210540119558573, + 0.008673337288200855, + 0.007584226783365011, + 0.005142036825418472, + 0.004992716014385223, + 0.015204577706754208, + 0.018508153036236763, + 0.04322534054517746, + 0.04068607836961746, + 0.02176235057413578, + 0.015573172830045223, + 0.03685314580798149, + 0.018784787505865097, + 0.01327354833483696, + 0.016423286870121956, + 0.014708545990288258, + 0.03525931388139725, + 0.00724966824054718, + 0.0054345265962183475, + 0.01273821759968996, + 0.0073186783120036125, + 0.004041133914142847, + 0.00287020206451416, + 0.00653938390314579, + 0.004865794442594051, + 0.0029878118075430393, + 0.0023559988476336002, + 0.01269841194152832, + 0.007120300550013781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.07000859081745148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071440763771534 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071440763771534, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071440763771534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04596914350986481, + 0.05754847824573517, + 0.06335066258907318, + 0.040522266179323196, + 0.05121740326285362, + 0.4058634638786316, + 0.061665404587984085, + 0.15063509345054626, + 0.05043278634548187, + 0.07279530167579651 + ], + "credit_map": [ + 0.005952784325927496, + 0.005691134836524725, + 0.0074105518870055676, + 0.0062246015295386314, + 0.005839368794113398, + 0.004720483906567097, + 0.005291995592415333, + 0.004838220309466124, + 0.00354007794521749, + 0.005585027392953634, + 0.01216284278780222, + 0.006361051462590694, + 0.0062136463820934296, + 0.014723912812769413, + 0.003047523321583867, + 0.005914398934692144, + 0.006942063570022583, + 0.008483966812491417, + 0.006095468532294035, + 0.00750752491876483, + 0.005086603108793497, + 0.006691611371934414, + 0.012990240007638931, + 0.009553184732794762, + 0.0035484812688082457, + 0.0044965362176299095, + 0.004499321803450584, + 0.003641860792413354, + 0.0053772469982504845, + 0.0038234300445765257, + 0.006190963555127382, + 0.008944426663219929, + 0.0062186322174966335, + 0.006508232560008764, + 0.0060880109667778015, + 0.010059009306132793, + 0.005391675513237715, + 0.005591251887381077, + 0.005432796664535999, + 0.005927796009927988, + 0.054565928876399994, + 0.08900772035121918, + 0.10016586631536484, + 0.06837310642004013, + 0.04164661094546318, + 0.018705492839217186, + 0.0192671287804842, + 0.01413161400705576, + 0.00729372026398778, + 0.014593904837965965, + 0.009295661002397537, + 0.006416383199393749, + 0.007449893746525049, + 0.005302696954458952, + 0.005930917803198099, + 0.00538222398608923, + 0.008841397240757942, + 0.012074652127921581, + 0.04840988665819168, + 0.04024611786007881, + 0.014392568729817867, + 0.008600123226642609, + 0.010986982844769955, + 0.007083353120833635, + 0.006750002969056368, + 0.008396114222705364, + 0.006816806737333536, + 0.011953935958445072, + 0.003824775805696845, + 0.003486375790089369, + 0.004845355171710253, + 0.004359422251582146, + 0.013162460178136826, + 0.00651317834854126, + 0.008938036859035492, + 0.007172355893999338, + 0.0061038038693368435, + 0.006503324955701828, + 0.015751050785183907, + 0.00865108985453844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.011071440763771534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, rendered with Octane for stunning physical realism, begins from a very low angle. A Subaru WRX STI ARX, covered in thick, wet mud, completes a breathtaking leap before crashing down onto the uneven gravel track. At the moment of impact, its specialized suspension system compresses to its limits, sending the car's wide body plummeting, nearly contacting its tires. Simultaneously, a shockwave of mud, water, and gravel erupts from the wheels, each particle visible in breathtaking detail thanks to physically accurate ray tracing. The car's classic blue livery and geometric decals appear faintly beneath the mud, the wet mud reflecting an oily texture in the bright, diffused light beneath the overcast sky. Then, the camera quickly and steadily tilts upwards and pans along with the car, capturing it as it instantly recovers its body posture. With the huge roar of the engine, the four-wheel drive system frantically digs up the ground, kicking up a cloud of dust, and its figure disappears into the dust and mist in the distance of the track. The whole picture is full of adrenaline-pumping tension and mechanical violent aesthetics, showing the moments of power and speed in racing to the fullest.", + "seed": 8610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_6/index_899_seed_8610_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": -0.021478707902133465, + "vlm_logprob": -0.021478707902133465, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.021478707902133465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015734568238258362 + ], + "opposite_logprob_tokens": [ + -4.390734672546387 + ], + "logprob_reward": -0.015734568238258362, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015734568238258362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07131432741880417, + 0.05534237623214722, + 0.043975092470645905, + 0.056213293224573135, + 0.35043996572494507, + 0.05088658258318901, + 0.18390892446041107, + 0.019377874210476875, + 0.12556111812591553, + 0.04298047348856926 + ], + "credit_map": [ + 0.009553506039083004, + 0.008372277021408081, + 0.011429660953581333, + 0.0070804934948682785, + 0.006340884603559971, + 0.008850842714309692, + 0.007557605393230915, + 0.012129057198762894, + 0.006126898806542158, + 0.00945145171135664, + 0.006170639768242836, + 0.007194775156676769, + 0.008961164392530918, + 0.004690858535468578, + 0.007079893257468939, + 0.005666694138199091, + 0.009420027956366539, + 0.007179603911936283, + 0.006038408726453781, + 0.004032664932310581, + 0.005165942013263702, + 0.004077324643731117, + 0.00324552645906806, + 0.004815590102225542, + 0.006577813532203436, + 0.005321821663528681, + 0.005809513386338949, + 0.012201037257909775, + 0.007725754287093878, + 0.004546362441033125, + 0.005977638531476259, + 0.008053346537053585, + 0.10713770985603333, + 0.032285794615745544, + 0.015179886482656002, + 0.01615333743393421, + 0.037986770272254944, + 0.11240510642528534, + 0.01502692699432373, + 0.014264429919421673, + 0.007373032625764608, + 0.005951454397290945, + 0.005499865394085646, + 0.006235112436115742, + 0.0030203508213162422, + 0.0043280706740915775, + 0.0055840834975242615, + 0.012894612736999989, + 0.03255072981119156, + 0.05963270366191864, + 0.01845194213092327, + 0.019330885261297226, + 0.011014239862561226, + 0.015167691744863987, + 0.0129190469160676, + 0.014841675758361816, + 0.00222119246609509, + 0.0033857873640954494, + 0.002395231043919921, + 0.003900616429746151, + 0.0016369533259421587, + 0.0017833379097282887, + 0.0020203765016049147, + 0.002034378470852971, + 0.02135884203016758, + 0.00949079543352127, + 0.03883682191371918, + 0.017136847600340843, + 0.008438519202172756, + 0.006270507350564003, + 0.010396553203463554, + 0.013632235117256641, + 0.005298350006341934, + 0.004974189214408398, + 0.0062469360418617725, + 0.00577817065641284, + 0.007413891144096851, + 0.00324043957516551, + 0.004873732570558786, + 0.005154761951416731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.015734568238258362 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497291311621666 + ], + "opposite_logprob_tokens": [ + -2.9299728870391846 + ], + "logprob_reward": -0.05497291311621666, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497291311621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06413478404283524, + 0.041191257536411285, + 0.034100305289030075, + 0.048018720000982285, + 0.47176438570022583, + 0.01845697872340679, + 0.061771586537361145, + 0.03366806358098984, + 0.16646477580070496, + 0.06042911857366562 + ], + "credit_map": [ + 0.007866238243877888, + 0.007322445511817932, + 0.009036512114107609, + 0.0068231001496315, + 0.006817844230681658, + 0.008212701417505741, + 0.009401693940162659, + 0.008654246106743813, + 0.005534838885068893, + 0.004232584498822689, + 0.004801465664058924, + 0.005777041427791119, + 0.006024729460477829, + 0.0034067598171532154, + 0.007110715378075838, + 0.004303121939301491, + 0.007729623466730118, + 0.004992456175386906, + 0.003933454863727093, + 0.003405527211725712, + 0.003718489082530141, + 0.0037622416857630014, + 0.002993972972035408, + 0.0035645412281155586, + 0.006013055797666311, + 0.00481070252135396, + 0.007072226610034704, + 0.006939078215509653, + 0.0064952620305120945, + 0.006080412305891514, + 0.005203706678003073, + 0.0054042749106884, + 0.1476837396621704, + 0.04401583597064018, + 0.02069968730211258, + 0.022293608635663986, + 0.049759309738874435, + 0.1464502215385437, + 0.020630354061722755, + 0.020231662318110466, + 0.0022416829597204924, + 0.002774161519482732, + 0.0032863435335457325, + 0.0020098614040762186, + 0.0013943834928795695, + 0.0020078502129763365, + 0.002327389782294631, + 0.0024153057020157576, + 0.014493430033326149, + 0.017581114545464516, + 0.005979740992188454, + 0.006087452173233032, + 0.003759955521672964, + 0.004644084721803665, + 0.003811969654634595, + 0.005413838662207127, + 0.004155420698225498, + 0.003701544599607587, + 0.004942235536873341, + 0.005956013221293688, + 0.0025477048475295305, + 0.003502735635265708, + 0.0032992681954056025, + 0.00556314317509532, + 0.035274796187877655, + 0.011379582807421684, + 0.05909500643610954, + 0.01922639086842537, + 0.008693215437233448, + 0.008426965214312077, + 0.013768298551440239, + 0.010600507259368896, + 0.009294774383306503, + 0.006299957167357206, + 0.01167781837284565, + 0.007415443193167448, + 0.006023643538355827, + 0.007812201511114836, + 0.005993261467665434, + 0.005912017542868853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.05497291311621666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963390693068504 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.005963390693068504, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963390693068504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05872754007577896, + 0.11801188439130783, + 0.08185040205717087, + 0.07814236730337143, + 0.2634737193584442, + 0.034687574952840805, + 0.16774329543113708, + 0.04504116252064705, + 0.04861166700720787, + 0.10371033847332001 + ], + "credit_map": [ + 0.007738387677818537, + 0.006720293778926134, + 0.007290937006473541, + 0.007308627478778362, + 0.007021029945462942, + 0.009766247123479843, + 0.006372260861098766, + 0.006509758532047272, + 0.015352851711213589, + 0.020433785393834114, + 0.012328111566603184, + 0.014617020264267921, + 0.0186899583786726, + 0.009789331816136837, + 0.014299444854259491, + 0.012501380406320095, + 0.015254173427820206, + 0.012355323880910873, + 0.009108597412705421, + 0.008776814676821232, + 0.010135187767446041, + 0.009724379517138004, + 0.006813161075115204, + 0.009682775475084782, + 0.009803589433431625, + 0.006501288153231144, + 0.009433740749955177, + 0.017451094463467598, + 0.008649223484098911, + 0.008894335478544235, + 0.008419876918196678, + 0.008989221416413784, + 0.07560580968856812, + 0.02443680353462696, + 0.012101048603653908, + 0.01289487536996603, + 0.02873772382736206, + 0.08368992805480957, + 0.011854024603962898, + 0.014153525233268738, + 0.004632032476365566, + 0.006592368707060814, + 0.005358262453228235, + 0.0026548937894403934, + 0.0025310118217021227, + 0.0034994857851415873, + 0.004735663067549467, + 0.004683858249336481, + 0.03531588613986969, + 0.06224551051855087, + 0.016447756439447403, + 0.014847427606582642, + 0.009074745699763298, + 0.014330204576253891, + 0.0073725697584450245, + 0.00810921099036932, + 0.00652727484703064, + 0.004969580098986626, + 0.008505241945385933, + 0.006788316648453474, + 0.00359923392534256, + 0.0058183069340884686, + 0.003612339496612549, + 0.005220872350037098, + 0.005844190716743469, + 0.004932088311761618, + 0.00837091263383627, + 0.008134773932397366, + 0.00486546428874135, + 0.003920471761375666, + 0.006717518903315067, + 0.005826251115649939, + 0.012450783513486385, + 0.008959073573350906, + 0.02038046158850193, + 0.015354499220848083, + 0.01076305378228426, + 0.012831457890570164, + 0.01127209234982729, + 0.01169892493635416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.005963390693068504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818554848432541 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818554848432541, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818554848432541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05806400626897812, + 0.05686891824007034, + 0.10668162256479263, + 0.10302377492189407, + 0.3333047330379486, + 0.03838712349534035, + 0.08517463505268097, + 0.05862504243850708, + 0.08123283833265305, + 0.07863730937242508 + ], + "credit_map": [ + 0.008236703462898731, + 0.005113087594509125, + 0.007288307882845402, + 0.008601713925600052, + 0.005264414008706808, + 0.007141559850424528, + 0.0062460992485284805, + 0.010172119364142418, + 0.007196152117103338, + 0.007991540245711803, + 0.005658148322254419, + 0.007257744204252958, + 0.009938977658748627, + 0.004059805069118738, + 0.008545875549316406, + 0.006220673676580191, + 0.020530810579657555, + 0.01827765256166458, + 0.012841799296438694, + 0.00854208692908287, + 0.013850842602550983, + 0.01138670276850462, + 0.009636200033128262, + 0.011615533381700516, + 0.010307087562978268, + 0.008913169614970684, + 0.011395534500479698, + 0.016510523855686188, + 0.013612015172839165, + 0.013183878734707832, + 0.012989597395062447, + 0.01611196994781494, + 0.0950184091925621, + 0.03132462501525879, + 0.02017565630376339, + 0.01649334654211998, + 0.0380416102707386, + 0.0984329879283905, + 0.01717505417764187, + 0.016643039882183075, + 0.0062909312546253204, + 0.0062718940898776054, + 0.005879207048565149, + 0.00427613640204072, + 0.0024802081752568483, + 0.004273335449397564, + 0.0040589189156889915, + 0.004856491461396217, + 0.02072286419570446, + 0.026686670258641243, + 0.008474305272102356, + 0.00638853432610631, + 0.0049192230217158794, + 0.0066868606954813, + 0.004561768379062414, + 0.006734411232173443, + 0.007545515429228544, + 0.007361699361354113, + 0.008851303718984127, + 0.009449095465242863, + 0.004444554913789034, + 0.005423710215836763, + 0.00606879498809576, + 0.009480367414653301, + 0.013870757073163986, + 0.007316939532756805, + 0.019986040890216827, + 0.010887986049056053, + 0.007088135927915573, + 0.005191511940211058, + 0.008436250500380993, + 0.008455213159322739, + 0.011979328468441963, + 0.007512220647186041, + 0.013246847316622734, + 0.010857848450541496, + 0.007645420730113983, + 0.009997805580496788, + 0.009405418299138546, + 0.007992418482899666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.01818554848432541 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Subaru WRX STI ARX covered in thick, wet mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the car perform a leap before crashing down onto an uneven gravel track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car's blue livery and geometric decals visible beneath the mud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start from a very low angle and then tilt upwards while panning along with the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the car kick up a cloud of dust and disappear into the distance after recovering its posture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253711897879839 + ], + "opposite_logprob_tokens": [ + -4.387537002563477 + ], + "logprob_reward": -0.01253711897879839, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253711897879839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06339351087808609, + 0.028795795515179634, + 0.04748760163784027, + 0.057789720594882965, + 0.15197306871414185, + 0.03558197245001793, + 0.10711656510829926, + 0.027614405378699303, + 0.2183624505996704, + 0.2618849277496338 + ], + "credit_map": [ + 0.009852134622633457, + 0.006885821931064129, + 0.006640185136348009, + 0.006412442773580551, + 0.007170086260885, + 0.005586383398622274, + 0.0067574637942016125, + 0.014088993892073631, + 0.0037190436851233244, + 0.003948942758142948, + 0.0035084779374301434, + 0.0033564723562449217, + 0.004606599919497967, + 0.002499374095350504, + 0.004297771956771612, + 0.0028591123409569263, + 0.006524338852614164, + 0.00779562396928668, + 0.005434931255877018, + 0.00629231845960021, + 0.005934413056820631, + 0.00649356609210372, + 0.004766776226460934, + 0.004245635122060776, + 0.0055090212263166904, + 0.006719605997204781, + 0.008838637731969357, + 0.00614494364708662, + 0.009983689524233341, + 0.007830861024558544, + 0.005231159273535013, + 0.007531799376010895, + 0.04854569956660271, + 0.014360549859702587, + 0.006892585661262274, + 0.007051009684801102, + 0.015994897112250328, + 0.04607577994465828, + 0.006574292667210102, + 0.006478252820670605, + 0.005065783858299255, + 0.004817948676645756, + 0.005274096503853798, + 0.003856812370941043, + 0.002391290618106723, + 0.0034027379006147385, + 0.004123340826481581, + 0.006649962160736322, + 0.01837148703634739, + 0.039544761180877686, + 0.010270361788570881, + 0.010192658752202988, + 0.006588438525795937, + 0.009386299178004265, + 0.006116347853094339, + 0.00664621451869607, + 0.004461304750293493, + 0.0033770152367651463, + 0.0053946892730891705, + 0.004638829734176397, + 0.0018622336210682988, + 0.0020892908796668053, + 0.0021424409933388233, + 0.0036486012395471334, + 0.028226330876350403, + 0.026425350457429886, + 0.05399700626730919, + 0.030758226290345192, + 0.017100440338253975, + 0.015284835360944271, + 0.02032249979674816, + 0.026247745379805565, + 0.027109671384096146, + 0.02168428897857666, + 0.04587984457612038, + 0.03962508589029312, + 0.02901562675833702, + 0.03261001035571098, + 0.03220311924815178, + 0.0337572917342186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.01253711897879839 + } + ], + "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/0000152/group_7/index_tensor([6286, 6286])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_tensor([6286, 6286])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cbb989ee3661c9728e3b3dda79c1974653100a5a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_tensor([6286, 6286])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a warm, Stardew Valley-esque pixelated scene, a fixed, overhead camera (a bird's-eye view) is pinpointed at the center of the scene. Atop a wood-grained tabletop composed of warm, blocky brown and tan pixels sits a minimalist, textured ceramic covered bowl constructed from ivory white pixels. At the bottom of the bowl, a round Pu'er tea cake, composed of tightly compressed dark brown, black, and ochre pixels, rests quietly, its uneven surface imbued with a handcrafted quality. Then, a stream of scalding hot water, composed of shimmering white and light blue pixels, steadily and gently pours down from the top of the frame, unerringly hitting the center of the tea cake. Almost instantly, a stream of deep ruby-red tea spreads out from the tea cake, spreading swiftly and elegantly like a drop of thick ink on rice paper. Its expanding edges display a gentle pixel gradient from crimson to amber. This rich color quickly expands with a satisfying dynamic effect, completely dyeing the clear water and finally filling the entire interior of the bowl seamlessly, forming a bowl of mellow and warm crimson pixelated tea. The whole process is imbued with a healing, peaceful and comfortable pastoral atmosphere under the soft light.", + "seed": 11493, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11493_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": -0.01209167269989848, + "vlm_logprob": -0.01209167269989848, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.01209167269989848, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02714540623128414 + ], + "opposite_logprob_tokens": [ + -4.5271453857421875 + ], + "logprob_reward": -0.02714540623128414, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02714540623128414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09033993631601334, + 0.028541652485728264, + 0.02006819099187851, + 0.08501053601503372, + 0.12252171337604523, + 0.186116024851799, + 0.2000659704208374, + 0.16809611022472382, + 0.06973929703235626, + 0.02950054034590721 + ], + "credit_map": [ + 0.011853983625769615, + 0.008608960546553135, + 0.010953188873827457, + 0.013409384526312351, + 0.012188399210572243, + 0.008016275241971016, + 0.01074210088700056, + 0.014567650854587555, + 0.003948616795241833, + 0.002859733998775482, + 0.0033281568903476, + 0.0032388383988291025, + 0.004472457338124514, + 0.0032178042456507683, + 0.002898890059441328, + 0.004577157087624073, + 0.0025069620460271835, + 0.0017193302046507597, + 0.0018279924988746643, + 0.002773952204734087, + 0.004042390733957291, + 0.0030414762441068888, + 0.0016694084042683244, + 0.002486679470166564, + 0.01491065975278616, + 0.009002589620649815, + 0.007015235722064972, + 0.00897586066275835, + 0.011362137272953987, + 0.010001986287534237, + 0.008792600594460964, + 0.01494947262108326, + 0.02342904917895794, + 0.013843361288309097, + 0.012999570928514004, + 0.012367323972284794, + 0.010529179126024246, + 0.008396142162382603, + 0.013437456451356411, + 0.027519647032022476, + 0.03939153254032135, + 0.02313505858182907, + 0.02062489278614521, + 0.04027027636766434, + 0.016249027103185654, + 0.009561626240611076, + 0.018659917637705803, + 0.018223708495497704, + 0.021630784496665, + 0.03639039769768715, + 0.015277508646249771, + 0.049522824585437775, + 0.024146810173988342, + 0.01042475737631321, + 0.025654694065451622, + 0.01701822131872177, + 0.018120666965842247, + 0.03936154767870903, + 0.016414424404501915, + 0.03368499130010605, + 0.01910506747663021, + 0.009521630592644215, + 0.017937365919351578, + 0.013950426131486893, + 0.007120397873222828, + 0.01517211552709341, + 0.007473852951079607, + 0.01519281417131424, + 0.007089862134307623, + 0.003526987275108695, + 0.008280757814645767, + 0.0058825137093663216, + 0.0035370285622775555, + 0.004236489534378052, + 0.003555067116394639, + 0.0065456898882985115, + 0.0029985408764332533, + 0.0018400211120024323, + 0.004521261900663376, + 0.0022664444986730814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.02714540623128414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329925447702408 + ], + "opposite_logprob_tokens": [ + -3.773299217224121 + ], + "logprob_reward": -0.02329925447702408, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329925447702408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08351331204175949, + 0.06850872933864594, + 0.06874999403953552, + 0.39051929116249084, + 0.09718005359172821, + 0.11827242374420166, + 0.05531501770019531, + 0.057884372770786285, + 0.018659306690096855, + 0.04139753431081772 + ], + "credit_map": [ + 0.014265724457800388, + 0.010014303959906101, + 0.006769833620637655, + 0.01240707840770483, + 0.015384194441139698, + 0.00619856221601367, + 0.008978687226772308, + 0.009494923055171967, + 0.008996431715786457, + 0.005607745610177517, + 0.005866239778697491, + 0.008180566132068634, + 0.013063258491456509, + 0.005244293250143528, + 0.00709305377677083, + 0.014457136392593384, + 0.01031617820262909, + 0.004883000161498785, + 0.004882411565631628, + 0.009123720228672028, + 0.010508631356060505, + 0.0068982564844191074, + 0.008904843591153622, + 0.013232954777777195, + 0.08759617805480957, + 0.033663127571344376, + 0.03242476284503937, + 0.03187287971377373, + 0.03494461998343468, + 0.040153343230485916, + 0.0312538743019104, + 0.09861051291227341, + 0.012921922840178013, + 0.011022764258086681, + 0.016301672905683517, + 0.01578151248395443, + 0.010827552527189255, + 0.006395920179784298, + 0.010518137365579605, + 0.013410571962594986, + 0.02901136688888073, + 0.01480770856142044, + 0.01395842619240284, + 0.02207898162305355, + 0.00815386138856411, + 0.006254824809730053, + 0.010249185375869274, + 0.013758072629570961, + 0.007948605343699455, + 0.008024120703339577, + 0.004483761731535196, + 0.011227736249566078, + 0.008314051665365696, + 0.0030886572785675526, + 0.006981635466217995, + 0.005246451590210199, + 0.006703357212245464, + 0.013221446424722672, + 0.0060759601183235645, + 0.011577649973332882, + 0.006857407279312611, + 0.00326049467548728, + 0.006188856437802315, + 0.003999198321253061, + 0.0017652343958616257, + 0.003817290998995304, + 0.0013899708865210414, + 0.003183781635016203, + 0.0022404869087040424, + 0.0015168023528531194, + 0.0021145008504390717, + 0.0026312393601983786, + 0.006464887410402298, + 0.007071912754327059, + 0.0037688615266233683, + 0.007657644338905811, + 0.005944826174527407, + 0.0020035235211253166, + 0.006284656003117561, + 0.0022012207191437483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.02329925447702408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019433676498010755 + ], + "opposite_logprob_tokens": [ + -6.251943588256836 + ], + "logprob_reward": -0.0019433676498010755, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019433676498010755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11478647589683533, + 0.04462277516722679, + 0.0602240264415741, + 0.25611764192581177, + 0.15668107569217682, + 0.06534632295370102, + 0.09491319209337234, + 0.11694911122322083, + 0.05688652768731117, + 0.03347276896238327 + ], + "credit_map": [ + 0.0214720256626606, + 0.00825433898717165, + 0.01095507387071848, + 0.011033863760530949, + 0.018666351214051247, + 0.010280001908540726, + 0.011778687126934528, + 0.022346140816807747, + 0.006029265001416206, + 0.0037288872990757227, + 0.003509705886244774, + 0.004983130842447281, + 0.010982952080667019, + 0.004266130737960339, + 0.003753244411200285, + 0.007369458209723234, + 0.006721629295498133, + 0.005304213147610426, + 0.003992617595940828, + 0.011316066607832909, + 0.010609200224280357, + 0.006409668829292059, + 0.006822412833571434, + 0.009048225358128548, + 0.0678086057305336, + 0.022530004382133484, + 0.021198106929659843, + 0.01937716454267502, + 0.01758876070380211, + 0.018247604370117188, + 0.014471324160695076, + 0.07489610463380814, + 0.043047770857810974, + 0.01650753989815712, + 0.012653031386435032, + 0.011393391527235508, + 0.009455696679651737, + 0.010614423081278801, + 0.013280120678246021, + 0.03972911089658737, + 0.020410187542438507, + 0.006995780393481255, + 0.007149450480937958, + 0.006605083122849464, + 0.003841220634058118, + 0.0038150232285261154, + 0.004871753044426441, + 0.011657829396426678, + 0.011811971664428711, + 0.018508855253458023, + 0.009835242293775082, + 0.018898488953709602, + 0.009487018920481205, + 0.005733494646847248, + 0.010726804845035076, + 0.009911322966217995, + 0.016387464478611946, + 0.022855505347251892, + 0.010188399814069271, + 0.0205693319439888, + 0.018368376418948174, + 0.005400757305324078, + 0.014305692166090012, + 0.008873584680259228, + 0.006641673389822245, + 0.013074521906673908, + 0.004527254030108452, + 0.011195800267159939, + 0.007412671111524105, + 0.003175424411892891, + 0.007064704783260822, + 0.0037944852374494076, + 0.004003417212516069, + 0.006245074328035116, + 0.0017561353743076324, + 0.006741715129464865, + 0.005492298863828182, + 0.001543198013678193, + 0.005642805714160204, + 0.002048127120360732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.0019433676498010755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013362773461267352 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013362773461267352, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013362773461267352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08981373906135559, + 0.08479039371013641, + 0.08189742267131805, + 0.14541427791118622, + 0.0690455213189125, + 0.06294010579586029, + 0.13992299139499664, + 0.23433181643486023, + 0.06433101743459702, + 0.02751273475587368 + ], + "credit_map": [ + 0.009065494872629642, + 0.00668736919760704, + 0.009610720910131931, + 0.00955041404813528, + 0.021438755095005035, + 0.009722582064568996, + 0.012236839160323143, + 0.011501572094857693, + 0.010953298769891262, + 0.00652430159971118, + 0.006449526641517878, + 0.0066963983699679375, + 0.024536335840821266, + 0.009581726975739002, + 0.008196111768484116, + 0.011852696537971497, + 0.012163270264863968, + 0.004926942754536867, + 0.005081085953861475, + 0.005937666166573763, + 0.018272029235959053, + 0.008569076657295227, + 0.009413021616637707, + 0.01753433421254158, + 0.03231341764330864, + 0.013814047910273075, + 0.010817200876772404, + 0.013961656019091606, + 0.014882768504321575, + 0.014090863056480885, + 0.01067888829857111, + 0.0348554402589798, + 0.008825000375509262, + 0.010776354931294918, + 0.006191236432641745, + 0.01054416038095951, + 0.01158542837947607, + 0.0049584293738007545, + 0.009675227105617523, + 0.006489683873951435, + 0.006200868636369705, + 0.010574845597147942, + 0.004774504341185093, + 0.02008815109729767, + 0.004605367314070463, + 0.0037125376984477043, + 0.008730177767574787, + 0.004253656603395939, + 0.014750979840755463, + 0.02361028455197811, + 0.011478891596198082, + 0.03238243609666824, + 0.016446078196167946, + 0.008336139842867851, + 0.01739710010588169, + 0.015521086752414703, + 0.027265984565019608, + 0.05222809687256813, + 0.023465832695364952, + 0.04741895571351051, + 0.027160434052348137, + 0.011936572380363941, + 0.025620324537158012, + 0.0192356389015913, + 0.005930766463279724, + 0.014725170098245144, + 0.0050560408271849155, + 0.012642877176404, + 0.008428367786109447, + 0.0038109987508505583, + 0.007542021106928587, + 0.006194777321070433, + 0.004151157569140196, + 0.00473402626812458, + 0.00314659858122468, + 0.005688587669283152, + 0.0033385155256837606, + 0.0012619299814105034, + 0.004059823229908943, + 0.001132097328081727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.0013362773461267352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734057795256376 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.006734057795256376, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734057795256376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22544527053833008, + 0.10693670064210892, + 0.036227039992809296, + 0.14827239513397217, + 0.12453651428222656, + 0.07859911769628525, + 0.045641206204891205, + 0.145567387342453, + 0.042222555726766586, + 0.046551838517189026 + ], + "credit_map": [ + 0.046411752700805664, + 0.016742676496505737, + 0.018555540591478348, + 0.018633084371685982, + 0.02840334363281727, + 0.01671256124973297, + 0.017881374806165695, + 0.06210494041442871, + 0.01989903673529625, + 0.0062697832472622395, + 0.008428535424172878, + 0.007132309954613447, + 0.022269321605563164, + 0.009815026074647903, + 0.008572378195822239, + 0.024550314992666245, + 0.006838671863079071, + 0.00334703316912055, + 0.0026958214584738016, + 0.003514098236337304, + 0.004464233759790659, + 0.0036989531945437193, + 0.0033326128032058477, + 0.008335614576935768, + 0.03260170295834541, + 0.014225629158318043, + 0.011264343746006489, + 0.013339533470571041, + 0.01540993619710207, + 0.014642122201621532, + 0.010049624368548393, + 0.03673950582742691, + 0.02976859360933304, + 0.012998444959521294, + 0.016143878921866417, + 0.017554746940732002, + 0.007904037833213806, + 0.006953367497771978, + 0.009357699193060398, + 0.023855747655034065, + 0.025006821379065514, + 0.008497070521116257, + 0.009287020191550255, + 0.0055876909755170345, + 0.00463934987783432, + 0.004914664663374424, + 0.005289322696626186, + 0.015377167612314224, + 0.005151350051164627, + 0.007364392280578613, + 0.004617257975041866, + 0.009242543019354343, + 0.0069656637497246265, + 0.0029051201418042183, + 0.0055720810778439045, + 0.0038227951154112816, + 0.017137235030531883, + 0.030031558126211166, + 0.013683716766536236, + 0.029499415308237076, + 0.01710117980837822, + 0.00794205255806446, + 0.016645340248942375, + 0.013526896014809608, + 0.005371930543333292, + 0.00858425721526146, + 0.003867201739922166, + 0.006326783914119005, + 0.006748488172888756, + 0.0025887154042720795, + 0.005324648227542639, + 0.0034105314407497644, + 0.005878821946680546, + 0.007180448155850172, + 0.0035044534597545862, + 0.008034811355173588, + 0.007664467208087444, + 0.002746633253991604, + 0.007745898794382811, + 0.003796304576098919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.006734057795256376 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Stardew Valley-esque pixelated scene, a fixed, overhead camera (a bird's-eye view) is pinpointed at the center of the scene. Atop a wood-grained tabletop composed of warm, blocky brown and tan pixels sits a minimalist, textured ceramic covered bowl constructed from ivory white pixels. At the bottom of the bowl, a round Pu'er tea cake, composed of tightly compressed dark brown, black, and ochre pixels, rests quietly, its uneven surface imbued with a handcrafted quality. Then, a stream of scalding hot water, composed of shimmering white and light blue pixels, steadily and gently pours down from the top of the frame, unerringly hitting the center of the tea cake. Almost instantly, a stream of deep ruby-red tea spreads out from the tea cake, spreading swiftly and elegantly like a drop of thick ink on rice paper. Its expanding edges display a gentle pixel gradient from crimson to amber. This rich color quickly expands with a satisfying dynamic effect, completely dyeing the clear water and finally filling the entire interior of the bowl seamlessly, forming a bowl of mellow and warm crimson pixelated tea. The whole process is imbued with a healing, peaceful and comfortable pastoral atmosphere under the soft light.", + "seed": 11494, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11494_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": -0.015952021721750496, + "vlm_logprob": -0.015952021721750496, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.015952021721750496, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016343889757990837 + ], + "opposite_logprob_tokens": [ + -5.26634407043457 + ], + "logprob_reward": -0.016343889757990837, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.016343889757990837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09372878074645996, + 0.04390880465507507, + 0.05133070796728134, + 0.03262162581086159, + 0.026778660714626312, + 0.06541357189416885, + 0.03608039766550064, + 0.1790337860584259, + 0.3481657803058624, + 0.122937873005867 + ], + "credit_map": [ + 0.011146275326609612, + 0.007373808417469263, + 0.013409409672021866, + 0.010092255659401417, + 0.010790839791297913, + 0.009744965471327305, + 0.007947513833642006, + 0.02322371117770672, + 0.00543938810005784, + 0.005436134058982134, + 0.005182535387575626, + 0.006083111744374037, + 0.00828978419303894, + 0.003606068901717663, + 0.004524351563304663, + 0.0053474316373467445, + 0.008256949484348297, + 0.004479683004319668, + 0.01052447035908699, + 0.005344767589122057, + 0.005810341332107782, + 0.0041442252695560455, + 0.00400291383266449, + 0.008767357096076012, + 0.00551578076556325, + 0.004252953454852104, + 0.006638054735958576, + 0.00223299372009933, + 0.003752640448510647, + 0.0023978387471288443, + 0.003227961715310812, + 0.0046034036204218864, + 0.003695723135024309, + 0.0032583854626864195, + 0.0040336912497878075, + 0.0026281578466296196, + 0.004286482464522123, + 0.0024187450762838125, + 0.0034714771900326014, + 0.0029859982896596193, + 0.014962132088840008, + 0.0036475921515375376, + 0.019001662731170654, + 0.003908141981810331, + 0.00544028589501977, + 0.005303399637341499, + 0.003783084684982896, + 0.009367266669869423, + 0.007420756388455629, + 0.0016451734118163586, + 0.009879058226943016, + 0.0033439798280596733, + 0.001409805496223271, + 0.0030858616810292006, + 0.002180320443585515, + 0.007115441374480724, + 0.07071051001548767, + 0.005046239588409662, + 0.08275087922811508, + 0.005658331792801619, + 0.003285605227574706, + 0.0066953967325389385, + 0.0023910137824714184, + 0.0024958301801234484, + 0.1393505185842514, + 0.009194408543407917, + 0.16415724158287048, + 0.010377084836363792, + 0.005358740221709013, + 0.010715527459979057, + 0.003681726986542344, + 0.005330539774149656, + 0.04577184095978737, + 0.003117835381999612, + 0.05445293337106705, + 0.005081992130726576, + 0.001723508001305163, + 0.0043925209902226925, + 0.002321847714483738, + 0.006075396202504635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.016343889757990837 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430946871638298 + ], + "opposite_logprob_tokens": [ + -3.1680946350097656 + ], + "logprob_reward": -0.0430946871638298, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430946871638298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02986987680196762, + 0.05022556334733963, + 0.07000812888145447, + 0.028008682653307915, + 0.027642957866191864, + 0.054361164569854736, + 0.0482313372194767, + 0.1561049520969391, + 0.4918839931488037, + 0.04366334155201912 + ], + "credit_map": [ + 0.003552972339093685, + 0.0029778715688735247, + 0.003627775702625513, + 0.003097556997090578, + 0.004982144106179476, + 0.0035425883252173662, + 0.0035327342338860035, + 0.004556231200695038, + 0.00642114644870162, + 0.005798359867185354, + 0.005936950910836458, + 0.007490009069442749, + 0.008122896775603294, + 0.004391100723296404, + 0.005589228123426437, + 0.006475865375250578, + 0.010282708331942558, + 0.007814044132828712, + 0.008422397077083588, + 0.014960146509110928, + 0.00902337022125721, + 0.006524552591145039, + 0.008894400671124458, + 0.004086505156010389, + 0.004449302796274424, + 0.003146435832604766, + 0.004540412221103907, + 0.002218385459855199, + 0.005057299509644508, + 0.00244896300137043, + 0.003146190196275711, + 0.0030016896780580282, + 0.0032032770104706287, + 0.004060924518853426, + 0.0034687526058405638, + 0.00248417304828763, + 0.005625417921692133, + 0.002542905742302537, + 0.003460814245045185, + 0.002796690911054611, + 0.011833151802420616, + 0.002728652209043503, + 0.016243694350123405, + 0.002581873908638954, + 0.004676894750446081, + 0.0043963720090687275, + 0.002305783797055483, + 0.009594733826816082, + 0.011709900572896004, + 0.0016741071594879031, + 0.01547747477889061, + 0.001985743409022689, + 0.0014788289554417133, + 0.004041164647787809, + 0.0015020782593637705, + 0.010362036526203156, + 0.06048719212412834, + 0.004267971031367779, + 0.07111344486474991, + 0.0050244382582604885, + 0.0038540007080882788, + 0.00579419732093811, + 0.0026907767169177532, + 0.002872925018891692, + 0.19543233513832092, + 0.013234570622444153, + 0.2302003800868988, + 0.016535433009266853, + 0.007470989134162664, + 0.01544201746582985, + 0.005884891375899315, + 0.007683340925723314, + 0.015748601406812668, + 0.0012048614444211125, + 0.018516013398766518, + 0.0014105403097346425, + 0.0014615835389122367, + 0.0018497195560485125, + 0.0009703973773866892, + 0.0025016185827553272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.0430946871638298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742938421666622 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006742938421666622, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742938421666622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032102636992931366, + 0.026968205347657204, + 0.041619110852479935, + 0.02161547727882862, + 0.05074537917971611, + 0.06504669785499573, + 0.16080895066261292, + 0.07070300728082657, + 0.45611488819122314, + 0.07427552342414856 + ], + "credit_map": [ + 0.005075254011899233, + 0.003054906614124775, + 0.004040357191115618, + 0.0043632760643959045, + 0.006227894686162472, + 0.002823661081492901, + 0.003320259042084217, + 0.0031970336567610502, + 0.0027797305956482887, + 0.0022628617007285357, + 0.003619623836129904, + 0.0033976545091718435, + 0.00383486645296216, + 0.0027166299987584352, + 0.0027796330396085978, + 0.005577208008617163, + 0.005719900131225586, + 0.0025981604121625423, + 0.008227597922086716, + 0.0051614996045827866, + 0.003975681494921446, + 0.004180973395705223, + 0.003628032049164176, + 0.008127274923026562, + 0.003044923534616828, + 0.0024786265566945076, + 0.00458178948611021, + 0.0017033405601978302, + 0.0021843721624463797, + 0.001952938036993146, + 0.0019758129492402077, + 0.0036936781834810972, + 0.005548165179789066, + 0.007658079266548157, + 0.005774160381406546, + 0.0056037153117358685, + 0.010596602223813534, + 0.004698662087321281, + 0.006269113160669804, + 0.0045968880876898766, + 0.013891042210161686, + 0.004338390659540892, + 0.01623621955513954, + 0.004158801399171352, + 0.009656337089836597, + 0.00508492486551404, + 0.005149881821125746, + 0.006531110964715481, + 0.04242122918367386, + 0.004044144414365292, + 0.05496221408247948, + 0.005295542534440756, + 0.0034375039394944906, + 0.013256698846817017, + 0.004303139168769121, + 0.033088501542806625, + 0.02529423125088215, + 0.0024081713054329157, + 0.029963605105876923, + 0.002955172210931778, + 0.0029332444537431, + 0.0028724430594593287, + 0.0023114525247365236, + 0.001964699011296034, + 0.18281055986881256, + 0.012024334631860256, + 0.21510231494903564, + 0.014186693355441093, + 0.00618156511336565, + 0.014321047812700272, + 0.004854958970099688, + 0.006633487530052662, + 0.027607880532741547, + 0.0018428033217787743, + 0.03292363882064819, + 0.0031121200881898403, + 0.0012231218861415982, + 0.002631374401971698, + 0.0013540370855480433, + 0.003580559743568301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.006742938421666622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949525628238916 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949525628238916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949525628238916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18126864731311798, + 0.04186198487877846, + 0.02765706740319729, + 0.04893660545349121, + 0.06688924133777618, + 0.047011811286211014, + 0.046307746320962906, + 0.09421304613351822, + 0.4160638153553009, + 0.02979006990790367 + ], + "credit_map": [ + 0.018812546506524086, + 0.014360730536282063, + 0.022896558046340942, + 0.014390824362635612, + 0.036827076226472855, + 0.019259661436080933, + 0.015677494928240776, + 0.03904375806450844, + 0.005912187974900007, + 0.0032131802290678024, + 0.005659330170601606, + 0.005247726570814848, + 0.006850834470242262, + 0.003726159455254674, + 0.004648495931178331, + 0.0066040693782269955, + 0.004164095502346754, + 0.0025371257215738297, + 0.005620442796498537, + 0.0022144552785903215, + 0.004721122328191996, + 0.0023187901824712753, + 0.0021501434966921806, + 0.0039308941923081875, + 0.008776389993727207, + 0.004341797437518835, + 0.007861406542360783, + 0.0053330580703914165, + 0.009780379943549633, + 0.005343570373952389, + 0.004662174731492996, + 0.0028378323186188936, + 0.011670047417283058, + 0.008102317340672016, + 0.014172016642987728, + 0.009256967343389988, + 0.007086044643074274, + 0.004684573970735073, + 0.005363170523196459, + 0.006554102525115013, + 0.007952747866511345, + 0.0025368540082126856, + 0.013085291720926762, + 0.0022677097003906965, + 0.0025851333048194647, + 0.004493129439651966, + 0.0021235349122434855, + 0.011967414058744907, + 0.013075103051960468, + 0.0013630668399855494, + 0.01623653993010521, + 0.0014176182448863983, + 0.0012667591217905283, + 0.003348371246829629, + 0.0011831214651465416, + 0.008417165838181973, + 0.03625331073999405, + 0.0025552960578352213, + 0.042409371584653854, + 0.003367279889062047, + 0.0024367780424654484, + 0.0032640499994158745, + 0.0017843256937339902, + 0.0021426426246762276, + 0.16604425013065338, + 0.011247062124311924, + 0.1951218992471695, + 0.014294267632067204, + 0.005321229808032513, + 0.013141877949237823, + 0.004874767269939184, + 0.006018449552357197, + 0.00944106187671423, + 0.0011825114488601685, + 0.011420788243412971, + 0.002409542677924037, + 0.0014039638917893171, + 0.0010892458958551288, + 0.0011792656732723117, + 0.0016636926447972655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.005949525628238916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762906763702631 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.00762906763702631, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762906763702631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02046777680516243, + 0.010053375735878944, + 0.020145233720541, + 0.010875634849071503, + 0.009961982257664204, + 0.08051218092441559, + 0.6178457140922546, + 0.09031409025192261, + 0.04448242858052254, + 0.09534156322479248 + ], + "credit_map": [ + 0.003013842971995473, + 0.0019410591339692473, + 0.0031146013643592596, + 0.0017993324436247349, + 0.003240819787606597, + 0.0023247471544891596, + 0.002065399195998907, + 0.0029679767321795225, + 0.0012228533159941435, + 0.0009041171870194376, + 0.001436165883205831, + 0.0013123759999871254, + 0.000928474182728678, + 0.0009250073926523328, + 0.0011153019731864333, + 0.0022090806160122156, + 0.0035426404792815447, + 0.001793665811419487, + 0.004123340826481581, + 0.002227475168183446, + 0.0022328421473503113, + 0.0017988620093092322, + 0.0015880095306783915, + 0.002838400425389409, + 0.0011465342249721289, + 0.0017043956322595477, + 0.0012284730328246951, + 0.0012678994098678231, + 0.0021016632672399282, + 0.0011355915339663625, + 0.001344835851341486, + 0.0009462429443374276, + 0.001554610556922853, + 0.0008242267649620771, + 0.0019735016394406557, + 0.0008621914312243462, + 0.0017341942293569446, + 0.0008180056465789676, + 0.0010775268310680985, + 0.001117726438678801, + 0.017867188900709152, + 0.003957969602197409, + 0.024495109915733337, + 0.0035382674541324377, + 0.005213027819991112, + 0.006366869900375605, + 0.002931175520643592, + 0.016142582520842552, + 0.16315321624279022, + 0.015838604420423508, + 0.2092265784740448, + 0.017893943935632706, + 0.01153823733329773, + 0.05112326145172119, + 0.01528333593159914, + 0.133788600564003, + 0.03432434797286987, + 0.002392688300460577, + 0.03998956084251404, + 0.0031787152402102947, + 0.0027964480686932802, + 0.003297880059108138, + 0.0017917099175974727, + 0.0025427532382309437, + 0.01319186296314001, + 0.0019455336732789874, + 0.014781870879232883, + 0.004414754919707775, + 0.003023111494258046, + 0.0018042895244434476, + 0.0023012161254882812, + 0.003019796684384346, + 0.03497476875782013, + 0.002545545808970928, + 0.041402485221624374, + 0.003039223374798894, + 0.0023615299724042416, + 0.004202762618660927, + 0.001618997543118894, + 0.005196265876293182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.00762906763702631 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Stardew Valley-esque pixelated scene, a fixed, overhead camera (a bird's-eye view) is pinpointed at the center of the scene. Atop a wood-grained tabletop composed of warm, blocky brown and tan pixels sits a minimalist, textured ceramic covered bowl constructed from ivory white pixels. At the bottom of the bowl, a round Pu'er tea cake, composed of tightly compressed dark brown, black, and ochre pixels, rests quietly, its uneven surface imbued with a handcrafted quality. Then, a stream of scalding hot water, composed of shimmering white and light blue pixels, steadily and gently pours down from the top of the frame, unerringly hitting the center of the tea cake. Almost instantly, a stream of deep ruby-red tea spreads out from the tea cake, spreading swiftly and elegantly like a drop of thick ink on rice paper. Its expanding edges display a gentle pixel gradient from crimson to amber. This rich color quickly expands with a satisfying dynamic effect, completely dyeing the clear water and finally filling the entire interior of the bowl seamlessly, forming a bowl of mellow and warm crimson pixelated tea. The whole process is imbued with a healing, peaceful and comfortable pastoral atmosphere under the soft light.", + "seed": 11495, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11495_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": -0.014111360628157854, + "vlm_logprob": -0.014111360628157854, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.014111360628157854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03304494544863701 + ], + "opposite_logprob_tokens": [ + -4.658044815063477 + ], + "logprob_reward": -0.03304494544863701, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.03304494544863701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06470569968223572, + 0.032917797565460205, + 0.0284073855727911, + 0.08316735923290253, + 0.0733213871717453, + 0.21514251828193665, + 0.30168554186820984, + 0.024615030735731125, + 0.05511358380317688, + 0.12092365324497223 + ], + "credit_map": [ + 0.011982972733676434, + 0.009045972488820553, + 0.007912633940577507, + 0.009480186738073826, + 0.007345696445554495, + 0.004993544891476631, + 0.004905050620436668, + 0.009039643220603466, + 0.005833227187395096, + 0.0035341286566108465, + 0.004772047977894545, + 0.0030722953379154205, + 0.005235348362475634, + 0.004138553515076637, + 0.002338527934625745, + 0.003993667662143707, + 0.0056840283796191216, + 0.0032155218068510294, + 0.0037198432255536318, + 0.0033369262237101793, + 0.0023198816925287247, + 0.002888372866436839, + 0.0035116062499582767, + 0.003731203731149435, + 0.01031701173633337, + 0.00992287416011095, + 0.010947445407509804, + 0.008217213675379753, + 0.014737465418875217, + 0.008664276450872421, + 0.00839727371931076, + 0.011963802389800549, + 0.012823591008782387, + 0.007848421111702919, + 0.01162520982325077, + 0.007689727004617453, + 0.014855471439659595, + 0.003691213671118021, + 0.0058654495514929295, + 0.008922298438847065, + 0.04294825717806816, + 0.013867553323507309, + 0.0281408429145813, + 0.022163230925798416, + 0.035443391650915146, + 0.02122381143271923, + 0.028624964877963066, + 0.022730469703674316, + 0.05116051062941551, + 0.046753522008657455, + 0.03795018419623375, + 0.03469694033265114, + 0.06431979686021805, + 0.021922526881098747, + 0.02174176461994648, + 0.023140298202633858, + 0.003350724931806326, + 0.005374804604798555, + 0.0030657039023935795, + 0.002251099096611142, + 0.005528950598090887, + 0.0018472885712981224, + 0.001348369987681508, + 0.0018480870639905334, + 0.008970499970018864, + 0.007844701409339905, + 0.0079886419698596, + 0.007417622022330761, + 0.007823766209185123, + 0.0040312400087714195, + 0.005938578862696886, + 0.005098529625684023, + 0.049737416207790375, + 0.005866030231118202, + 0.043539922684431076, + 0.0045392317697405815, + 0.00508579146116972, + 0.004173325840383768, + 0.003363208845257759, + 0.004618737380951643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.03304494544863701 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318471387028694 + ], + "opposite_logprob_tokens": [ + -3.7733185291290283 + ], + "logprob_reward": -0.023318471387028694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318471387028694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04279780387878418, + 0.0880788266658783, + 0.05210769549012184, + 0.11472392082214355, + 0.10423459112644196, + 0.1613219827413559, + 0.2646694779396057, + 0.05741264671087265, + 0.05377345532178879, + 0.06087963283061981 + ], + "credit_map": [ + 0.007368927821516991, + 0.005215243436396122, + 0.006080167833715677, + 0.0045707933604717255, + 0.0037832956295460463, + 0.003930731676518917, + 0.0060089509934186935, + 0.005839687306433916, + 0.012687841430306435, + 0.00862216204404831, + 0.012536095455288887, + 0.02100467123091221, + 0.0078102052211761475, + 0.0052264719270169735, + 0.008705844171345234, + 0.011485526338219643, + 0.009239684790372849, + 0.006886570248752832, + 0.008823633193969727, + 0.010066032409667969, + 0.003286329796537757, + 0.0034303420688956976, + 0.005937271285802126, + 0.004437833558768034, + 0.01201650034636259, + 0.013369087129831314, + 0.01332626212388277, + 0.015794018283486366, + 0.017831292003393173, + 0.011822567321360111, + 0.01208733394742012, + 0.01847684569656849, + 0.012162433005869389, + 0.006230971310287714, + 0.012730064801871777, + 0.009977081790566444, + 0.013094660826027393, + 0.014869914390146732, + 0.01019591186195612, + 0.024973539635539055, + 0.03272610902786255, + 0.010293528437614441, + 0.019829707220196724, + 0.016642291098833084, + 0.02558182179927826, + 0.017594190314412117, + 0.020295852795243263, + 0.018358474597334862, + 0.04308333247900009, + 0.04273585230112076, + 0.032454635947942734, + 0.02974723093211651, + 0.057547274976968765, + 0.01943211443722248, + 0.01923971064388752, + 0.02042931318283081, + 0.008201553486287594, + 0.009306816384196281, + 0.0069058737717568874, + 0.006577624008059502, + 0.010168105363845825, + 0.007833227515220642, + 0.003373953513801098, + 0.005045483820140362, + 0.005982998758554459, + 0.008118550293147564, + 0.006288027856498957, + 0.007007816806435585, + 0.010126861743628979, + 0.004654993303120136, + 0.006662978325039148, + 0.0049312240444123745, + 0.019454028457403183, + 0.006469094660133123, + 0.017473207786679268, + 0.0037948135286569595, + 0.005884809419512749, + 0.0024463783483952284, + 0.002565080067142844, + 0.0027922112494707108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.023318471387028694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259012337774038 + ], + "opposite_logprob_tokens": [ + -5.255259037017822 + ], + "logprob_reward": -0.005259012337774038, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259012337774038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05805172026157379, + 0.1064162477850914, + 0.036202482879161835, + 0.07233629375696182, + 0.09376226365566254, + 0.16122165322303772, + 0.2634936273097992, + 0.0353778637945652, + 0.03129822015762329, + 0.1418396383523941 + ], + "credit_map": [ + 0.010252103209495544, + 0.007073282264173031, + 0.008655393496155739, + 0.009834672324359417, + 0.006637886632233858, + 0.0045288424007594585, + 0.0054086255840957165, + 0.0056609162129461765, + 0.016011245548725128, + 0.010219489224255085, + 0.016174426302313805, + 0.02431613765656948, + 0.010096532292664051, + 0.006530448794364929, + 0.010166282765567303, + 0.01290169171988964, + 0.00592687726020813, + 0.004098409786820412, + 0.005099359434098005, + 0.007817136123776436, + 0.0018540500896051526, + 0.0026920405216515064, + 0.0033162657637149096, + 0.005398344714194536, + 0.013840020634233952, + 0.006705016363412142, + 0.01409098505973816, + 0.015349569730460644, + 0.005395675078034401, + 0.003040364244952798, + 0.008699217811226845, + 0.005215437617152929, + 0.015338597819209099, + 0.011654548346996307, + 0.014805436134338379, + 0.012070314027369022, + 0.014725500717759132, + 0.0057328203693032265, + 0.006039851810783148, + 0.013395199552178383, + 0.032336506992578506, + 0.011274825781583786, + 0.019031904637813568, + 0.017208637669682503, + 0.02178361639380455, + 0.01950162462890148, + 0.01849057711660862, + 0.021593960002064705, + 0.04346197843551636, + 0.0422653965651989, + 0.032507482916116714, + 0.029808087274432182, + 0.056937284767627716, + 0.019655568525195122, + 0.018442302942276, + 0.020415518432855606, + 0.005037875846028328, + 0.004493981134146452, + 0.003653008956462145, + 0.0030363716650754213, + 0.006077835336327553, + 0.005318888463079929, + 0.003069709986448288, + 0.004690189380198717, + 0.004602898843586445, + 0.004664313048124313, + 0.0038489799480885267, + 0.004520633723586798, + 0.004755526315420866, + 0.002549063879996538, + 0.0034515170846134424, + 0.0029052894096821547, + 0.05053256079554558, + 0.009526791051030159, + 0.043989866971969604, + 0.009812840260565281, + 0.010644249618053436, + 0.004906048998236656, + 0.00724261999130249, + 0.0051846555434167385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.005259012337774038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197947818785906 + ], + "opposite_logprob_tokens": [ + -6.127197742462158 + ], + "logprob_reward": -0.002197947818785906, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197947818785906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06379353255033493, + 0.0379524901509285, + 0.034871120005846024, + 0.06815754622220993, + 0.10490910708904266, + 0.262438029050827, + 0.39169982075691223, + 0.007778220809996128, + 0.012203386053442955, + 0.01619676873087883 + ], + "credit_map": [ + 0.014264649711549282, + 0.005973543506115675, + 0.014821787364780903, + 0.009212415665388107, + 0.005419939290732145, + 0.004631039220839739, + 0.00451215635985136, + 0.00495799258351326, + 0.0058725373819470406, + 0.003838882315903902, + 0.005790277384221554, + 0.004225112497806549, + 0.006329130847007036, + 0.00527177145704627, + 0.002666610525920987, + 0.003958167042583227, + 0.005901076830923557, + 0.003413067664951086, + 0.004144925158470869, + 0.003924576565623283, + 0.0032633126247674227, + 0.0030387085862457752, + 0.006462619639933109, + 0.004722831305116415, + 0.009848788380622864, + 0.007243085652589798, + 0.010866843163967133, + 0.008500242605805397, + 0.011681748554110527, + 0.006371886935085058, + 0.005469061899930239, + 0.00817588809877634, + 0.018192337825894356, + 0.011959967203438282, + 0.015668675303459167, + 0.011497722938656807, + 0.02083500102162361, + 0.006228405516594648, + 0.008820770308375359, + 0.011706225574016571, + 0.050909373909235, + 0.016020290553569794, + 0.03200983256101608, + 0.028252622112631798, + 0.04054608196020126, + 0.02938072197139263, + 0.033633507788181305, + 0.031685564666986465, + 0.06342993676662445, + 0.06319981068372726, + 0.0481191910803318, + 0.04469362273812294, + 0.08485379070043564, + 0.029333628714084625, + 0.027834486216306686, + 0.030235348269343376, + 0.0015368429012596607, + 0.0009703048854134977, + 0.0010143356630578637, + 0.0010230905609205365, + 0.001027383143082261, + 0.0009950085077434778, + 0.0004754028341267258, + 0.0007358522852882743, + 0.0017861862434074283, + 0.0015755549538880587, + 0.002075106604024768, + 0.0015329705784097314, + 0.0016589416190981865, + 0.001182767329737544, + 0.001432083547115326, + 0.0009597752359695733, + 0.003956010565161705, + 0.0018322458490729332, + 0.003580438904464245, + 0.0013164408737793565, + 0.002352598123252392, + 0.0007404734496958554, + 0.001264597405679524, + 0.0011539626866579056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.002197947818785906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05566302686929703, + 0.03882044926285744, + 0.030768727883696556, + 0.08252452313899994, + 0.07083236426115036, + 0.2602927088737488, + 0.39200109243392944, + 0.0267587099224329, + 0.0104126688092947, + 0.03192577138543129 + ], + "credit_map": [ + 0.010028377175331116, + 0.006674254313111305, + 0.010070206597447395, + 0.009671374224126339, + 0.0060641830787062645, + 0.0038546635769307613, + 0.0036885139998048544, + 0.0056114536710083485, + 0.005782637745141983, + 0.00416890112683177, + 0.0055523524060845375, + 0.006204559467732906, + 0.0050872513093054295, + 0.0036780345253646374, + 0.0035851681604981422, + 0.004761544056236744, + 0.004874111618846655, + 0.00375865725800395, + 0.003683367045596242, + 0.0030558831058442593, + 0.005153401754796505, + 0.003169408766552806, + 0.003969937562942505, + 0.003103962168097496, + 0.012518052943050861, + 0.007493030279874802, + 0.014242459088563919, + 0.01668165996670723, + 0.008784673176705837, + 0.006651642732322216, + 0.006014854647219181, + 0.010138150304555893, + 0.011816822923719883, + 0.005310596898198128, + 0.009647708386182785, + 0.008777412585914135, + 0.011270974762737751, + 0.005327693652361631, + 0.005094811320304871, + 0.01358633954077959, + 0.05108267813920975, + 0.01667587272822857, + 0.03147629275918007, + 0.02785581722855568, + 0.039105843752622604, + 0.02954639308154583, + 0.032873962074518204, + 0.031675852835178375, + 0.0632501170039177, + 0.06388551741838455, + 0.04812368378043175, + 0.04480880871415138, + 0.08490318804979324, + 0.028804367408156395, + 0.028114276006817818, + 0.030111119151115417, + 0.003787315683439374, + 0.0033690305426716805, + 0.003072305116802454, + 0.002404404105618596, + 0.005459503270685673, + 0.0034098541364073753, + 0.0020272487308830023, + 0.0032290504314005375, + 0.001720120431855321, + 0.0020010564476251602, + 0.0009939295705407858, + 0.0010970385046675801, + 0.0018986540380865335, + 0.0008899988606572151, + 0.0007778030703775585, + 0.0010340686421841383, + 0.012375619262456894, + 0.0021335911005735397, + 0.010743385180830956, + 0.0011652972316369414, + 0.002138970885425806, + 0.0013285912573337555, + 0.0008417586795985699, + 0.0011985603487119079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.006736426148563623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Stardew Valley-esque pixelated scene, a fixed, overhead camera (a bird's-eye view) is pinpointed at the center of the scene. Atop a wood-grained tabletop composed of warm, blocky brown and tan pixels sits a minimalist, textured ceramic covered bowl constructed from ivory white pixels. At the bottom of the bowl, a round Pu'er tea cake, composed of tightly compressed dark brown, black, and ochre pixels, rests quietly, its uneven surface imbued with a handcrafted quality. Then, a stream of scalding hot water, composed of shimmering white and light blue pixels, steadily and gently pours down from the top of the frame, unerringly hitting the center of the tea cake. Almost instantly, a stream of deep ruby-red tea spreads out from the tea cake, spreading swiftly and elegantly like a drop of thick ink on rice paper. Its expanding edges display a gentle pixel gradient from crimson to amber. This rich color quickly expands with a satisfying dynamic effect, completely dyeing the clear water and finally filling the entire interior of the bowl seamlessly, forming a bowl of mellow and warm crimson pixelated tea. The whole process is imbued with a healing, peaceful and comfortable pastoral atmosphere under the soft light.", + "seed": 11496, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_7/index_6286_seed_11496_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": -0.008666487177833915, + "vlm_logprob": -0.008666487177833915, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.008666487177833915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02202339470386505 + ], + "opposite_logprob_tokens": [ + -5.1470232009887695 + ], + "logprob_reward": -0.02202339470386505, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.02202339470386505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12990796566009521, + 0.10764914751052856, + 0.0814538523554802, + 0.10124284029006958, + 0.21932806074619293, + 0.18522214889526367, + 0.0369284488260746, + 0.05046412721276283, + 0.03255308419466019, + 0.055250342935323715 + ], + "credit_map": [ + 0.02123469114303589, + 0.013045837171375751, + 0.01329098828136921, + 0.01310418825596571, + 0.013066143728792667, + 0.016978852450847626, + 0.014168462716042995, + 0.025018788874149323, + 0.022533364593982697, + 0.009608471766114235, + 0.011410074308514595, + 0.008799261413514614, + 0.014027060940861702, + 0.026177728548645973, + 0.005434995982795954, + 0.00965818576514721, + 0.012468449771404266, + 0.0073104132898151875, + 0.009151806123554707, + 0.006773422006517649, + 0.014549397863447666, + 0.012235188856720924, + 0.008382345549762249, + 0.010582830756902695, + 0.017896834760904312, + 0.00677031883969903, + 0.009942016564309597, + 0.014857250265777111, + 0.007367828395217657, + 0.006494589149951935, + 0.018400557339191437, + 0.019513443112373352, + 0.055667031556367874, + 0.015915662050247192, + 0.017526181414723396, + 0.020165154710412025, + 0.007122045382857323, + 0.014983258210122585, + 0.023345738649368286, + 0.06460297107696533, + 0.04246741533279419, + 0.014608772471547127, + 0.0205017551779747, + 0.023918744176626205, + 0.008840889669954777, + 0.02687707170844078, + 0.030588902533054352, + 0.017418596893548965, + 0.00844920240342617, + 0.002690291963517666, + 0.003035386558622122, + 0.005889161489903927, + 0.0025249114260077477, + 0.004871688317507505, + 0.004885616712272167, + 0.004582191817462444, + 0.010161656886339188, + 0.003635646542534232, + 0.007179667241871357, + 0.009042386896908283, + 0.0027346883434802294, + 0.006382374558597803, + 0.007829333655536175, + 0.003498368663713336, + 0.005609617568552494, + 0.003069898346439004, + 0.004714036360383034, + 0.004088074434548616, + 0.0025267607998102903, + 0.00491742929443717, + 0.005127457901835442, + 0.002499809954315424, + 0.007787606213241816, + 0.003989838529378176, + 0.00691691879183054, + 0.009045314975082874, + 0.0037440473679453135, + 0.009875673800706863, + 0.01070643775165081, + 0.0031845062039792538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02202339470386505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558073736727238 + ], + "opposite_logprob_tokens": [ + -4.3875579833984375 + ], + "logprob_reward": -0.012558073736727238, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558073736727238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1291445791721344, + 0.048372525721788406, + 0.07378113269805908, + 0.04139361158013344, + 0.3111969232559204, + 0.11124354600906372, + 0.1300232708454132, + 0.06161896884441376, + 0.030448274686932564, + 0.06277716159820557 + ], + "credit_map": [ + 0.027761759236454964, + 0.020120400935411453, + 0.014848959632217884, + 0.01575060933828354, + 0.013577744364738464, + 0.008980024605989456, + 0.016600677743554115, + 0.011504399590194225, + 0.005976671352982521, + 0.005920489318668842, + 0.00532950833439827, + 0.006671166978776455, + 0.008909312076866627, + 0.003975595813244581, + 0.005850952118635178, + 0.00573883019387722, + 0.010959606617689133, + 0.0071950345300138, + 0.008909553289413452, + 0.00875375047326088, + 0.009806028567254543, + 0.011758746579289436, + 0.0079265795648098, + 0.008471834473311901, + 0.0064909737557172775, + 0.0030311602167785168, + 0.005688262637704611, + 0.006428546272218227, + 0.0037055269349366426, + 0.0034572570584714413, + 0.008703514002263546, + 0.003888368373736739, + 0.06432007253170013, + 0.0342479944229126, + 0.021428991109132767, + 0.04518662020564079, + 0.017390772700309753, + 0.0222675409168005, + 0.042613666504621506, + 0.06374126672744751, + 0.0166202150285244, + 0.009691364131867886, + 0.010960853658616543, + 0.018335111439228058, + 0.004896469414234161, + 0.01415175385773182, + 0.027157459408044815, + 0.009430318139493465, + 0.0261631291359663, + 0.01690167933702469, + 0.011864099651575089, + 0.017973149195313454, + 0.009697764180600643, + 0.013328402303159237, + 0.019772453233599663, + 0.014322583563625813, + 0.016992712393403053, + 0.004529083147644997, + 0.008053192868828773, + 0.0067347511649131775, + 0.00424992386251688, + 0.008576535619795322, + 0.006715461611747742, + 0.0057673146948218346, + 0.00536381546407938, + 0.00207911035977304, + 0.004400685429573059, + 0.0055642020888626575, + 0.001970556564629078, + 0.003316356334835291, + 0.005111048463732004, + 0.0026424997486174107, + 0.011716648004949093, + 0.004771821666508913, + 0.010573594830930233, + 0.008796227164566517, + 0.00410076230764389, + 0.008223500102758408, + 0.00987622607499361, + 0.004718377720564604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.012558073736727238 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036180056631565094 + ], + "opposite_logprob_tokens": [ + -5.628618240356445 + ], + "logprob_reward": -0.0036180056631565094, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036180056631565094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1289079338312149, + 0.06689359992742538, + 0.1019793227314949, + 0.09688077867031097, + 0.15587960183620453, + 0.2241426557302475, + 0.05056418105959892, + 0.06734693050384521, + 0.02567291632294655, + 0.08173207193613052 + ], + "credit_map": [ + 0.015374458394944668, + 0.016173597425222397, + 0.013761786743998528, + 0.01441999152302742, + 0.021133292466402054, + 0.022483784705400467, + 0.015881763771176338, + 0.009679264388978481, + 0.0077247838489711285, + 0.008132467046380043, + 0.007826454006135464, + 0.010619280859827995, + 0.012326481752097607, + 0.006647774018347263, + 0.006280132569372654, + 0.007336225360631943, + 0.016565576195716858, + 0.007677214685827494, + 0.015138112008571625, + 0.013411191292107105, + 0.010751941241323948, + 0.01577826403081417, + 0.010730082169175148, + 0.011926942504942417, + 0.013193357735872269, + 0.0062024095095694065, + 0.009841983206570148, + 0.014727099798619747, + 0.011730912141501904, + 0.007529604248702526, + 0.022345494478940964, + 0.011309931054711342, + 0.030577311292290688, + 0.013228284195065498, + 0.011541995219886303, + 0.021646887063980103, + 0.007886282168328762, + 0.011642690747976303, + 0.023483505472540855, + 0.035872649401426315, + 0.056232061237096786, + 0.016952956095337868, + 0.0240121278911829, + 0.030913399532437325, + 0.010585160925984383, + 0.028833182528614998, + 0.03904473036527634, + 0.017569053918123245, + 0.009723156690597534, + 0.003734994912520051, + 0.00752505986019969, + 0.005875878501683474, + 0.003798590274527669, + 0.010817171074450016, + 0.005131605546921492, + 0.003957726061344147, + 0.015473111532628536, + 0.00535342562943697, + 0.011369633488357067, + 0.007571007590740919, + 0.003987546544522047, + 0.010445429012179375, + 0.006457271985709667, + 0.00668951403349638, + 0.003478832310065627, + 0.0028267656452953815, + 0.003217332763597369, + 0.003573325462639332, + 0.001968354219570756, + 0.004846212919801474, + 0.00352047779597342, + 0.0022416175343096256, + 0.014219743199646473, + 0.006557073909789324, + 0.010214170441031456, + 0.01317974179983139, + 0.005522067192941904, + 0.012217761017382145, + 0.015246372669935226, + 0.004575149621814489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.0036180056631565094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019407500512897968 + ], + "opposite_logprob_tokens": [ + -6.251940727233887 + ], + "logprob_reward": -0.0019407500512897968, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019407500512897968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09686342626810074, + 0.10708018392324448, + 0.026477888226509094, + 0.13820604979991913, + 0.13804751634597778, + 0.20189985632896423, + 0.20565097033977509, + 0.027023902162909508, + 0.03155672922730446, + 0.027193531394004822 + ], + "credit_map": [ + 0.019997488707304, + 0.010225998237729073, + 0.01124007347971201, + 0.007488483563065529, + 0.010327570140361786, + 0.022231921553611755, + 0.008325982838869095, + 0.007025897037237883, + 0.011223336681723595, + 0.012697121128439903, + 0.009891391731798649, + 0.012269017286598682, + 0.026997989043593407, + 0.015640979632735252, + 0.009810619056224823, + 0.008549720980226994, + 0.004625008907169104, + 0.0022657597437500954, + 0.003224499523639679, + 0.003985108342021704, + 0.003165061352774501, + 0.004102976992726326, + 0.0023796872701495886, + 0.0027297823689877987, + 0.023367146030068398, + 0.009133282117545605, + 0.017162777483463287, + 0.017352798953652382, + 0.018032746389508247, + 0.016235360875725746, + 0.01926487870514393, + 0.017657049000263214, + 0.022996686398983, + 0.019877029582858086, + 0.01239654328674078, + 0.02143099345266819, + 0.009844941087067127, + 0.010865528136491776, + 0.018256952986121178, + 0.0223788321018219, + 0.036267295479774475, + 0.019847100600600243, + 0.023117875680327415, + 0.027705971151590347, + 0.011230530217289925, + 0.0282101072371006, + 0.03493352234363556, + 0.020587434992194176, + 0.046376537531614304, + 0.027699584141373634, + 0.02219174988567829, + 0.021642111241817474, + 0.017592916265130043, + 0.028796102851629257, + 0.020927447825670242, + 0.02042449824512005, + 0.005822289269417524, + 0.0021746924612671137, + 0.003910471219569445, + 0.003679777029901743, + 0.001937903929501772, + 0.003970818128436804, + 0.003144646529108286, + 0.002383301965892315, + 0.007185501512140036, + 0.0023881623055785894, + 0.0036140764132142067, + 0.005040363874286413, + 0.002700583077967167, + 0.0038357044104486704, + 0.004153334069997072, + 0.0026390007697045803, + 0.004776559304445982, + 0.0024308953434228897, + 0.003241896163672209, + 0.0041818758472800255, + 0.001985355047509074, + 0.0036156941205263138, + 0.005182908847928047, + 0.0017783419461920857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.0019407500512897968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera view fixed and overhead (bird's-eye view) at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a minimalist textured ceramic bowl made of ivory white pixels sitting on a wood-grained tabletop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a round Pu'er tea cake composed of dark brown, black, and ochre pixels resting inside the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does hot water pour from the top of the frame onto the tea cake, followed by deep ruby-red tea spreading out to fill the bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style pixelated with a warm, Stardew Valley-esque aesthetic?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1386573761701584, + 0.10536396503448486, + 0.12871325016021729, + 0.06037188693881035, + 0.13769756257534027, + 0.07462155073881149, + 0.15172238647937775, + 0.050959959626197815, + 0.04132171720266342, + 0.11057025194168091 + ], + "credit_map": [ + 0.03036375902593136, + 0.010795853100717068, + 0.01734931953251362, + 0.01896483637392521, + 0.01705257035791874, + 0.021203720942139626, + 0.013076727278530598, + 0.009850581176578999, + 0.012364641763269901, + 0.011919264681637287, + 0.010246777907013893, + 0.015215935185551643, + 0.02147381566464901, + 0.014626042917370796, + 0.009193557314574718, + 0.010323937050998211, + 0.025328680872917175, + 0.01017714012414217, + 0.016522325575351715, + 0.01610027439892292, + 0.017511656507849693, + 0.016126016154885292, + 0.012833823449909687, + 0.014113345183432102, + 0.009971070103347301, + 0.004897003527730703, + 0.005848615430295467, + 0.0047831423580646515, + 0.01260423008352518, + 0.006083478219807148, + 0.007938460446894169, + 0.008245890028774738, + 0.036677900701761246, + 0.011903660371899605, + 0.012804715894162655, + 0.013146507553756237, + 0.006938030943274498, + 0.010532606393098831, + 0.01147715374827385, + 0.0342169925570488, + 0.011774042621254921, + 0.00731330830603838, + 0.009713947772979736, + 0.011217824183404446, + 0.0053745838813483715, + 0.011441251263022423, + 0.011495773680508137, + 0.006290814373642206, + 0.02515633963048458, + 0.010438863188028336, + 0.0148105314001441, + 0.030296456068754196, + 0.006941807921975851, + 0.01388749573379755, + 0.031210273504257202, + 0.018980607390403748, + 0.009932861663401127, + 0.0036075557582080364, + 0.008532242849469185, + 0.003971450496464968, + 0.0030230358242988586, + 0.012500708922743797, + 0.004811303224414587, + 0.004580801352858543, + 0.006926525849848986, + 0.003182235872372985, + 0.005841401405632496, + 0.004323175642639399, + 0.0028771343640983105, + 0.010104794055223465, + 0.004818813409656286, + 0.003247635904699564, + 0.02578863315284252, + 0.0059048631228506565, + 0.012585366144776344, + 0.013970850966870785, + 0.006203876342624426, + 0.023648610338568687, + 0.016088148579001427, + 0.006379903294146061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0031922117341309786 + } + ], + "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/0000152/group_8/index_tensor([3208, 3208])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_tensor([3208, 3208])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1a402f0af1567c788e3d3d67c78841809b4dd5ca --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_tensor([3208, 3208])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a level angle, the camera extremely slowly zooms in, focusing on the hand of an elderly man reaching toward an ancient tree trunk. The hand is covered in wrinkles and light brown age spots, its skin thinned by age. The golden light of the afternoon sun reveals the distinct outlines of the knuckles. The elderly man, wearing a dark gray wool cardigan with frayed cuffs, cautiously and hesitantly reaches out with an uncontrollable, slight tremor, aiming for a deep, time-polished indentation of an old name carved into the tree's bark. Soft light filters through the branches, casting mottled shadows on the rough, furrowed bark. Dust particles seem to dance in the light. Finally, with utterly gentleness and reverence, the trembling fingertips touch the cool, rough surface of the wood, as if awakening a long-forgotten memory through the touch. The entire scene is immersed in a tranquil and profound atmosphere of nostalgia, with warm tones, a slow rhythm, and a sense of silent storytelling.", + "seed": 19552, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_3208_seed_19552_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": -0.04958253940567374, + "vlm_logprob": -0.04958253940567374, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04958253940567374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.016537025570869446 + ], + "opposite_logprob_tokens": [ + -5.016537189483643 + ], + "logprob_reward": -0.016537025570869446, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016537025570869446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4716154932975769, + 0.05420984327793121, + 0.18858970701694489, + 0.046379122883081436, + 0.07167767733335495, + 0.03168848901987076, + 0.0658765509724617, + 0.02075774408876896, + 0.02796822227537632, + 0.02123711071908474 + ], + "credit_map": [ + 0.03347772732377052, + 0.023166202008724213, + 0.036239996552467346, + 0.07188842445611954, + 0.02775351144373417, + 0.16076712310314178, + 0.06507226079702377, + 0.05325024202466011, + 0.003943485673516989, + 0.006504247896373272, + 0.007414704654365778, + 0.008166966959834099, + 0.007190411444753408, + 0.00552250724285841, + 0.00934549979865551, + 0.006122018210589886, + 0.011213348247110844, + 0.005211642012000084, + 0.03417706862092018, + 0.007327815517783165, + 0.015288718976080418, + 0.07629681378602982, + 0.012188994325697422, + 0.026885302737355232, + 0.0071994587779045105, + 0.004862240515649319, + 0.006158841773867607, + 0.005132559686899185, + 0.0037832455709576607, + 0.007197607308626175, + 0.006381787825375795, + 0.005663383286446333, + 0.008855817839503288, + 0.006873210892081261, + 0.012506944127380848, + 0.00857462640851736, + 0.007764608599245548, + 0.00818853359669447, + 0.00982479378581047, + 0.009089140221476555, + 0.004219802096486092, + 0.004516332410275936, + 0.004215349443256855, + 0.0037586300168186426, + 0.0038045719265937805, + 0.0041768779046833515, + 0.004289418458938599, + 0.0027075055986642838, + 0.005435755010694265, + 0.008389192633330822, + 0.010912579484283924, + 0.006961440201848745, + 0.009033307433128357, + 0.005500673316419125, + 0.006430177018046379, + 0.013213428668677807, + 0.0027959803119301796, + 0.002411952242255211, + 0.0026061220560222864, + 0.002406965009868145, + 0.0038927861023694277, + 0.0025058044120669365, + 0.001811868860386312, + 0.00232626567594707, + 0.003914108499884605, + 0.0021931149531155825, + 0.004162301309406757, + 0.0027423256542533636, + 0.0039756447076797485, + 0.002680714474990964, + 0.002488507889211178, + 0.00581150408834219, + 0.0025448091328144073, + 0.002692932728677988, + 0.0031719068065285683, + 0.0037648333236575127, + 0.002565728034824133, + 0.002257267013192177, + 0.00209970329888165, + 0.0021399292163550854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.016537025570869446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.00527632562443614 + ], + "opposite_logprob_tokens": [ + -5.255276203155518 + ], + "logprob_reward": -0.00527632562443614, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527632562443614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22963590919971466, + 0.05042298510670662, + 0.18838168680667877, + 0.05448830500245094, + 0.2560384273529053, + 0.03135332837700844, + 0.1321982443332672, + 0.013175358064472675, + 0.027007197961211205, + 0.017298582941293716 + ], + "credit_map": [ + 0.018390631303191185, + 0.012213886715471745, + 0.016170550137758255, + 0.03626984730362892, + 0.013382205739617348, + 0.08669761568307877, + 0.020253244787454605, + 0.026257943361997604, + 0.0033371278550475836, + 0.0036826198920607567, + 0.007435700856149197, + 0.008143494836986065, + 0.0077455854043364525, + 0.007114751730114222, + 0.006436015013605356, + 0.006527688819915056, + 0.010758064687252045, + 0.0039595323614776134, + 0.023401346057653427, + 0.005602959543466568, + 0.011955510824918747, + 0.078459732234478, + 0.013232390396296978, + 0.041012149304151535, + 0.006600554566830397, + 0.004780057817697525, + 0.009052111767232418, + 0.00503090163692832, + 0.0039320336654782295, + 0.006652015261352062, + 0.011039219796657562, + 0.007401409558951855, + 0.02328963205218315, + 0.022907575592398643, + 0.04370500519871712, + 0.02709989994764328, + 0.03640570864081383, + 0.04199232906103134, + 0.027205342426896095, + 0.03343294933438301, + 0.004205824341624975, + 0.005617497488856316, + 0.003987243864685297, + 0.004563510417938232, + 0.0029543049167841673, + 0.0027451245114207268, + 0.005025098565965891, + 0.002254723571240902, + 0.01086578331887722, + 0.01262670662254095, + 0.024429552257061005, + 0.010296274907886982, + 0.022925835102796555, + 0.01355686504393816, + 0.01666255295276642, + 0.02083466574549675, + 0.0015578422462567687, + 0.0012080107117071748, + 0.001701954985037446, + 0.0014242465840652585, + 0.0019035135628655553, + 0.001551631954498589, + 0.0017379228956997395, + 0.0020902359392493963, + 0.003418277483433485, + 0.0019818353466689587, + 0.005926711484789848, + 0.001973249251022935, + 0.0019069095142185688, + 0.0023593564983457327, + 0.0028611738234758377, + 0.006579684093594551, + 0.0019691460765898228, + 0.0020217106211930513, + 0.002107079140841961, + 0.0022632295731455088, + 0.0028915044385939837, + 0.0016706575406715274, + 0.002170783234760165, + 0.0022044708020985126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00527632562443614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.113017737865448 + ], + "opposite_logprob_tokens": [ + -2.2380177974700928 + ], + "logprob_reward": -0.113017737865448, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.113017737865448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10148032009601593, + 0.03926517814397812, + 0.41316306591033936, + 0.18176883459091187, + 0.044076718389987946, + 0.053945284336805344, + 0.08674654364585876, + 0.02742406167089939, + 0.03300044685602188, + 0.019129550084471703 + ], + "credit_map": [ + 0.013105674646794796, + 0.008319426327943802, + 0.011183260008692741, + 0.010396067053079605, + 0.006888582371175289, + 0.03020518831908703, + 0.01014777086675167, + 0.011234349571168423, + 0.003211399307474494, + 0.0036193598061800003, + 0.008018388412892818, + 0.003426878247410059, + 0.007101209834218025, + 0.004720552358776331, + 0.00468035601079464, + 0.004487035796046257, + 0.0263349711894989, + 0.012379063293337822, + 0.04161401465535164, + 0.01599186286330223, + 0.023735154420137405, + 0.15821419656276703, + 0.032605722546577454, + 0.10228807479143143, + 0.02602318860590458, + 0.017380820587277412, + 0.023945076391100883, + 0.016074543818831444, + 0.022536205127835274, + 0.03167295083403587, + 0.025229277089238167, + 0.018906766548752785, + 0.008466089144349098, + 0.004265285562723875, + 0.007497934624552727, + 0.004076527431607246, + 0.0036967520136386156, + 0.0042231883853673935, + 0.005870835855603218, + 0.005980107467621565, + 0.00969996489584446, + 0.006682739593088627, + 0.007291024550795555, + 0.008214312605559826, + 0.005519413854926825, + 0.005723451729863882, + 0.006700483150780201, + 0.004113896284252405, + 0.007950656116008759, + 0.010983143001794815, + 0.014961140230298042, + 0.009625129401683807, + 0.011779510416090488, + 0.006382139399647713, + 0.010018919594585896, + 0.015045903623104095, + 0.0025504527147859335, + 0.0021115653216838837, + 0.0040143439546227455, + 0.0021449115592986345, + 0.004204390104860067, + 0.002906302222982049, + 0.0026173496153205633, + 0.006874744314700365, + 0.003461421001702547, + 0.0023861990775913, + 0.006493614055216312, + 0.0025698284152895212, + 0.003935422282665968, + 0.003053920576348901, + 0.0022612041793763638, + 0.008838839828968048, + 0.002529176650568843, + 0.0022931064013391733, + 0.0025365944020450115, + 0.003077020402997732, + 0.0021096004638820887, + 0.002324020955711603, + 0.002295656595379114, + 0.0019643749110400677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.113017737865448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.07002837210893631 + ], + "opposite_logprob_tokens": [ + -2.695028305053711 + ], + "logprob_reward": -0.07002837210893631, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002837210893631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22637318074703217, + 0.10059434920549393, + 0.39021921157836914, + 0.07631347328424454, + 0.053212352097034454, + 0.032428208738565445, + 0.04678243398666382, + 0.027765745297074318, + 0.01769823022186756, + 0.028612874448299408 + ], + "credit_map": [ + 0.02557150460779667, + 0.014062488451600075, + 0.02029518038034439, + 0.02265283651649952, + 0.0156808253377676, + 0.08799472451210022, + 0.014786195941269398, + 0.02532939240336418, + 0.00666689733043313, + 0.008477205410599709, + 0.02259061485528946, + 0.017928285524249077, + 0.009477507323026657, + 0.009684388525784016, + 0.013639617711305618, + 0.012129820883274078, + 0.02424672059714794, + 0.008694772608578205, + 0.05362575128674507, + 0.010840469971299171, + 0.026875736191868782, + 0.16616766154766083, + 0.028704948723316193, + 0.07106310874223709, + 0.009991667233407497, + 0.006889174226671457, + 0.01320606004446745, + 0.0080623934045434, + 0.006264463532716036, + 0.011060144752264023, + 0.01141132041811943, + 0.009428242221474648, + 0.0066008418798446655, + 0.006685641594231129, + 0.007747534196823835, + 0.007748754229396582, + 0.005847922991961241, + 0.005878359079360962, + 0.007631836924701929, + 0.005071456078439951, + 0.0026534036733210087, + 0.002439186442643404, + 0.004411483183503151, + 0.004189209546893835, + 0.004782333970069885, + 0.0042446572333574295, + 0.006068001501262188, + 0.0036399317905306816, + 0.004957729950547218, + 0.003516110824421048, + 0.006583565846085548, + 0.004679820965975523, + 0.006568976677954197, + 0.005966463591903448, + 0.008905051276087761, + 0.005604707170277834, + 0.002894916571676731, + 0.0033007196616381407, + 0.004121730569750071, + 0.00198832037858665, + 0.0036857991944998503, + 0.0019403372425585985, + 0.0021979650482535362, + 0.007635952904820442, + 0.0020441620144993067, + 0.0017343696672469378, + 0.0023758551105856895, + 0.0023310387041419744, + 0.0038108406588435173, + 0.00180762296076864, + 0.0017186219338327646, + 0.0018757183570414782, + 0.0029402729123830795, + 0.005016909912228584, + 0.002508444944396615, + 0.0047830138355493546, + 0.00428383145481348, + 0.0029847121331840754, + 0.0030080066062510014, + 0.003087681718170643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07002837210893631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.04305323585867882 + ], + "opposite_logprob_tokens": [ + -3.168053150177002 + ], + "logprob_reward": -0.04305323585867882, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305323585867882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.223391592502594, + 0.031702615320682526, + 0.5181524753570557, + 0.0447695292532444, + 0.039678700268268585, + 0.04639340937137604, + 0.05485193431377411, + 0.014197994954884052, + 0.01808425970375538, + 0.008777454495429993 + ], + "credit_map": [ + 0.01604677177965641, + 0.012029744684696198, + 0.0187580194324255, + 0.032542746514081955, + 0.014278476126492023, + 0.06806886941194534, + 0.03648844361305237, + 0.025178533047437668, + 0.002451839856803417, + 0.003908238373696804, + 0.004196316469460726, + 0.004548629280179739, + 0.005034790839999914, + 0.002795365871861577, + 0.004558178596198559, + 0.00420925859361887, + 0.030672699213027954, + 0.012059083208441734, + 0.09155610203742981, + 0.016671549528837204, + 0.03964066505432129, + 0.215496227145195, + 0.03553255647420883, + 0.07652357220649719, + 0.004972133785486221, + 0.003658704459667206, + 0.010936634615063667, + 0.00448852963745594, + 0.003784269094467163, + 0.004935960751026869, + 0.006837924942374229, + 0.005155373830348253, + 0.00402454100549221, + 0.004481010604649782, + 0.005243450403213501, + 0.003982855007052422, + 0.00633491575717926, + 0.005007231142371893, + 0.006285692565143108, + 0.0043190051801502705, + 0.006022290326654911, + 0.008559483103454113, + 0.006729242857545614, + 0.006022336892783642, + 0.0052942619659006596, + 0.003652812447398901, + 0.006796394474804401, + 0.003316588466987014, + 0.0060671851970255375, + 0.003660439280793071, + 0.008877268061041832, + 0.0063779097981750965, + 0.006920425221323967, + 0.009705794975161552, + 0.008657616563141346, + 0.004585301503539085, + 0.0012105508940294385, + 0.0011348766274750233, + 0.0021130878012627363, + 0.0013087181141600013, + 0.002074335003271699, + 0.0015926429769024253, + 0.0015103939222171903, + 0.0032533900812268257, + 0.002117699244990945, + 0.0019163568504154682, + 0.002032917458564043, + 0.002103125676512718, + 0.0038133191410452127, + 0.0020942625124007463, + 0.0016164855333045125, + 0.002390093170106411, + 0.001619870075955987, + 0.001207544468343258, + 0.000859844614751637, + 0.0015848102048039436, + 0.0006132286507636309, + 0.0011043432168662548, + 0.0009782157139852643, + 0.0008095977827906609 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.04305323585867882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera extremely slowly zooms in, focusing on the hand of an elderly man reaching toward an ancient tree trunk. The hand is covered in wrinkles and light brown age spots, its skin thinned by age. The golden light of the afternoon sun reveals the distinct outlines of the knuckles. The elderly man, wearing a dark gray wool cardigan with frayed cuffs, cautiously and hesitantly reaches out with an uncontrollable, slight tremor, aiming for a deep, time-polished indentation of an old name carved into the tree's bark. Soft light filters through the branches, casting mottled shadows on the rough, furrowed bark. Dust particles seem to dance in the light. Finally, with utterly gentleness and reverence, the trembling fingertips touch the cool, rough surface of the wood, as if awakening a long-forgotten memory through the touch. The entire scene is immersed in a tranquil and profound atmosphere of nostalgia, with warm tones, a slow rhythm, and a sense of silent storytelling.", + "seed": 19553, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_3208_seed_19553_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": -0.04716495554894209, + "vlm_logprob": -0.04716495554894209, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.04716495554894209, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.01540323905646801 + ], + "opposite_logprob_tokens": [ + -4.7654032707214355 + ], + "logprob_reward": -0.01540323905646801, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01540323905646801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13476425409317017, + 0.05196608230471611, + 0.08273081481456757, + 0.06028909608721733, + 0.14714375138282776, + 0.27717310190200806, + 0.05059808865189552, + 0.08744632452726364, + 0.06793948262929916, + 0.03994905203580856 + ], + "credit_map": [ + 0.014789205975830555, + 0.010905781760811806, + 0.015279429033398628, + 0.024338221177458763, + 0.015238876454532146, + 0.016787854954600334, + 0.02225114218890667, + 0.015173734165728092, + 0.0070914230309426785, + 0.0042168437503278255, + 0.010664231143891811, + 0.004977529868483543, + 0.005241174250841141, + 0.009200207889080048, + 0.004289931617677212, + 0.006284737028181553, + 0.00794915109872818, + 0.008172379806637764, + 0.009724431671202183, + 0.010441197082400322, + 0.011857583187520504, + 0.014740746468305588, + 0.009477539919316769, + 0.01036777812987566, + 0.007005364168435335, + 0.008730422705411911, + 0.008519518189132214, + 0.005769798997789621, + 0.007692614570260048, + 0.008120170794427395, + 0.006134529132395983, + 0.008316676132380962, + 0.012497245334088802, + 0.01155311893671751, + 0.019652562215924263, + 0.008912739343941212, + 0.01884629763662815, + 0.025789527222514153, + 0.025629937648773193, + 0.024262316524982452, + 0.0428890585899353, + 0.027227461338043213, + 0.03989315778017044, + 0.04054204747080803, + 0.026581360027194023, + 0.03593635931611061, + 0.029708780348300934, + 0.03439486026763916, + 0.007362949196249247, + 0.005984937306493521, + 0.00739979138597846, + 0.006913571152836084, + 0.00431632436811924, + 0.008573153987526894, + 0.004146287217736244, + 0.005901070311665535, + 0.005416208412498236, + 0.006929811090230942, + 0.007916253991425037, + 0.008868256583809853, + 0.008115490898489952, + 0.020446889102458954, + 0.006495140492916107, + 0.0232582688331604, + 0.009967481717467308, + 0.007201441563665867, + 0.006388205569237471, + 0.00934302806854248, + 0.012924423441290855, + 0.008563516661524773, + 0.006465977523475885, + 0.007085401564836502, + 0.0029338588938117027, + 0.0027833583299070597, + 0.004181773867458105, + 0.004322147462517023, + 0.005626076366752386, + 0.0069354381412267685, + 0.004968062974512577, + 0.008198335766792297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.01540323905646801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.014234201051294804 + ], + "opposite_logprob_tokens": [ + -4.2642340660095215 + ], + "logprob_reward": -0.014234201051294804, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014234201051294804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12729841470718384, + 0.07867832481861115, + 0.0838049128651619, + 0.06262196600437164, + 0.10542715340852737, + 0.07510778307914734, + 0.03095097653567791, + 0.14006046950817108, + 0.17975620925426483, + 0.1162937805056572 + ], + "credit_map": [ + 0.016927162185311317, + 0.008766476064920425, + 0.01318180002272129, + 0.02839868888258934, + 0.012095865793526173, + 0.016052400693297386, + 0.018063565716147423, + 0.01381244882941246, + 0.007538077887147665, + 0.0072332420386374, + 0.013635551556944847, + 0.008333883248269558, + 0.010270249098539352, + 0.008312314748764038, + 0.007475226651877165, + 0.015879778191447258, + 0.008481844328343868, + 0.010389800183475018, + 0.01024620421230793, + 0.010526413097977638, + 0.012652028352022171, + 0.012799162417650223, + 0.009487918578088284, + 0.009221547283232212, + 0.007551497779786587, + 0.008894375525414944, + 0.010379057377576828, + 0.006609264295548201, + 0.008615903556346893, + 0.007930809631943703, + 0.004602022469043732, + 0.008039035834372044, + 0.0048367297276854515, + 0.006227487698197365, + 0.017266174778342247, + 0.00399510795250535, + 0.005807104054838419, + 0.026078905910253525, + 0.00653312262147665, + 0.03468251973390579, + 0.008499915711581707, + 0.005852381698787212, + 0.010965412482619286, + 0.00897175446152687, + 0.008392282761633396, + 0.011479012668132782, + 0.011392387561500072, + 0.009554634802043438, + 0.003889745566993952, + 0.0034580097999423742, + 0.00504092825576663, + 0.003432692028582096, + 0.0029268222860991955, + 0.005546976812183857, + 0.002438374562188983, + 0.0042174276895821095, + 0.008343636058270931, + 0.00809499155730009, + 0.01458047330379486, + 0.007212043274194002, + 0.00765964575111866, + 0.05230387672781944, + 0.0060599977150559425, + 0.035805802792310715, + 0.02369958907365799, + 0.017641650512814522, + 0.014457297511398792, + 0.02073042467236519, + 0.0416744202375412, + 0.025119412690401077, + 0.016574425622820854, + 0.019858989864587784, + 0.015969956293702126, + 0.008005140349268913, + 0.008675796911120415, + 0.017418386414647102, + 0.02099067158997059, + 0.017137955874204636, + 0.014976684004068375, + 0.013119191862642765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.014234201051294804 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.08918023854494095 + ], + "opposite_logprob_tokens": [ + -2.4641802310943604 + ], + "logprob_reward": -0.08918023854494095, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08918023854494095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06789407879114151, + 0.021429095417261124, + 0.0230411384254694, + 0.02549874782562256, + 0.13499625027179718, + 0.06748340278863907, + 0.017156880348920822, + 0.5785205960273743, + 0.043600473552942276, + 0.020379379391670227 + ], + "credit_map": [ + 0.007265402469784021, + 0.005986139643937349, + 0.01720973290503025, + 0.011480904184281826, + 0.004674962256103754, + 0.006150784902274609, + 0.010624335147440434, + 0.004501814488321543, + 0.002510739490389824, + 0.001392412232235074, + 0.005860908888280392, + 0.00250927172601223, + 0.0029050034936517477, + 0.002206446835771203, + 0.0016608198639005423, + 0.0023834898602217436, + 0.0024221378844231367, + 0.004306776914745569, + 0.002745932200923562, + 0.0031559979543089867, + 0.002887438051402569, + 0.0026252875104546547, + 0.0021845733281224966, + 0.002712992485612631, + 0.0028014995623379946, + 0.0043013920076191425, + 0.0033841172698885202, + 0.003335096873342991, + 0.003359563648700714, + 0.0025999003555625677, + 0.0032254524994641542, + 0.0024917249102145433, + 0.009510748088359833, + 0.00863607693463564, + 0.022336022928357124, + 0.006323518231511116, + 0.011987486854195595, + 0.02814875915646553, + 0.008430678397417068, + 0.0396229587495327, + 0.009451241232454777, + 0.007422246970236301, + 0.00939112901687622, + 0.00758266681805253, + 0.007537041790783405, + 0.009015854448080063, + 0.007142882794141769, + 0.009940332733094692, + 0.0015962565084919333, + 0.002262626774609089, + 0.0022690065670758486, + 0.0021077152341604233, + 0.0022365092299878597, + 0.0020902385003864765, + 0.0018862319411709905, + 0.0027082934975624084, + 0.03340617194771767, + 0.015605957247316837, + 0.055322155356407166, + 0.016884636133909225, + 0.022758351638913155, + 0.24291472136974335, + 0.016711285337805748, + 0.1749172806739807, + 0.00721038319170475, + 0.00488758971914649, + 0.004074644297361374, + 0.005855673924088478, + 0.0066453092731535435, + 0.0048761144280433655, + 0.0050449916161596775, + 0.005005763377994299, + 0.0027134926058351994, + 0.002187923761084676, + 0.0021694451570510864, + 0.0029251498635858297, + 0.0023732916451990604, + 0.00296982959844172, + 0.002435076981782913, + 0.0026051700115203857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.08918023854494095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.05497528240084648 + ], + "opposite_logprob_tokens": [ + -2.9299752712249756 + ], + "logprob_reward": -0.05497528240084648, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497528240084648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1098538413643837, + 0.06722161173820496, + 0.10524635016918182, + 0.10252011567354202, + 0.09080302715301514, + 0.09463271498680115, + 0.050954677164554596, + 0.12595216929912567, + 0.11759250611066818, + 0.13522300124168396 + ], + "credit_map": [ + 0.01266565639525652, + 0.011357847601175308, + 0.010337579064071178, + 0.019222302362322807, + 0.00897615309804678, + 0.0198639128357172, + 0.011213068850338459, + 0.016217319294810295, + 0.008366761729121208, + 0.005892204586416483, + 0.008226221427321434, + 0.006597746163606644, + 0.00907187070697546, + 0.010359120555222034, + 0.005165978334844112, + 0.013541708700358868, + 0.008743281476199627, + 0.011414003558456898, + 0.011701186187565327, + 0.00984127540141344, + 0.016228707507252693, + 0.019482171162962914, + 0.011092280969023705, + 0.01674344576895237, + 0.010997187346220016, + 0.012783421203494072, + 0.015721013769507408, + 0.013848485425114632, + 0.013640674762427807, + 0.009790686890482903, + 0.01430787518620491, + 0.011430774815380573, + 0.008008609525859356, + 0.009542780928313732, + 0.01293773390352726, + 0.005260717123746872, + 0.009646265767514706, + 0.021730337291955948, + 0.007365979719907045, + 0.01631060056388378, + 0.014317980036139488, + 0.009927171282470226, + 0.013668014667928219, + 0.015615595504641533, + 0.008470611646771431, + 0.01161138154566288, + 0.010147681459784508, + 0.010874280706048012, + 0.004457037430256605, + 0.0059655033983290195, + 0.007546905893832445, + 0.005958868190646172, + 0.008069846779108047, + 0.009746766649186611, + 0.00419627595692873, + 0.005013473331928253, + 0.009975844994187355, + 0.012893373146653175, + 0.02317935787141323, + 0.013251438736915588, + 0.013892277143895626, + 0.029876098036766052, + 0.011305976659059525, + 0.011577797122299671, + 0.014577939175069332, + 0.015240458771586418, + 0.010826689191162586, + 0.012795360758900642, + 0.022330623120069504, + 0.016748730093240738, + 0.008567741140723228, + 0.016504961997270584, + 0.00972293596714735, + 0.009345184080302715, + 0.012581768445670605, + 0.018624693155288696, + 0.022626062855124474, + 0.02420436404645443, + 0.012145756743848324, + 0.02597222477197647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.05497528240084648 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.0620318166911602 + ], + "opposite_logprob_tokens": [ + -2.8120317459106445 + ], + "logprob_reward": -0.0620318166911602, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620318166911602, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08755464851856232, + 0.04632924497127533, + 0.131647989153862, + 0.11333249509334564, + 0.05275889113545418, + 0.12609076499938965, + 0.17283384501934052, + 0.08808979392051697, + 0.13750876486301422, + 0.04385362192988396 + ], + "credit_map": [ + 0.012997069396078587, + 0.00852400716394186, + 0.00646323524415493, + 0.01124181505292654, + 0.010904024355113506, + 0.012416435405611992, + 0.011736025102436543, + 0.013272037729620934, + 0.005664035677909851, + 0.004860938061028719, + 0.01181532721966505, + 0.005546464119106531, + 0.0041106888093054295, + 0.004791529383510351, + 0.0036978530697524548, + 0.005842407234013081, + 0.0156718660145998, + 0.0309408288449049, + 0.019561558961868286, + 0.014965845271945, + 0.015918854624032974, + 0.01248975656926632, + 0.010789250023663044, + 0.011310027912259102, + 0.014396335929632187, + 0.017242655158042908, + 0.019598282873630524, + 0.012911215424537659, + 0.012380542233586311, + 0.012204024009406567, + 0.011330709792673588, + 0.013268724083900452, + 0.00662205507978797, + 0.006602698937058449, + 0.009254437871277332, + 0.0036878548562526703, + 0.006907460745424032, + 0.009302548132836819, + 0.003243440994992852, + 0.007138397544622421, + 0.014583791606128216, + 0.010780167765915394, + 0.018660282716155052, + 0.013214681297540665, + 0.015837928280234337, + 0.019332509487867355, + 0.017026200890541077, + 0.016655202955007553, + 0.02221124991774559, + 0.019707350060343742, + 0.024485208094120026, + 0.020962094888091087, + 0.02117757685482502, + 0.027964046224951744, + 0.015081670135259628, + 0.02124464139342308, + 0.007429469842463732, + 0.009336614049971104, + 0.016480904072523117, + 0.009375824593007565, + 0.006664780434221029, + 0.024655358865857124, + 0.006063658744096756, + 0.008083181455731392, + 0.01898210309445858, + 0.016689008101820946, + 0.015470236539840698, + 0.022561630234122276, + 0.02268216386437416, + 0.01636810600757599, + 0.00998029112815857, + 0.014775215648114681, + 0.008146068081259727, + 0.0051255002617836, + 0.004599504638463259, + 0.005031078588217497, + 0.005272251553833485, + 0.006090698763728142, + 0.003348792437463999, + 0.006239728070795536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.0620318166911602 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera extremely slowly zooms in, focusing on the hand of an elderly man reaching toward an ancient tree trunk. The hand is covered in wrinkles and light brown age spots, its skin thinned by age. The golden light of the afternoon sun reveals the distinct outlines of the knuckles. The elderly man, wearing a dark gray wool cardigan with frayed cuffs, cautiously and hesitantly reaches out with an uncontrollable, slight tremor, aiming for a deep, time-polished indentation of an old name carved into the tree's bark. Soft light filters through the branches, casting mottled shadows on the rough, furrowed bark. Dust particles seem to dance in the light. Finally, with utterly gentleness and reverence, the trembling fingertips touch the cool, rough surface of the wood, as if awakening a long-forgotten memory through the touch. The entire scene is immersed in a tranquil and profound atmosphere of nostalgia, with warm tones, a slow rhythm, and a sense of silent storytelling.", + "seed": 19554, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_3208_seed_19554_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": -0.04239086229354143, + "vlm_logprob": -0.04239086229354143, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04239086229354143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.01122740563005209 + ], + "opposite_logprob_tokens": [ + -5.261227607727051 + ], + "logprob_reward": -0.01122740563005209, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01122740563005209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2526794672012329, + 0.11395208537578583, + 0.0953962653875351, + 0.09598728269338608, + 0.09405378252267838, + 0.11584266275167465, + 0.0504586398601532, + 0.04518174007534981, + 0.057294536381959915, + 0.07915352284908295 + ], + "credit_map": [ + 0.044385239481925964, + 0.0761437937617302, + 0.023601844906806946, + 0.02584783546626568, + 0.01656777784228325, + 0.015646744519472122, + 0.02446088008582592, + 0.026025351136922836, + 0.018075210973620415, + 0.012324973940849304, + 0.01930980756878853, + 0.011370407417416573, + 0.008346485905349255, + 0.013628649525344372, + 0.014551019296050072, + 0.016345536336302757, + 0.013735728338360786, + 0.010627217590808868, + 0.014971550554037094, + 0.015112739987671375, + 0.007582451682537794, + 0.010888129472732544, + 0.013853642158210278, + 0.008624816313385963, + 0.010382290929555893, + 0.006345789879560471, + 0.010856136679649353, + 0.006754689384251833, + 0.011657819151878357, + 0.01685405522584915, + 0.013705951161682606, + 0.019430555403232574, + 0.011050322093069553, + 0.008297170512378216, + 0.016142139211297035, + 0.014021887443959713, + 0.01261874008923769, + 0.009405476041138172, + 0.008661620318889618, + 0.013856430537998676, + 0.010993185453116894, + 0.009987096302211285, + 0.016244273632764816, + 0.012459937483072281, + 0.010968321934342384, + 0.0216695386916399, + 0.013099933043122292, + 0.020420383661985397, + 0.003189780516549945, + 0.002848232164978981, + 0.00562091963365674, + 0.006684388034045696, + 0.009279384277760983, + 0.0060725342482328415, + 0.005672783590853214, + 0.01109061948955059, + 0.0031407750211656094, + 0.0036000912077724934, + 0.003668778808787465, + 0.006031065713614225, + 0.00539690675213933, + 0.007511146366596222, + 0.01012025959789753, + 0.005712716840207577, + 0.0075967214070260525, + 0.00658831000328064, + 0.005261705256998539, + 0.008594117127358913, + 0.006305253598839045, + 0.008504170924425125, + 0.006935636047273874, + 0.007508625276386738, + 0.008013026788830757, + 0.005639384966343641, + 0.020449379459023476, + 0.005440437234938145, + 0.005443375091999769, + 0.009435721673071384, + 0.005201324354857206, + 0.019530871883034706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01122740563005209 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.004657847806811333 + ], + "opposite_logprob_tokens": [ + -5.379657745361328 + ], + "logprob_reward": -0.004657847806811333, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004657847806811333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21393458545207977, + 0.08545145392417908, + 0.06502523273229599, + 0.06089957058429718, + 0.07326852530241013, + 0.1912813037633896, + 0.07516099512577057, + 0.052352506667375565, + 0.06638804078102112, + 0.1162378191947937 + ], + "credit_map": [ + 0.039606474339962006, + 0.04477985203266144, + 0.02089795470237732, + 0.02482648752629757, + 0.01624741405248642, + 0.013649111613631248, + 0.02725200168788433, + 0.026675283908843994, + 0.013224968686699867, + 0.010087882168591022, + 0.00900364387780428, + 0.009059983305633068, + 0.0063071949407458305, + 0.014075515791773796, + 0.012560835108160973, + 0.011131425388157368, + 0.008596663363277912, + 0.007258924189954996, + 0.010940269567072392, + 0.01179464440792799, + 0.004744936712086201, + 0.005910773761570454, + 0.007013294845819473, + 0.008765730075538158, + 0.0067682648077607155, + 0.006509710103273392, + 0.007346577476710081, + 0.006703829392790794, + 0.008987260051071644, + 0.010441216640174389, + 0.007180138025432825, + 0.006962575949728489, + 0.006881394889205694, + 0.010446034371852875, + 0.008384941145777702, + 0.009417214430868626, + 0.0136643685400486, + 0.008214576169848442, + 0.006363268010318279, + 0.009896723553538322, + 0.017665404826402664, + 0.014745516702532768, + 0.033838484436273575, + 0.02376028522849083, + 0.01851864904165268, + 0.03249228000640869, + 0.01924838125705719, + 0.031012311577796936, + 0.004475381225347519, + 0.004566351417452097, + 0.0057625374756753445, + 0.007731251884251833, + 0.01714230142533779, + 0.008813128806650639, + 0.006483318284153938, + 0.02018672227859497, + 0.004759376868605614, + 0.004168593790382147, + 0.006823547184467316, + 0.00766364112496376, + 0.008268875069916248, + 0.007204702589660883, + 0.007031143177300692, + 0.00643262080848217, + 0.008804039098322392, + 0.006304276175796986, + 0.007021918892860413, + 0.0107920216396451, + 0.007826115936040878, + 0.009895410388708115, + 0.008218959905207157, + 0.00752530200406909, + 0.014346683397889137, + 0.011135471984744072, + 0.009993805550038815, + 0.016193168237805367, + 0.011915368027985096, + 0.012273840606212616, + 0.014401202090084553, + 0.025978289544582367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.004657847806811333 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.07905983179807663 + ], + "opposite_logprob_tokens": [ + -2.5790598392486572 + ], + "logprob_reward": -0.07905983179807663, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905983179807663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28845804929733276, + 0.0797484740614891, + 0.11336491256952286, + 0.05283123627305031, + 0.14868107438087463, + 0.12486894428730011, + 0.03987916558980942, + 0.04977543652057648, + 0.053446169942617416, + 0.048946525901556015 + ], + "credit_map": [ + 0.04065472632646561, + 0.04302468150854111, + 0.03003082051873207, + 0.04540327191352844, + 0.03143691271543503, + 0.022726545110344887, + 0.04031510278582573, + 0.03486596792936325, + 0.013027393259108067, + 0.01067151315510273, + 0.011752655729651451, + 0.009494449011981487, + 0.006453066598623991, + 0.006932830438017845, + 0.009560877457261086, + 0.011855682358145714, + 0.016058657318353653, + 0.012101972475647926, + 0.019024299457669258, + 0.02208593674004078, + 0.008078133687376976, + 0.012143741361796856, + 0.00991076324135065, + 0.013961407355964184, + 0.0050101992674171925, + 0.00361253647133708, + 0.005740439984947443, + 0.005576556548476219, + 0.006699028890579939, + 0.011243878863751888, + 0.0058115022256970406, + 0.009137092158198357, + 0.012648701667785645, + 0.013322128914296627, + 0.024074792861938477, + 0.01620510034263134, + 0.02546532265841961, + 0.016856633126735687, + 0.011899924837052822, + 0.02820846065878868, + 0.012492037378251553, + 0.009113348089158535, + 0.02208918333053589, + 0.016715534031391144, + 0.013369347900152206, + 0.016568614169955254, + 0.012731371447443962, + 0.02178950048983097, + 0.004042789805680513, + 0.0033752783201634884, + 0.009152887389063835, + 0.004363061860203743, + 0.004759742878377438, + 0.004883307032287121, + 0.004074291326105595, + 0.005227805580943823, + 0.005063855554908514, + 0.005524584092199802, + 0.00928551610559225, + 0.005381900817155838, + 0.006533697247505188, + 0.007102089002728462, + 0.005923138931393623, + 0.0049606566317379475, + 0.006625850219279528, + 0.006841612048447132, + 0.005634102039039135, + 0.007431074045598507, + 0.005270922556519508, + 0.0077728875912725925, + 0.005597445648163557, + 0.008272276259958744, + 0.005016877315938473, + 0.004392217844724655, + 0.015033387579023838, + 0.005231752526015043, + 0.0023589839693158865, + 0.0038114716298878193, + 0.0033547268249094486, + 0.00974710751324892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.07905983179807663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.06206587329506874 + ], + "opposite_logprob_tokens": [ + -2.812065839767456 + ], + "logprob_reward": -0.06206587329506874, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206587329506874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26051992177963257, + 0.10428142547607422, + 0.10857026278972626, + 0.1116851270198822, + 0.05441250652074814, + 0.09703414142131805, + 0.06187323108315468, + 0.06672415882349014, + 0.08130084723234177, + 0.05359834432601929 + ], + "credit_map": [ + 0.04384731501340866, + 0.0782415121793747, + 0.017881738021969795, + 0.02214718796312809, + 0.019129546359181404, + 0.02029709704220295, + 0.027970870956778526, + 0.03100467473268509, + 0.01495885569602251, + 0.010921384207904339, + 0.012163258157670498, + 0.011581556871533394, + 0.011048185639083385, + 0.014717272482812405, + 0.013827560469508171, + 0.015063364058732986, + 0.016049548983573914, + 0.011359955184161663, + 0.01502256840467453, + 0.012532887049019337, + 0.011020311154425144, + 0.010291771031916142, + 0.019621359184384346, + 0.012671875767409801, + 0.012021791189908981, + 0.008540802635252476, + 0.015813743695616722, + 0.015852222219109535, + 0.014754993841052055, + 0.014068333432078362, + 0.019064228981733322, + 0.011569017544388771, + 0.007345282007008791, + 0.004407977685332298, + 0.009281273931264877, + 0.006094207987189293, + 0.008043993264436722, + 0.0076737673953175545, + 0.004577640909701586, + 0.006988368462771177, + 0.012381923384964466, + 0.006504646968096495, + 0.022092819213867188, + 0.011595222167670727, + 0.013587579131126404, + 0.011644931510090828, + 0.008455829694867134, + 0.010771193541586399, + 0.005234929732978344, + 0.005322055891156197, + 0.006097805220633745, + 0.005186139605939388, + 0.012926928699016571, + 0.008399815298616886, + 0.0066360970959067345, + 0.012069466523826122, + 0.0066147493198513985, + 0.006862854119390249, + 0.01038183830678463, + 0.007792366202920675, + 0.009438198991119862, + 0.009831815958023071, + 0.0087177325040102, + 0.007084606681019068, + 0.011994880624115467, + 0.009783545508980751, + 0.00969731342047453, + 0.011379523202776909, + 0.008059916086494923, + 0.009678875096142292, + 0.009650529362261295, + 0.011056272312998772, + 0.007996945641934872, + 0.008200172334909439, + 0.005913945380598307, + 0.0051680696196854115, + 0.005124830640852451, + 0.008667543530464172, + 0.004856589715927839, + 0.007670252118259668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.06206587329506874 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.054943352937698364 + ], + "opposite_logprob_tokens": [ + -2.929943323135376 + ], + "logprob_reward": -0.054943352937698364, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054943352937698364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12068381160497665, + 0.10085352510213852, + 0.06350505352020264, + 0.1207784116268158, + 0.09705212712287903, + 0.313305139541626, + 0.053118735551834106, + 0.03685135766863823, + 0.07220565527677536, + 0.02164621651172638 + ], + "credit_map": [ + 0.01483670063316822, + 0.013708086684346199, + 0.013583012856543064, + 0.017247885465621948, + 0.01697107031941414, + 0.01197007019072771, + 0.02058202028274536, + 0.011784950271248817, + 0.014556811191141605, + 0.010492303408682346, + 0.013772212900221348, + 0.012847560457885265, + 0.010270626284182072, + 0.012980249710381031, + 0.012279524467885494, + 0.013654232025146484, + 0.010029428638517857, + 0.009126661345362663, + 0.005561204627156258, + 0.006761595606803894, + 0.005151059944182634, + 0.005950327962636948, + 0.01007496751844883, + 0.010849805548787117, + 0.015566217713057995, + 0.008860214613378048, + 0.016661858186125755, + 0.01214523334056139, + 0.009906576946377754, + 0.017828045412898064, + 0.020921772345900536, + 0.018888486549258232, + 0.008912368677556515, + 0.012386715039610863, + 0.015013864263892174, + 0.012398415245115757, + 0.014355270192027092, + 0.013020981103181839, + 0.010170391760766506, + 0.01079411432147026, + 0.027414942160248756, + 0.023997433483600616, + 0.04962319880723953, + 0.044526778161525726, + 0.035370659083127975, + 0.04628997668623924, + 0.030414827167987823, + 0.05566732585430145, + 0.004029776901006699, + 0.003291507950052619, + 0.008424295112490654, + 0.006546150427311659, + 0.007512623909860849, + 0.007942844182252884, + 0.00786361750215292, + 0.0075079211965203285, + 0.003836077870801091, + 0.0026306910440325737, + 0.005658443551510572, + 0.004436619579792023, + 0.0039018348325043917, + 0.007266980595886707, + 0.004977197851985693, + 0.004143511410802603, + 0.006577093619853258, + 0.0054682581685483456, + 0.007772727869451046, + 0.012787058018147945, + 0.008507982827723026, + 0.011852623894810677, + 0.009212159551680088, + 0.010027758777141571, + 0.003007526509463787, + 0.0016145848203450441, + 0.0034846696071326733, + 0.0022523393854498863, + 0.001979393418878317, + 0.0026535214856266975, + 0.0023642643354833126, + 0.004289916716516018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.054943352937698364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a level angle, the camera extremely slowly zooms in, focusing on the hand of an elderly man reaching toward an ancient tree trunk. The hand is covered in wrinkles and light brown age spots, its skin thinned by age. The golden light of the afternoon sun reveals the distinct outlines of the knuckles. The elderly man, wearing a dark gray wool cardigan with frayed cuffs, cautiously and hesitantly reaches out with an uncontrollable, slight tremor, aiming for a deep, time-polished indentation of an old name carved into the tree's bark. Soft light filters through the branches, casting mottled shadows on the rough, furrowed bark. Dust particles seem to dance in the light. Finally, with utterly gentleness and reverence, the trembling fingertips touch the cool, rough surface of the wood, as if awakening a long-forgotten memory through the touch. The entire scene is immersed in a tranquil and profound atmosphere of nostalgia, with warm tones, a slow rhythm, and a sense of silent storytelling.", + "seed": 19555, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_8/index_3208_seed_19555_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": -0.05024749003350735, + "vlm_logprob": -0.05024749003350735, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.05024749003350735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.025858532637357712 + ], + "opposite_logprob_tokens": [ + -4.150858402252197 + ], + "logprob_reward": -0.025858532637357712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.025858532637357712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3196679353713989, + 0.3539635241031647, + 0.08046577125787735, + 0.03197300061583519, + 0.027156751602888107, + 0.03485475853085518, + 0.0248413123190403, + 0.04562797397375107, + 0.055223941802978516, + 0.026225091889500618 + ], + "credit_map": [ + 0.04489491879940033, + 0.028818292543292046, + 0.03145328909158707, + 0.03293641284108162, + 0.023991180583834648, + 0.03332402929663658, + 0.06140616163611412, + 0.06284364312887192, + 0.09577714651823044, + 0.0182994082570076, + 0.019906165078282356, + 0.012387248687446117, + 0.016680721193552017, + 0.027715077623724937, + 0.028466133400797844, + 0.13473162055015564, + 0.010971886105835438, + 0.007536900229752064, + 0.014329599216580391, + 0.013122441247105598, + 0.0094071039929986, + 0.007830914109945297, + 0.009683097712695599, + 0.007583831902593374, + 0.0042287749238312244, + 0.0041020093485713005, + 0.004946387372910976, + 0.004464777186512947, + 0.004193497356027365, + 0.0025839023292064667, + 0.003806873457506299, + 0.003646779339760542, + 0.0033268039114773273, + 0.004015512764453888, + 0.0038553935009986162, + 0.0035000278148800135, + 0.00349613418802619, + 0.0028920331969857216, + 0.0025228848680853844, + 0.003547962987795472, + 0.005998462438583374, + 0.0026929890736937523, + 0.0073740920051932335, + 0.003935455344617367, + 0.0035931637976318598, + 0.0037340158596634865, + 0.0036350274458527565, + 0.003891554195433855, + 0.0030683595687150955, + 0.002489869249984622, + 0.004745489917695522, + 0.002295881975442171, + 0.0023573406506329775, + 0.0030595154967159033, + 0.004326806403696537, + 0.002498048823326826, + 0.006392687559127808, + 0.005368759389966726, + 0.007617495488375425, + 0.0029468676075339317, + 0.002425906015560031, + 0.00880721490830183, + 0.005988640710711479, + 0.006080402061343193, + 0.006686887238174677, + 0.008174432441592216, + 0.008561876602470875, + 0.004406480584293604, + 0.005274598486721516, + 0.009412853978574276, + 0.007012559566646814, + 0.005694253370165825, + 0.004326647613197565, + 0.0021343196276575327, + 0.0031360818538814783, + 0.002921590581536293, + 0.002571478486061096, + 0.0046758148819208145, + 0.0036449937615543604, + 0.002814165549352765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.025858532637357712 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.005279408767819405 + ], + "opposite_logprob_tokens": [ + -5.255279541015625 + ], + "logprob_reward": -0.005279408767819405, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005279408767819405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17841726541519165, + 0.3948152959346771, + 0.1769310086965561, + 0.0353936031460762, + 0.05033964663743973, + 0.037130165845155716, + 0.021025123074650764, + 0.052253358066082, + 0.03203573450446129, + 0.021658705547451973 + ], + "credit_map": [ + 0.02360350638628006, + 0.017196211963891983, + 0.02856321819126606, + 0.022458169609308243, + 0.014387787319719791, + 0.018440265208482742, + 0.027815423905849457, + 0.025952700525522232, + 0.10460231453180313, + 0.021343136206269264, + 0.019752517342567444, + 0.013965505175292492, + 0.019271917641162872, + 0.030927937477827072, + 0.03154479339718819, + 0.15340720117092133, + 0.02017822489142418, + 0.016680806875228882, + 0.02397884614765644, + 0.03449155390262604, + 0.026223115622997284, + 0.015198043547570705, + 0.02387012355029583, + 0.016310304403305054, + 0.005705197341740131, + 0.004865189548581839, + 0.006063494831323624, + 0.004149907734245062, + 0.0030829543247818947, + 0.0032105601858347654, + 0.0035323568154126406, + 0.004783945623785257, + 0.008711536414921284, + 0.004532469902187586, + 0.008573229424655437, + 0.006674166303128004, + 0.007451266050338745, + 0.003574161557480693, + 0.005158242769539356, + 0.005664576310664415, + 0.005486892070621252, + 0.0034224865958094597, + 0.007126959972083569, + 0.002544178394600749, + 0.002335675759240985, + 0.0044408454559743404, + 0.0034147382248193026, + 0.008358395658433437, + 0.0037200718652457, + 0.001993519952520728, + 0.0038395794108510017, + 0.0028616348281502724, + 0.0018657754408195615, + 0.0018870693165808916, + 0.002443076577037573, + 0.002414398593828082, + 0.007489851675927639, + 0.003716773819178343, + 0.006748488172888756, + 0.006529528181999922, + 0.005386429838836193, + 0.007259008474647999, + 0.008452181704342365, + 0.006671102251857519, + 0.004587093368172646, + 0.0043802158907055855, + 0.0036189102102071047, + 0.003174163168296218, + 0.0035225138999521732, + 0.0051373085007071495, + 0.004166491329669952, + 0.003449039300903678, + 0.0037668694276362658, + 0.0020958748646080494, + 0.0020173268858343363, + 0.0028222573455423117, + 0.0018522199243307114, + 0.0035560771357268095, + 0.003322223899886012, + 0.002225857460871339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.005279408767819405 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.07905267179012299 + ], + "opposite_logprob_tokens": [ + -2.579052686691284 + ], + "logprob_reward": -0.07905267179012299, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905267179012299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4589160978794098, + 0.07082971930503845, + 0.056192997843027115, + 0.047574110329151154, + 0.12168000638484955, + 0.11589323729276657, + 0.03871738538146019, + 0.0403071790933609, + 0.01958736591041088, + 0.030301902443170547 + ], + "credit_map": [ + 0.0634978711605072, + 0.06781566888093948, + 0.04771356284618378, + 0.07831192761659622, + 0.04929594695568085, + 0.028832631185650826, + 0.06453020125627518, + 0.058918315917253494, + 0.010009028017520905, + 0.007418571971356869, + 0.012686232104897499, + 0.011437089182436466, + 0.008569341152906418, + 0.005763771943747997, + 0.008296552114188671, + 0.006649133283644915, + 0.008252059109508991, + 0.0067256237380206585, + 0.007665927987545729, + 0.00869485642760992, + 0.005154887679964304, + 0.006406235042959452, + 0.006483176723122597, + 0.006810234393924475, + 0.00621385034173727, + 0.006635017693042755, + 0.010324271395802498, + 0.004622200969606638, + 0.003901862306520343, + 0.0033214320428669453, + 0.006549234967678785, + 0.00600624131038785, + 0.020183008164167404, + 0.015541169792413712, + 0.019397249445319176, + 0.01631002314388752, + 0.015916315838694572, + 0.00789694394916296, + 0.013730926439166069, + 0.012704364955425262, + 0.02016579918563366, + 0.015062698163092136, + 0.0274568609893322, + 0.007739617954939604, + 0.006397889461368322, + 0.01217565592378378, + 0.011128916405141354, + 0.015765804797410965, + 0.005492519587278366, + 0.0041033681482076645, + 0.008467785082757473, + 0.003650043625384569, + 0.0026492378674447536, + 0.005227317567914724, + 0.004376808647066355, + 0.00475030392408371, + 0.006195337511599064, + 0.006884139031171799, + 0.005302546080201864, + 0.0035937197972089052, + 0.0029612358193844557, + 0.0056304014287889, + 0.005783111788332462, + 0.003956688102334738, + 0.003021394368261099, + 0.003119530389085412, + 0.0026124173309653997, + 0.0016706944443285465, + 0.0016817839350551367, + 0.002452095504850149, + 0.002228977857157588, + 0.002800471382215619, + 0.004897975828498602, + 0.0034169156569987535, + 0.004311258438974619, + 0.0022488445974886417, + 0.002659344347193837, + 0.005416930187493563, + 0.004007347859442234, + 0.0033432859927415848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.07905267179012299 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.07900717854499817 + ], + "opposite_logprob_tokens": [ + -2.579007148742676 + ], + "logprob_reward": -0.07900717854499817, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900717854499817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5971941351890564, + 0.13337263464927673, + 0.0721694678068161, + 0.02921481616795063, + 0.031171459704637527, + 0.032848723232746124, + 0.03444676101207733, + 0.02274106815457344, + 0.018651220947504044, + 0.028189746662974358 + ], + "credit_map": [ + 0.07914488762617111, + 0.08496011048555374, + 0.07060763239860535, + 0.10956770926713943, + 0.06428820639848709, + 0.03400270640850067, + 0.08139131218194962, + 0.07323154807090759, + 0.030112620443105698, + 0.008408695459365845, + 0.011317957192659378, + 0.009755597449839115, + 0.007891547866165638, + 0.010349739342927933, + 0.012532142922282219, + 0.04300433024764061, + 0.01027996838092804, + 0.006712449714541435, + 0.010750887915492058, + 0.007078740280121565, + 0.01078513078391552, + 0.008845247328281403, + 0.008734479546546936, + 0.008982566185295582, + 0.0036142850294709206, + 0.0038760399911552668, + 0.00563023379072547, + 0.003841586410999298, + 0.003492004470899701, + 0.0025546280667185783, + 0.0029426515102386475, + 0.003263387130573392, + 0.004291785880923271, + 0.0039354003965854645, + 0.003002422396093607, + 0.0040876069106161594, + 0.004940451588481665, + 0.003186678048223257, + 0.004008724819868803, + 0.0037183882668614388, + 0.003419034881517291, + 0.0031650708988308907, + 0.007248700130730867, + 0.002722569042816758, + 0.0024528237991034985, + 0.004388884641230106, + 0.0035821907222270966, + 0.005869449116289616, + 0.004505649674683809, + 0.0033849759493023157, + 0.009581902995705605, + 0.00399216590449214, + 0.0028464512433856726, + 0.0031721151899546385, + 0.003387672360986471, + 0.0035758307203650475, + 0.003351242281496525, + 0.0025237046647816896, + 0.004528000485152006, + 0.0013426279183477163, + 0.001429416355676949, + 0.003879668889567256, + 0.002985761733725667, + 0.002700646873563528, + 0.0031658944208174944, + 0.0019903224892914295, + 0.002496799686923623, + 0.0021859481930732727, + 0.0016042847419157624, + 0.002409335458651185, + 0.002798496512696147, + 0.0020001381635665894, + 0.005291398148983717, + 0.00299098645336926, + 0.0027688087429851294, + 0.003582830773666501, + 0.0027089351788163185, + 0.003801813116297126, + 0.0034536540042608976, + 0.003591323271393776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07900717854499817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly man's hand reaching toward an ancient tree trunk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the hand covered in wrinkles, light brown age spots, and thinned skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the elderly man wearing a dark gray wool cardigan with frayed cuffs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand exhibit a slight tremor while reaching for a carved indentation in the bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an extremely slow zoom in from a level 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": [ + -0.062039658427238464 + ], + "opposite_logprob_tokens": [ + -2.812039613723755 + ], + "logprob_reward": -0.062039658427238464, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039658427238464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28024882078170776, + 0.4467895030975342, + 0.04485396295785904, + 0.02279304340481758, + 0.02404783107340336, + 0.018778527155518532, + 0.02011028677225113, + 0.05049244314432144, + 0.06731070578098297, + 0.024574900045990944 + ], + "credit_map": [ + 0.03600722551345825, + 0.04257094860076904, + 0.02983083948493004, + 0.04760144650936127, + 0.03060731664299965, + 0.017505090683698654, + 0.0418512336909771, + 0.03427472338080406, + 0.11427929252386093, + 0.02523200213909149, + 0.028866996988654137, + 0.017689377069473267, + 0.024111885577440262, + 0.03561944141983986, + 0.035723280161619186, + 0.1652672290802002, + 0.008764205500483513, + 0.0033014079090207815, + 0.007678270805627108, + 0.004381616599857807, + 0.00734914792701602, + 0.004223617725074291, + 0.004903246182948351, + 0.004252451006323099, + 0.0035369503311812878, + 0.0025053266435861588, + 0.0030928226187825203, + 0.00347764790058136, + 0.0031443366315215826, + 0.0025121422950178385, + 0.0022864837665110826, + 0.002237333683297038, + 0.0033922067377716303, + 0.0030858421232551336, + 0.004116363357752562, + 0.003010332118719816, + 0.003038605907931924, + 0.0019536723848432302, + 0.0031672529876232147, + 0.002283555455505848, + 0.002470310078933835, + 0.0013660471886396408, + 0.0045860749669373035, + 0.0015966969076544046, + 0.0017400054493919015, + 0.001777541358023882, + 0.0022453919518738985, + 0.0029964596033096313, + 0.0025321098510175943, + 0.0021371468901634216, + 0.004172797314822674, + 0.002439844887703657, + 0.0017104438738897443, + 0.002842177636921406, + 0.0021709755528718233, + 0.002104791346937418, + 0.007405204698443413, + 0.005397249478846788, + 0.006853130646049976, + 0.003065345110371709, + 0.0035534529015421867, + 0.010586501099169254, + 0.006695204880088568, + 0.006936356425285339, + 0.008336744271218777, + 0.00781903974711895, + 0.01254637073725462, + 0.004385177046060562, + 0.005205442197620869, + 0.013163185678422451, + 0.008408600464463234, + 0.007446144241839647, + 0.004183795303106308, + 0.0021096330601722, + 0.004072795622050762, + 0.0026071283500641584, + 0.001994194695726037, + 0.003443231573328376, + 0.0036193844862282276, + 0.0025447364896535873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.062039658427238464 + } + ], + "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/0000152/group_9/index_tensor([9013, 9013])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_tensor([9013, 9013])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2055aa253fb7b962e75261529759906b2fe4b763 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_tensor([9013, 9013])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up, the camera remains still, capturing a face draped in heavy clown makeup from a level angle. Thick, slightly cracked white paint covers the entire face. Exaggerated black eyeliner creates a giant, upward teardrop pattern, while deep red lips are deliberately drawn into a mournful, downward curve. Under a cold, focused spotlight from above and side, his unkempt, bright blue wig and bright red, bulbous prosthetic nose appear particularly glaring. Amidst this still, tragic scene, a full, crystalline tear slowly wells up from his bloodshot right eye, trembling and hanging on the edge of his lower lid. Finally, it slips silently, carefully carving a clear, wet path across his painted cheek, breaking through the thick white barrier and revealing the true color of a small patch of skin beneath. The tear mark shimmers in the strong light, forming a devastating contrast with the ridiculously stiff makeup around it, and the whole picture is shrouded in an atmosphere of heartbreak, loneliness and silent sadness.", + "seed": 26668, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_9013_seed_26668_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": -0.1368107189424336, + "vlm_logprob": -0.1368107189424336, + "vlm_yes_no_margin": 3.700000011920929, + "vlm_token_credit": -0.1368107189424336, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023503080010414124 + ], + "opposite_logprob_tokens": [ + -7.377350330352783 + ], + "logprob_reward": -0.0023503080010414124, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0023503080010414124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16368788480758667, + 0.05002753436565399, + 0.047045979648828506, + 0.16346116364002228, + 0.06999258697032928, + 0.08618402481079102, + 0.0836743712425232, + 0.15210941433906555, + 0.12083395570516586, + 0.06298308074474335 + ], + "credit_map": [ + 0.026787804439663887, + 0.01616491749882698, + 0.020581522956490517, + 0.023207977414131165, + 0.01641565002501011, + 0.01744561828672886, + 0.0192706398665905, + 0.02381376549601555, + 0.009956499561667442, + 0.005879757460206747, + 0.008793417364358902, + 0.005438723601400852, + 0.005264637991786003, + 0.005238133482635021, + 0.004693704191595316, + 0.004762670490890741, + 0.009178925305604935, + 0.004603495355695486, + 0.006924154236912727, + 0.005478020757436752, + 0.004413526505231857, + 0.004343749489635229, + 0.006152649875730276, + 0.0059514641761779785, + 0.024599038064479828, + 0.01991906389594078, + 0.01837873086333275, + 0.020411167293787003, + 0.02037147618830204, + 0.021133022382855415, + 0.01354723796248436, + 0.02510143630206585, + 0.011348632164299488, + 0.007007812149822712, + 0.01077745109796524, + 0.00941514316946268, + 0.006615220103412867, + 0.007499255705624819, + 0.007035459391772747, + 0.010293619707226753, + 0.014499378390610218, + 0.008007184602320194, + 0.018530936911702156, + 0.010591596364974976, + 0.008802630007266998, + 0.007692426908761263, + 0.009353898465633392, + 0.008705981075763702, + 0.01417552400380373, + 0.008276904933154583, + 0.010389995761215687, + 0.010097813792526722, + 0.009603515267372131, + 0.008048143237829208, + 0.009292847476899624, + 0.013789638876914978, + 0.02713802456855774, + 0.017519930377602577, + 0.025447463616728783, + 0.02021895721554756, + 0.017419392243027687, + 0.015667157247662544, + 0.016360385343432426, + 0.012338126078248024, + 0.022513587027788162, + 0.010938012972474098, + 0.02999105118215084, + 0.014078420586884022, + 0.007803732063621283, + 0.009542346931993961, + 0.006669430527836084, + 0.019297385588288307, + 0.013097476214170456, + 0.007907337509095669, + 0.008549059741199017, + 0.007850169204175472, + 0.005120547022670507, + 0.0054687135852873325, + 0.004140473436564207, + 0.010849310085177422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0023503080010414124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894476652145386 + ], + "opposite_logprob_tokens": [ + -1.0539448261260986 + ], + "logprob_reward": -0.42894476652145386, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42894476652145386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20250290632247925, + 0.08950699120759964, + 0.09591521322727203, + 0.12464772909879684, + 0.05701065436005592, + 0.04605709761381149, + 0.13611212372779846, + 0.12713973224163055, + 0.06490101665258408, + 0.056206513196229935 + ], + "credit_map": [ + 0.04582161828875542, + 0.026108000427484512, + 0.028510916978120804, + 0.02804305963218212, + 0.01765497773885727, + 0.021404258906841278, + 0.016913477331399918, + 0.018046600744128227, + 0.01617102324962616, + 0.011925291270017624, + 0.012179161421954632, + 0.011902257800102234, + 0.00927924644201994, + 0.010181884281337261, + 0.008010538294911385, + 0.009857582859694958, + 0.01912318728864193, + 0.012386476621031761, + 0.01781478151679039, + 0.013411149382591248, + 0.008008303120732307, + 0.008652815595269203, + 0.007279966026544571, + 0.009238538332283497, + 0.01815568469464779, + 0.015751155093312263, + 0.020369909703731537, + 0.016116322949528694, + 0.01516677439212799, + 0.014618830755352974, + 0.012880138121545315, + 0.011588904075324535, + 0.009495488367974758, + 0.00605347752571106, + 0.008450224064290524, + 0.007701924070715904, + 0.0056225755251944065, + 0.0048299515619874, + 0.004999544005841017, + 0.009857467375695705, + 0.009945154190063477, + 0.00447732163593173, + 0.010782520286738873, + 0.006312574725598097, + 0.003996925428509712, + 0.0036636984441429377, + 0.003234311705455184, + 0.0036445893347263336, + 0.023620283231139183, + 0.016107359901070595, + 0.02180035412311554, + 0.01865443028509617, + 0.016929570585489273, + 0.013254723511636257, + 0.014438210055232048, + 0.011307206936180592, + 0.01956101879477501, + 0.012244882993400097, + 0.016931677237153053, + 0.016739562153816223, + 0.012753311544656754, + 0.014594098553061485, + 0.013352700509130955, + 0.020962469279766083, + 0.009048211388289928, + 0.005900584626942873, + 0.010774454101920128, + 0.00854627974331379, + 0.007217595353722572, + 0.007343826349824667, + 0.007090582046657801, + 0.00897948443889618, + 0.010897164233028889, + 0.006979846861213446, + 0.0077862730249762535, + 0.006442889105528593, + 0.004252857528626919, + 0.004792275372892618, + 0.0047761485911905766, + 0.010279054753482342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.42894476652145386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024966993369162083 + ], + "opposite_logprob_tokens": [ + -6.002496719360352 + ], + "logprob_reward": -0.0024966993369162083, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024966993369162083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2222445011138916, + 0.14313557744026184, + 0.11045416444540024, + 0.15393178164958954, + 0.08281286060810089, + 0.04850105568766594, + 0.06797545403242111, + 0.059613462537527084, + 0.03993840143084526, + 0.07139278948307037 + ], + "credit_map": [ + 0.036739200353622437, + 0.03197065368294716, + 0.023579521104693413, + 0.029222430661320686, + 0.02854214981198311, + 0.026406988501548767, + 0.019184298813343048, + 0.02659924514591694, + 0.025027846917510033, + 0.019531693309545517, + 0.01580219902098179, + 0.016845403239130974, + 0.014740848913788795, + 0.015674201771616936, + 0.01787499710917473, + 0.017638355493545532, + 0.02911374159157276, + 0.010057142935693264, + 0.018032405525445938, + 0.012863298878073692, + 0.00840808730572462, + 0.00844720471650362, + 0.011845557950437069, + 0.011686710640788078, + 0.025926150381565094, + 0.01737949624657631, + 0.018575163558125496, + 0.020329054445028305, + 0.012136143632233143, + 0.017096633091568947, + 0.01036665216088295, + 0.03212246671319008, + 0.0142228277400136, + 0.007686407305300236, + 0.010990943759679794, + 0.009338092058897018, + 0.006439507938921452, + 0.00802392978221178, + 0.009379557333886623, + 0.016731586307287216, + 0.008108403533697128, + 0.00693114660680294, + 0.0051245721988379955, + 0.006691627204418182, + 0.006229554768651724, + 0.005266011692583561, + 0.004129636101424694, + 0.0060200984589755535, + 0.014116122387349606, + 0.008164615370333195, + 0.0097974743694067, + 0.007484473288059235, + 0.007714970503002405, + 0.007152965292334557, + 0.006130550988018513, + 0.007414279505610466, + 0.009085562080144882, + 0.005764279048889875, + 0.010280732996761799, + 0.008085218258202076, + 0.0075507936999201775, + 0.007116289343684912, + 0.005676668602973223, + 0.006053912453353405, + 0.008164091035723686, + 0.005195033270865679, + 0.007538385223597288, + 0.004703321028500795, + 0.0033270777203142643, + 0.00421158829703927, + 0.0026772592682391405, + 0.004121643956750631, + 0.009060926735401154, + 0.008323170244693756, + 0.008760877884924412, + 0.0098758889362216, + 0.00715078366920352, + 0.006774093955755234, + 0.006337127182632685, + 0.015109912492334843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0024966993369162083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159761607646942 + ], + "opposite_logprob_tokens": [ + -1.7015975713729858 + ], + "logprob_reward": -0.20159761607646942, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159761607646942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32383978366851807, + 0.10167855769395828, + 0.07421663403511047, + 0.06997404992580414, + 0.0835474282503128, + 0.0643499568104744, + 0.04854888841509819, + 0.15595577657222748, + 0.028376203030347824, + 0.049512725323438644 + ], + "credit_map": [ + 0.0628974586725235, + 0.03981301188468933, + 0.062448900192976, + 0.03712986409664154, + 0.031495608389377594, + 0.029212813824415207, + 0.02546921744942665, + 0.035372890532016754, + 0.017223522067070007, + 0.012212426401674747, + 0.016299325972795486, + 0.014475852251052856, + 0.01033248845487833, + 0.011086523532867432, + 0.007641654461622238, + 0.01240676548331976, + 0.01306996587663889, + 0.009220384061336517, + 0.014924787916243076, + 0.009234484285116196, + 0.007050157990306616, + 0.007096261717379093, + 0.005506924353539944, + 0.008113672956824303, + 0.011212802492082119, + 0.008549656718969345, + 0.01008378341794014, + 0.008354372344911098, + 0.007753858342766762, + 0.00851866789162159, + 0.006723369006067514, + 0.008777539245784283, + 0.015381241217255592, + 0.009655740112066269, + 0.018401719629764557, + 0.00852672103792429, + 0.007409315090626478, + 0.007123306859284639, + 0.005726322997361422, + 0.011323061771690845, + 0.008345214650034904, + 0.007543470244854689, + 0.009971218183636665, + 0.008908284828066826, + 0.009613078087568283, + 0.007446273695677519, + 0.006632146891206503, + 0.00589027302339673, + 0.011359472759068012, + 0.005623371805995703, + 0.0059443372301757336, + 0.005553863476961851, + 0.004845289513468742, + 0.004222513176500797, + 0.004445972386747599, + 0.006554067600518465, + 0.023730870336294174, + 0.014270458370447159, + 0.03260822966694832, + 0.019354848191142082, + 0.018711645156145096, + 0.017106331884860992, + 0.013578804209828377, + 0.016594598069787025, + 0.006508920807391405, + 0.0028094525914639235, + 0.0078451381996274, + 0.0027869408950209618, + 0.0018427062314003706, + 0.0016819166485220194, + 0.0023261450696736574, + 0.002574982587248087, + 0.007451208774000406, + 0.004476345144212246, + 0.006189420353621244, + 0.007209230214357376, + 0.0044504026882350445, + 0.005068756174296141, + 0.005060733295977116, + 0.009606631472706795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.20159761607646942 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866420477628708 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866420477628708, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866420477628708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3281584680080414, + 0.11753325164318085, + 0.08711321651935577, + 0.12056718021631241, + 0.08476855605840683, + 0.01627727784216404, + 0.034055955708026886, + 0.02823304943740368, + 0.1265922486782074, + 0.056700803339481354 + ], + "credit_map": [ + 0.052272729575634, + 0.05132131651043892, + 0.04713565483689308, + 0.04542110860347748, + 0.03689385950565338, + 0.03169472515583038, + 0.028314942494034767, + 0.035104140639305115, + 0.025375530123710632, + 0.013648039661347866, + 0.022096775472164154, + 0.014663271605968475, + 0.007843974977731705, + 0.010550077073276043, + 0.00971007626503706, + 0.013645505532622337, + 0.016936669126152992, + 0.0096462182700634, + 0.012278134003281593, + 0.011656955815851688, + 0.007434174884110689, + 0.009445343166589737, + 0.007004715036600828, + 0.01271100528538227, + 0.018766190856695175, + 0.01414477452635765, + 0.01426600106060505, + 0.014374567195773125, + 0.013750945217907429, + 0.014845400117337704, + 0.011005183681845665, + 0.01941411755979061, + 0.01375641580671072, + 0.008087302558124065, + 0.012629115954041481, + 0.010854092426598072, + 0.00766103295609355, + 0.008952350355684757, + 0.009704310446977615, + 0.013123946264386177, + 0.002886029426008463, + 0.0022698405664414167, + 0.0018704956164583564, + 0.002324837725609541, + 0.001381122274324298, + 0.0019217318622395396, + 0.0015195158775895834, + 0.0021037051919847727, + 0.006837763823568821, + 0.0046233213506639, + 0.004699941258877516, + 0.004013533238321543, + 0.003518905257806182, + 0.00408801157027483, + 0.0026133633218705654, + 0.0036611156538128853, + 0.00410094391554594, + 0.003558881115168333, + 0.003961721435189247, + 0.003988686483353376, + 0.0034160471986979246, + 0.002650998532772064, + 0.0026191703509539366, + 0.003936599940061569, + 0.023764802142977715, + 0.02194296009838581, + 0.018339956179261208, + 0.018009789288043976, + 0.011657031252980232, + 0.011748929508030415, + 0.00951238814741373, + 0.01161638367921114, + 0.00909820944070816, + 0.008613689802587032, + 0.008951997384428978, + 0.008769703097641468, + 0.006397450342774391, + 0.004640104249119759, + 0.004553237929940224, + 0.005676410626620054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.04866420477628708 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, capturing a face draped in heavy clown makeup from a level angle. Thick, slightly cracked white paint covers the entire face. Exaggerated black eyeliner creates a giant, upward teardrop pattern, while deep red lips are deliberately drawn into a mournful, downward curve. Under a cold, focused spotlight from above and side, his unkempt, bright blue wig and bright red, bulbous prosthetic nose appear particularly glaring. Amidst this still, tragic scene, a full, crystalline tear slowly wells up from his bloodshot right eye, trembling and hanging on the edge of his lower lid. Finally, it slips silently, carefully carving a clear, wet path across his painted cheek, breaking through the thick white barrier and revealing the true color of a small patch of skin beneath. The tear mark shimmers in the strong light, forming a devastating contrast with the ridiculously stiff makeup around it, and the whole picture is shrouded in an atmosphere of heartbreak, loneliness and silent sadness.", + "seed": 26669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_9013_seed_26669_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": -0.12610434070229531, + "vlm_logprob": -0.12610434070229531, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.12610434070229531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003459423314779997 + ], + "opposite_logprob_tokens": [ + -7.378459453582764 + ], + "logprob_reward": -0.003459423314779997, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.003459423314779997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13873495161533356, + 0.06187838688492775, + 0.058854278177022934, + 0.030146028846502304, + 0.0703662782907486, + 0.14144451916217804, + 0.3084774613380432, + 0.10130814462900162, + 0.06531033664941788, + 0.023479660972952843 + ], + "credit_map": [ + 0.01524435356259346, + 0.017837682738900185, + 0.01669759303331375, + 0.016311518847942352, + 0.01594020612537861, + 0.025352718308568, + 0.014381972141563892, + 0.01696890778839588, + 0.006359748542308807, + 0.006353064905852079, + 0.007775107864290476, + 0.00470293452963233, + 0.00521615706384182, + 0.016299083828926086, + 0.0061754523776471615, + 0.008996828459203243, + 0.006840122397989035, + 0.0061637479811906815, + 0.007678101304918528, + 0.0046427068300545216, + 0.0056229084730148315, + 0.01122472994029522, + 0.006963028106838465, + 0.009718925692141056, + 0.002823758637532592, + 0.0025635624770075083, + 0.0032351426780223846, + 0.0032318795565515757, + 0.0032379215117543936, + 0.007614195346832275, + 0.0028888077940791845, + 0.004550758749246597, + 0.007846424356102943, + 0.009404096752405167, + 0.007836421020328999, + 0.008421339094638824, + 0.006989616435021162, + 0.01163418684154749, + 0.007976477034389973, + 0.010257710702717304, + 0.01616566628217697, + 0.01587114855647087, + 0.01587100699543953, + 0.017108242958784103, + 0.01609315164387226, + 0.023617254570126534, + 0.018826769664883614, + 0.01789126917719841, + 0.036003585904836655, + 0.03090791590511799, + 0.03455958887934685, + 0.03179434686899185, + 0.03064104914665222, + 0.06410707533359528, + 0.038561947643756866, + 0.041901908814907074, + 0.013205179944634438, + 0.011056493036448956, + 0.011511848308146, + 0.01210872270166874, + 0.010029289871454239, + 0.016792379319667816, + 0.012588203884661198, + 0.014016025699675083, + 0.006735168397426605, + 0.009511719457805157, + 0.005704273469746113, + 0.00680880481377244, + 0.007203863933682442, + 0.012007327750325203, + 0.008780144155025482, + 0.008559029549360275, + 0.0024204482324421406, + 0.001880176831036806, + 0.002209545811638236, + 0.0017284220084547997, + 0.002469446510076523, + 0.0067572761327028275, + 0.002438394585624337, + 0.0035759471356868744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.003459423314779997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28140053153038025 + ], + "opposite_logprob_tokens": [ + -1.4064005613327026 + ], + "logprob_reward": -0.28140053153038025, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28140053153038025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0822581797838211, + 0.029989058151841164, + 0.04360222443938255, + 0.0500061921775341, + 0.036005839705467224, + 0.10678794980049133, + 0.054010506719350815, + 0.3717907965183258, + 0.2140919715166092, + 0.011457334272563457 + ], + "credit_map": [ + 0.008428300730884075, + 0.007440883666276932, + 0.009500805288553238, + 0.007867944426834583, + 0.007641719654202461, + 0.018353639170527458, + 0.008976638317108154, + 0.014048249460756779, + 0.0030350990127772093, + 0.004287345334887505, + 0.0031599518842995167, + 0.0032115159556269646, + 0.0030609755776822567, + 0.005505464505404234, + 0.0034600344952195883, + 0.004268673248589039, + 0.0050029633566737175, + 0.0031590070575475693, + 0.006060704588890076, + 0.0038675805553793907, + 0.003840523073449731, + 0.010563922114670277, + 0.004184874705970287, + 0.006922651082277298, + 0.005678766872733831, + 0.003344351891428232, + 0.006349056959152222, + 0.004368618596345186, + 0.004577501676976681, + 0.01306461077183485, + 0.005352547857910395, + 0.00727074034512043, + 0.0033178143203258514, + 0.005676994565874338, + 0.002882895525544882, + 0.003125166054815054, + 0.003746951464563608, + 0.008476625196635723, + 0.003984570037573576, + 0.004794823005795479, + 0.011088232509791851, + 0.009777183644473553, + 0.01260445173829794, + 0.013180156238377094, + 0.0127264941111207, + 0.01922239363193512, + 0.013643577694892883, + 0.014545452781021595, + 0.0053373402915894985, + 0.005037591326981783, + 0.00588030694052577, + 0.006191691383719444, + 0.007356530986726284, + 0.00907142460346222, + 0.0067352489568293095, + 0.008400371298193932, + 0.04099345579743385, + 0.031041834503412247, + 0.044575031846761703, + 0.03652510046958923, + 0.04108773544430733, + 0.08146164566278458, + 0.04431281238794327, + 0.0517931766808033, + 0.022580856457352638, + 0.022468749433755875, + 0.021915512159466743, + 0.019448915496468544, + 0.021122541278600693, + 0.052258703857660294, + 0.02629518322646618, + 0.028001507744193077, + 0.0010298459092155099, + 0.0015783096896484494, + 0.0009896617848426104, + 0.0012555258581414819, + 0.0018095606938004494, + 0.0015931135276332498, + 0.0016457957681268454, + 0.0015555205754935741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.28140053153038025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022080582566559315 + ], + "opposite_logprob_tokens": [ + -6.127208232879639 + ], + "logprob_reward": -0.0022080582566559315, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022080582566559315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2026926428079605, + 0.28761744499206543, + 0.11817583441734314, + 0.08342216908931732, + 0.02931472286581993, + 0.04752439633011818, + 0.0911770612001419, + 0.06721080094575882, + 0.031777169555425644, + 0.04108786582946777 + ], + "credit_map": [ + 0.020137488842010498, + 0.022042499855160713, + 0.02229556068778038, + 0.01858578436076641, + 0.021223224699497223, + 0.040603019297122955, + 0.025582538917660713, + 0.03222251310944557, + 0.030868228524923325, + 0.019471393898129463, + 0.04030691832304001, + 0.030025480315089226, + 0.024082401767373085, + 0.0638812929391861, + 0.03194398060441017, + 0.04703770950436592, + 0.012241492047905922, + 0.01002528052777052, + 0.015324333682656288, + 0.013330335728824139, + 0.009746734984219074, + 0.026782061904668808, + 0.012874969281256199, + 0.017850615084171295, + 0.008577500469982624, + 0.007555688731372356, + 0.01088607870042324, + 0.00850340723991394, + 0.00702859228476882, + 0.01805741898715496, + 0.010254370048642159, + 0.01255910936743021, + 0.002498325891792774, + 0.003269225126132369, + 0.0029563538264483213, + 0.0035811911802738905, + 0.003232054878026247, + 0.0056454394944012165, + 0.0038773592095822096, + 0.004254771396517754, + 0.0035243649035692215, + 0.009170474484562874, + 0.004087342415004969, + 0.004889866337180138, + 0.004617171827703714, + 0.009237266145646572, + 0.006431677378714085, + 0.0055662281811237335, + 0.010441274382174015, + 0.00901968963444233, + 0.010426335968077183, + 0.009941629134118557, + 0.009638965129852295, + 0.01786717213690281, + 0.010963652282953262, + 0.012878328561782837, + 0.00842992588877678, + 0.006295458413660526, + 0.007642177399247885, + 0.007549557834863663, + 0.006254079286009073, + 0.013963164761662483, + 0.008097621612250805, + 0.008978812023997307, + 0.003263551741838455, + 0.0036816373467445374, + 0.002576592145487666, + 0.003428712720051408, + 0.0034981868229806423, + 0.006638430990278721, + 0.004548958037048578, + 0.004141097888350487, + 0.003039761446416378, + 0.004429491236805916, + 0.0036329193972051144, + 0.004805402364581823, + 0.007472729776054621, + 0.005548846907913685, + 0.006918189115822315, + 0.005240525119006634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0022080582566559315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281386137008667 + ], + "opposite_logprob_tokens": [ + -1.406386137008667 + ], + "logprob_reward": -0.281386137008667, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281386137008667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23749078810214996, + 0.07526218891143799, + 0.0662805438041687, + 0.08543016761541367, + 0.04147753491997719, + 0.06947778910398483, + 0.06360575556755066, + 0.21508929133415222, + 0.11676578223705292, + 0.029120154678821564 + ], + "credit_map": [ + 0.022660454735159874, + 0.025429146364331245, + 0.029164277017116547, + 0.021920282393693924, + 0.021566106006503105, + 0.04527302831411362, + 0.03098171018064022, + 0.040495771914720535, + 0.007769871037453413, + 0.011391432024538517, + 0.0057874396443367004, + 0.008016262203454971, + 0.010804850608110428, + 0.01219252496957779, + 0.009219528175890446, + 0.010080285370349884, + 0.007598075550049543, + 0.012545354664325714, + 0.0059030367992818356, + 0.008056120947003365, + 0.00814944040030241, + 0.008326135575771332, + 0.007855361327528954, + 0.007847017608582973, + 0.009873389266431332, + 0.009389250539243221, + 0.009077602997422218, + 0.010146771557629108, + 0.0081766527146101, + 0.016404226422309875, + 0.008968714624643326, + 0.013393567875027657, + 0.0041109854355454445, + 0.006118247751146555, + 0.0033402356784790754, + 0.006947127170860767, + 0.0043209814466536045, + 0.005181893240660429, + 0.005435737781226635, + 0.006022326182574034, + 0.006972484290599823, + 0.00725964829325676, + 0.007634888868778944, + 0.00729470606893301, + 0.0071863820776343346, + 0.01558051910251379, + 0.007742787711322308, + 0.009806371293962002, + 0.006628882139921188, + 0.006774278823286295, + 0.0062639787793159485, + 0.007950836792588234, + 0.008621230721473694, + 0.010130790993571281, + 0.007593479473143816, + 0.009642270393669605, + 0.024615300819277763, + 0.01751251332461834, + 0.024533776566386223, + 0.022276900708675385, + 0.023962263017892838, + 0.048085324466228485, + 0.024388721212744713, + 0.029714500531554222, + 0.012903936207294464, + 0.010439305566251278, + 0.011191741563379765, + 0.011845423839986324, + 0.011732891201972961, + 0.029347652569413185, + 0.013982818461954594, + 0.015322008170187473, + 0.002364601707085967, + 0.0028930932749062777, + 0.002722158096730709, + 0.002949425019323826, + 0.004710640758275986, + 0.005280951503664255, + 0.004189486615359783, + 0.0040097967721521854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.281386137008667 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206755340099335 + ], + "opposite_logprob_tokens": [ + -2.8120675086975098 + ], + "logprob_reward": -0.06206755340099335, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206755340099335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24674369394779205, + 0.10271035879850388, + 0.07075165957212448, + 0.2298528254032135, + 0.11883118748664856, + 0.047700315713882446, + 0.038460616022348404, + 0.03146285191178322, + 0.06716036051511765, + 0.046326130628585815 + ], + "credit_map": [ + 0.02646312117576599, + 0.023077191784977913, + 0.029398897662758827, + 0.024453099817037582, + 0.02712966501712799, + 0.05539699271321297, + 0.026978425681591034, + 0.03384631127119064, + 0.011390702798962593, + 0.009301296435296535, + 0.014410614967346191, + 0.008045322261750698, + 0.007271098904311657, + 0.023900559172034264, + 0.01102416031062603, + 0.017366603016853333, + 0.0069705164059996605, + 0.008586460724473, + 0.007089261896908283, + 0.008560623973608017, + 0.006741047371178865, + 0.012398257851600647, + 0.009847762063145638, + 0.010557731613516808, + 0.02865530550479889, + 0.016901616007089615, + 0.028634088113904, + 0.020389776676893234, + 0.02019231766462326, + 0.05847163498401642, + 0.025210732594132423, + 0.03139733895659447, + 0.015595567412674427, + 0.009776351973414421, + 0.014622007496654987, + 0.010631522163748741, + 0.009980901144444942, + 0.028700167313218117, + 0.012578429654240608, + 0.016946233808994293, + 0.004609635565429926, + 0.007692510727792978, + 0.00479359645396471, + 0.00569124473258853, + 0.004721977282315493, + 0.008703615516424179, + 0.004767272155731916, + 0.0067204637452960014, + 0.004502867814153433, + 0.004651532042771578, + 0.0042933691293001175, + 0.0036579768639057875, + 0.003530218033120036, + 0.0074427155777812, + 0.005425528157502413, + 0.004956407006829977, + 0.003788911970332265, + 0.0035618862602859735, + 0.002389125758782029, + 0.0032755755819380283, + 0.0032970993779599667, + 0.006006746087223291, + 0.004301481880247593, + 0.004842027090489864, + 0.006420159246772528, + 0.006846505217254162, + 0.005986787844449282, + 0.006312539801001549, + 0.00647825188934803, + 0.018665604293346405, + 0.005616092588752508, + 0.010834417305886745, + 0.003987288102507591, + 0.003572623711079359, + 0.00477175135165453, + 0.005128433462232351, + 0.006305763963609934, + 0.010236034169793129, + 0.0059350719675421715, + 0.006389163434505463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.06206755340099335 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, capturing a face draped in heavy clown makeup from a level angle. Thick, slightly cracked white paint covers the entire face. Exaggerated black eyeliner creates a giant, upward teardrop pattern, while deep red lips are deliberately drawn into a mournful, downward curve. Under a cold, focused spotlight from above and side, his unkempt, bright blue wig and bright red, bulbous prosthetic nose appear particularly glaring. Amidst this still, tragic scene, a full, crystalline tear slowly wells up from his bloodshot right eye, trembling and hanging on the edge of his lower lid. Finally, it slips silently, carefully carving a clear, wet path across his painted cheek, breaking through the thick white barrier and revealing the true color of a small patch of skin beneath. The tear mark shimmers in the strong light, forming a devastating contrast with the ridiculously stiff makeup around it, and the whole picture is shrouded in an atmosphere of heartbreak, loneliness and silent sadness.", + "seed": 26670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_9013_seed_26670_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": -0.1651513150893152, + "vlm_logprob": -0.1651513150893152, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1651513150893152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004541677888482809 + ], + "opposite_logprob_tokens": [ + -7.004541873931885 + ], + "logprob_reward": -0.004541677888482809, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.004541677888482809, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10459258407354355, + 0.06466161459684372, + 0.04968436807394028, + 0.0874418318271637, + 0.03555620089173317, + 0.04324934631586075, + 0.21298280358314514, + 0.22578100860118866, + 0.03287747874855995, + 0.14317280054092407 + ], + "credit_map": [ + 0.007749869488179684, + 0.01002077292650938, + 0.012196787633001804, + 0.010172337293624878, + 0.01946176029741764, + 0.01632848009467125, + 0.01149076595902443, + 0.017171811312437057, + 0.005212646909058094, + 0.008163751102983952, + 0.005172960925847292, + 0.006506394594907761, + 0.011303689330816269, + 0.012717143632471561, + 0.006930477451533079, + 0.008654550649225712, + 0.0034098767209798098, + 0.0037414247635751963, + 0.007716410327702761, + 0.0048940423876047134, + 0.004533730447292328, + 0.008929713629186153, + 0.004222430754452944, + 0.012236740440130234, + 0.007453028578311205, + 0.007588539272546768, + 0.010700138285756111, + 0.007326084189116955, + 0.009293368086218834, + 0.014736348763108253, + 0.00692004943266511, + 0.02342427894473076, + 0.004132305737584829, + 0.003070772159844637, + 0.0035586475860327482, + 0.002791219623759389, + 0.004295542370527983, + 0.006521590985357761, + 0.003680382389575243, + 0.007505736779421568, + 0.0023111787158995867, + 0.0045556738041341305, + 0.002454675966873765, + 0.0034707284066826105, + 0.007390504237264395, + 0.010178768076002598, + 0.0054095773957669735, + 0.007478238549083471, + 0.017368663102388382, + 0.020079253241419792, + 0.025072423741221428, + 0.010585104115307331, + 0.028814805671572685, + 0.030352842062711716, + 0.04278295487165451, + 0.03792676329612732, + 0.016295406967401505, + 0.009667571634054184, + 0.03009217232465744, + 0.011188577860593796, + 0.019146859645843506, + 0.04679349809885025, + 0.01445180643349886, + 0.0781451165676117, + 0.0035849784035235643, + 0.002611906500533223, + 0.0032227097544819117, + 0.0027554889675229788, + 0.006074895150959492, + 0.006667216308414936, + 0.003179854014888406, + 0.004780428484082222, + 0.009421329014003277, + 0.007966220378875732, + 0.016794631257653236, + 0.007715635467320681, + 0.013347968459129333, + 0.04152246192097664, + 0.014989189803600311, + 0.03141535073518753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.004541677888482809 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34872376918792725 + ], + "opposite_logprob_tokens": [ + -1.2237237691879272 + ], + "logprob_reward": -0.34872376918792725, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34872376918792725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0939197987318039, + 0.06080176681280136, + 0.0818151980638504, + 0.07705111801624298, + 0.10317160934209824, + 0.034766070544719696, + 0.1449117362499237, + 0.06597881019115448, + 0.07596079260110855, + 0.2616230249404907 + ], + "credit_map": [ + 0.00759120611473918, + 0.006890762597322464, + 0.005997140891849995, + 0.007616252638399601, + 0.014293781481683254, + 0.02143906056880951, + 0.012166345492005348, + 0.017925260588526726, + 0.003712455974891782, + 0.002998195355758071, + 0.01274092961102724, + 0.003191307419911027, + 0.005041038151830435, + 0.01081534568220377, + 0.0035213970113545656, + 0.018781105056405067, + 0.005498168524354696, + 0.0037212884053587914, + 0.01627976819872856, + 0.004530252888798714, + 0.005559003446251154, + 0.015088830143213272, + 0.005269525106996298, + 0.025868376716971397, + 0.0050710332579910755, + 0.007083933334797621, + 0.005332334898412228, + 0.006399073172360659, + 0.01293692272156477, + 0.015278580598533154, + 0.010042632929980755, + 0.0149066261947155, + 0.009384484961628914, + 0.005135547369718552, + 0.008601649664342403, + 0.00658109225332737, + 0.019868580624461174, + 0.02301517315208912, + 0.010118737816810608, + 0.020466359332203865, + 0.002676375675946474, + 0.0033637150190770626, + 0.005070481915026903, + 0.0028016124852001667, + 0.004026285372674465, + 0.0069702318869531155, + 0.004461115226149559, + 0.005396257620304823, + 0.013356838375329971, + 0.012537280097603798, + 0.013520908541977406, + 0.007275584619492292, + 0.017504330724477768, + 0.02787705883383751, + 0.0231952928006649, + 0.02964446134865284, + 0.006969055626541376, + 0.005111660808324814, + 0.007561521604657173, + 0.0051401094533503056, + 0.009449736215174198, + 0.012527077458798885, + 0.007009175606071949, + 0.012210486456751823, + 0.007436882704496384, + 0.005338483955711126, + 0.005987252574414015, + 0.006201924290508032, + 0.012380026280879974, + 0.01420859806239605, + 0.007585152518004179, + 0.016822488978505135, + 0.015589666552841663, + 0.012515844777226448, + 0.023232024163007736, + 0.010836327448487282, + 0.024978959932923317, + 0.0962359756231308, + 0.03074428252875805, + 0.0474899560213089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.34872376918792725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062064528465270996 + ], + "opposite_logprob_tokens": [ + -2.8120646476745605 + ], + "logprob_reward": -0.062064528465270996, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062064528465270996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05911475047469139, + 0.07993198186159134, + 0.0749940425157547, + 0.16646207869052887, + 0.1652173101902008, + 0.048939384520053864, + 0.2214023321866989, + 0.0688825398683548, + 0.045385055243968964, + 0.06967058032751083 + ], + "credit_map": [ + 0.006154465489089489, + 0.008954033255577087, + 0.005146747455000877, + 0.007095259614288807, + 0.007874824106693268, + 0.00779040576890111, + 0.007772544864565134, + 0.008326468989253044, + 0.004801768343895674, + 0.005451289936900139, + 0.014459799975156784, + 0.004621326923370361, + 0.005383373238146305, + 0.014688045717775822, + 0.0049376338720321655, + 0.025588734075427055, + 0.00557875307276845, + 0.007907828316092491, + 0.009717282839119434, + 0.006382290739566088, + 0.00772804394364357, + 0.01223455835133791, + 0.007025142665952444, + 0.018420137465000153, + 0.009593384340405464, + 0.005422262940555811, + 0.027081536129117012, + 0.007666413672268391, + 0.013854341581463814, + 0.03223865479230881, + 0.008379444479942322, + 0.06222601234912872, + 0.013299725018441677, + 0.009625242091715336, + 0.014108401723206043, + 0.012332204729318619, + 0.029176140204072, + 0.03494761139154434, + 0.01728135161101818, + 0.03444661572575569, + 0.004200642928481102, + 0.004671081900596619, + 0.00741978595033288, + 0.002750441199168563, + 0.008254987187683582, + 0.008589446544647217, + 0.006168147083371878, + 0.006884847767651081, + 0.01706370711326599, + 0.00978051871061325, + 0.04950174689292908, + 0.009541185572743416, + 0.02040811814367771, + 0.03747452050447464, + 0.044141415506601334, + 0.033491101115942, + 0.0061798011884093285, + 0.004165826365351677, + 0.007504178676754236, + 0.005039301700890064, + 0.007413611281663179, + 0.01432033907622099, + 0.005822970066219568, + 0.018436502665281296, + 0.003112100064754486, + 0.002749179257079959, + 0.003511951072141528, + 0.003308675019070506, + 0.00672430032864213, + 0.013368586078286171, + 0.003932901658117771, + 0.008677361533045769, + 0.005125932861119509, + 0.004825942683964968, + 0.007503192871809006, + 0.005960504990071058, + 0.009007425047457218, + 0.01659371145069599, + 0.009603237733244896, + 0.011050625704228878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.062064528465270996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871288597583771 + ], + "opposite_logprob_tokens": [ + -1.1371288299560547 + ], + "logprob_reward": -0.3871288597583771, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871288597583771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05411543324589729, + 0.12100065499544144, + 0.0847889631986618, + 0.19094619154930115, + 0.13914579153060913, + 0.04815973341464996, + 0.2306683212518692, + 0.03431643173098564, + 0.06798259168863297, + 0.02887580171227455 + ], + "credit_map": [ + 0.00519917719066143, + 0.005934101063758135, + 0.005023036152124405, + 0.00510226609185338, + 0.007516076322644949, + 0.00999362301081419, + 0.005816732067614794, + 0.009530426934361458, + 0.008088178001344204, + 0.007009417284280062, + 0.023335041478276253, + 0.005707481876015663, + 0.010017486289143562, + 0.022421156987547874, + 0.007317470386624336, + 0.03710443526506424, + 0.006521223578602076, + 0.009869429282844067, + 0.010954474098980427, + 0.008587973192334175, + 0.009433483704924583, + 0.011873148381710052, + 0.010054721496999264, + 0.017494508996605873, + 0.011081606149673462, + 0.006388826761394739, + 0.03010817989706993, + 0.009286350570619106, + 0.016019225120544434, + 0.03665373846888542, + 0.012347683310508728, + 0.06906060874462128, + 0.01409054920077324, + 0.0068793995305895805, + 0.011211123317480087, + 0.011377731338143349, + 0.022971026599407196, + 0.030751315876841545, + 0.011233980767428875, + 0.030630674213171005, + 0.0031671500764787197, + 0.004337413236498833, + 0.0044069006107747555, + 0.0036778890062123537, + 0.005997894331812859, + 0.009726649150252342, + 0.006127271335572004, + 0.010718567296862602, + 0.01842290349304676, + 0.013017315417528152, + 0.058595918118953705, + 0.010158883407711983, + 0.019154170528054237, + 0.033038314431905746, + 0.036768991500139236, + 0.04151184484362602, + 0.0033985718619078398, + 0.003244150895625353, + 0.0030486653558909893, + 0.003255164250731468, + 0.003881344571709633, + 0.007033545523881912, + 0.003316816408187151, + 0.007138175889849663, + 0.006144464015960693, + 0.004920667503029108, + 0.0049814037047326565, + 0.006099062506109476, + 0.009724280796945095, + 0.01585230976343155, + 0.0062116337940096855, + 0.014048776589334011, + 0.003023450495675206, + 0.0017897572834044695, + 0.0018926759948953986, + 0.0016069983830675483, + 0.002903294051066041, + 0.005582680460065603, + 0.003871331922709942, + 0.008205616846680641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.3871288597583771 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039525095373392105, + 0.04055190458893776, + 0.04033823683857918, + 0.21187347173690796, + 0.12651154398918152, + 0.039749182760715485, + 0.17920497059822083, + 0.051872819662094116, + 0.20103339850902557, + 0.06933940201997757 + ], + "credit_map": [ + 0.005515206605195999, + 0.004685383290052414, + 0.003310644766315818, + 0.003910167142748833, + 0.0051796953193843365, + 0.007384310942143202, + 0.0045147002674639225, + 0.005024987272918224, + 0.0030185312498360872, + 0.0033029350452125072, + 0.0046014199033379555, + 0.0029320342000573874, + 0.004515255335718393, + 0.008501085452735424, + 0.003474221797659993, + 0.010206418111920357, + 0.002836112165823579, + 0.004943490028381348, + 0.0035592413041740656, + 0.0036720663774758577, + 0.005797250661998987, + 0.007628986146301031, + 0.004188227467238903, + 0.007712864316999912, + 0.01347539946436882, + 0.006449699867516756, + 0.03140586242079735, + 0.009882877580821514, + 0.01922670006752014, + 0.040929440408945084, + 0.012932505458593369, + 0.07757099717855453, + 0.01186364982277155, + 0.007816880941390991, + 0.010512604378163815, + 0.0072993808425962925, + 0.025276968255639076, + 0.027199724689126015, + 0.014879376627504826, + 0.021662956103682518, + 0.00270601618103683, + 0.0066725630313158035, + 0.004846483003348112, + 0.0041247145272791386, + 0.00561909843236208, + 0.005774322431534529, + 0.004953048657625914, + 0.0050529371947050095, + 0.010769633576273918, + 0.011106395162642002, + 0.011805333197116852, + 0.008163567632436752, + 0.01810777187347412, + 0.06282290071249008, + 0.024654243141412735, + 0.031775109469890594, + 0.0067690336145460606, + 0.003763840300962329, + 0.006962804589420557, + 0.0041811419650912285, + 0.007108698133379221, + 0.009245580993592739, + 0.004892745520919561, + 0.00894897524267435, + 0.02962060272693634, + 0.015367511659860611, + 0.019926028326153755, + 0.014911064878106117, + 0.03032604046165943, + 0.037324488162994385, + 0.016014860942959785, + 0.03754279762506485, + 0.0050388723611831665, + 0.004630019888281822, + 0.00735294446349144, + 0.005430377554148436, + 0.007624911610037088, + 0.0201931893825531, + 0.007423484232276678, + 0.011645602993667126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.023297740146517754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up, the camera remains still, capturing a face draped in heavy clown makeup from a level angle. Thick, slightly cracked white paint covers the entire face. Exaggerated black eyeliner creates a giant, upward teardrop pattern, while deep red lips are deliberately drawn into a mournful, downward curve. Under a cold, focused spotlight from above and side, his unkempt, bright blue wig and bright red, bulbous prosthetic nose appear particularly glaring. Amidst this still, tragic scene, a full, crystalline tear slowly wells up from his bloodshot right eye, trembling and hanging on the edge of his lower lid. Finally, it slips silently, carefully carving a clear, wet path across his painted cheek, breaking through the thick white barrier and revealing the true color of a small patch of skin beneath. The tear mark shimmers in the strong light, forming a devastating contrast with the ridiculously stiff makeup around it, and the whole picture is shrouded in an atmosphere of heartbreak, loneliness and silent sadness.", + "seed": 26671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000152/group_9/index_9013_seed_26671_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": -0.15221447832882404, + "vlm_logprob": -0.15221447832882404, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.15221447832882404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0029226704500615597 + ], + "opposite_logprob_tokens": [ + -7.252922534942627 + ], + "logprob_reward": -0.0029226704500615597, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0029226704500615597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14092977344989777, + 0.11611979454755783, + 0.22397935390472412, + 0.05944451689720154, + 0.05738355219364166, + 0.044911421835422516, + 0.07935947924852371, + 0.04372918978333473, + 0.01258634589612484, + 0.22155654430389404 + ], + "credit_map": [ + 0.02382301725447178, + 0.005601551383733749, + 0.03787308186292648, + 0.004453422501683235, + 0.005575949791818857, + 0.018232129514217377, + 0.004608893766999245, + 0.04076172411441803, + 0.024942034855484962, + 0.005826860200613737, + 0.02720985934138298, + 0.004459073301404715, + 0.0046351272612810135, + 0.016503307968378067, + 0.00361367454752326, + 0.028929855674505234, + 0.04529281333088875, + 0.013711913488805294, + 0.03429417312145233, + 0.02053322270512581, + 0.016320379450917244, + 0.038455426692962646, + 0.01885216310620308, + 0.03651925548911095, + 0.004943046718835831, + 0.0023379677440971136, + 0.019005749374628067, + 0.0032981191761791706, + 0.0022841603495180607, + 0.004121797624975443, + 0.0029369161929935217, + 0.020516762509942055, + 0.008098861202597618, + 0.0042084259912371635, + 0.012457103468477726, + 0.006005348637700081, + 0.003783547319471836, + 0.006571835372596979, + 0.004005954135209322, + 0.01225247886031866, + 0.004907690454274416, + 0.004556844476610422, + 0.010241254232823849, + 0.0032117993105202913, + 0.004402198363095522, + 0.00595657667145133, + 0.0026078529190272093, + 0.009027203544974327, + 0.01315203495323658, + 0.007975720800459385, + 0.01209237053990364, + 0.010106670670211315, + 0.005865213926881552, + 0.010521318763494492, + 0.009414132684469223, + 0.010232021100819111, + 0.007167517673224211, + 0.004592706449329853, + 0.007199693936854601, + 0.003478736849501729, + 0.004097312688827515, + 0.005294909700751305, + 0.0045853666961193085, + 0.0073129464872181416, + 0.0015287298010662198, + 0.0010159978410229087, + 0.0024146749638020992, + 0.0012000498827546835, + 0.0008101053535938263, + 0.001905165845528245, + 0.001013413304463029, + 0.0026982100680470467, + 0.014749577268958092, + 0.0038930773735046387, + 0.08912121504545212, + 0.006746550090610981, + 0.005228999070823193, + 0.009383671917021275, + 0.002970075933262706, + 0.08946338295936584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0029226704500615597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871184289455414 + ], + "opposite_logprob_tokens": [ + -1.1371184587478638 + ], + "logprob_reward": -0.3871184289455414, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871184289455414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17111878097057343, + 0.044457174837589264, + 0.06156158074736595, + 0.4194207787513733, + 0.031242653727531433, + 0.046598829329013824, + 0.0329955630004406, + 0.05317215621471405, + 0.02522912807762623, + 0.1142033040523529 + ], + "credit_map": [ + 0.03604289889335632, + 0.007801794912666082, + 0.04257579520344734, + 0.005459214095026255, + 0.0060978839173913, + 0.023217113688588142, + 0.00480249896645546, + 0.04512159153819084, + 0.009996420703828335, + 0.00224936380982399, + 0.009016099385917187, + 0.0025341734290122986, + 0.0020231856033205986, + 0.00669459393247962, + 0.003072358202189207, + 0.008870979771018028, + 0.013832099735736847, + 0.003751972457394004, + 0.009786048904061317, + 0.0047516245394945145, + 0.004280088469386101, + 0.010784715414047241, + 0.004420076031237841, + 0.009954951703548431, + 0.024995800107717514, + 0.006347869522869587, + 0.16858525574207306, + 0.010676207952201366, + 0.007505913730710745, + 0.015495636500418186, + 0.004389096517115831, + 0.18142502009868622, + 0.0037289767060428858, + 0.0024724348913878202, + 0.007399825844913721, + 0.002242732560262084, + 0.0022477812599390745, + 0.004223768133670092, + 0.0014894900377839804, + 0.00743764266371727, + 0.003771804040297866, + 0.002866765484213829, + 0.014350412413477898, + 0.0020795653108507395, + 0.003216435434296727, + 0.0045019108802080154, + 0.0017290010582655668, + 0.014082936570048332, + 0.005522971972823143, + 0.0032217479310929775, + 0.005464761983603239, + 0.003529492998495698, + 0.002387134823948145, + 0.004329077899456024, + 0.003360672853887081, + 0.005179700441658497, + 0.008004382252693176, + 0.004709850996732712, + 0.008004448376595974, + 0.006318769883364439, + 0.0045126802287995815, + 0.007500613573938608, + 0.007165616378188133, + 0.006955794524401426, + 0.002867569215595722, + 0.002376489108428359, + 0.0044336398132145405, + 0.0026554132346063852, + 0.0016912295250222087, + 0.0037872674874961376, + 0.0022917999885976315, + 0.005125720519572496, + 0.008877607993781567, + 0.0023527389857918024, + 0.04427565634250641, + 0.003071356797590852, + 0.0023403922095894814, + 0.005669004749506712, + 0.001624402473680675, + 0.04599214717745781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.3871184289455414 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005961731541901827 + ], + "opposite_logprob_tokens": [ + -5.130961894989014 + ], + "logprob_reward": -0.005961731541901827, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005961731541901827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07175231724977493, + 0.10903722792863846, + 0.17353670299053192, + 0.1377885937690735, + 0.0653524249792099, + 0.03597593680024147, + 0.04650643467903137, + 0.05246647074818611, + 0.03528979793190956, + 0.27229398488998413 + ], + "credit_map": [ + 0.00757733266800642, + 0.005251340568065643, + 0.015227987430989742, + 0.007037211675196886, + 0.005755032412707806, + 0.008266108110547066, + 0.006886092945933342, + 0.01575121097266674, + 0.01509218942373991, + 0.007370668463408947, + 0.024783531203866005, + 0.008346320129930973, + 0.01000223308801651, + 0.009626797400414944, + 0.0064408197067677975, + 0.027374686673283577, + 0.037124019116163254, + 0.009026885032653809, + 0.036051541566848755, + 0.007203663699328899, + 0.008334783837199211, + 0.029239028692245483, + 0.007504026871174574, + 0.03905278071761131, + 0.009375199675559998, + 0.002960213227197528, + 0.050809044390916824, + 0.004276486579328775, + 0.0036241388879716396, + 0.007372416090220213, + 0.003057836787775159, + 0.05631328374147415, + 0.00830569677054882, + 0.0038558284286409616, + 0.016189102083444595, + 0.006915945094078779, + 0.003741301130503416, + 0.008006692864000797, + 0.003940894268453121, + 0.014396977610886097, + 0.006592999678105116, + 0.0024130032397806644, + 0.005765295121818781, + 0.0037199591752141714, + 0.0025556725449860096, + 0.006907658185809851, + 0.002695897128432989, + 0.00532545568421483, + 0.008240773342549801, + 0.004171299282461405, + 0.007980032823979855, + 0.0054099983535707, + 0.0030775126069784164, + 0.006722568534314632, + 0.004678713623434305, + 0.006225539371371269, + 0.007668027188628912, + 0.0049216607585549355, + 0.008247374556958675, + 0.004881600383669138, + 0.004496123641729355, + 0.007979020476341248, + 0.005424847360700369, + 0.008847820572555065, + 0.004673142917454243, + 0.0031856258865445852, + 0.005724221467971802, + 0.004279365763068199, + 0.0024047852493822575, + 0.005094630643725395, + 0.003264511004090309, + 0.006663519889116287, + 0.018290460109710693, + 0.00739195104688406, + 0.1007831022143364, + 0.011114189401268959, + 0.007796266581863165, + 0.012950541451573372, + 0.008983194828033447, + 0.10498432070016861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.005961731541901827 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211602449417114 + ], + "opposite_logprob_tokens": [ + -1.5021159648895264 + ], + "logprob_reward": -0.25211602449417114, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211602449417114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04238298907876015, + 0.1348428875207901, + 0.3175351023674011, + 0.1554238349199295, + 0.09871643036603928, + 0.10777483880519867, + 0.030413048341870308, + 0.01898360438644886, + 0.02338358201086521, + 0.07054371386766434 + ], + "credit_map": [ + 0.006582737434655428, + 0.0034417551942169666, + 0.00625465577468276, + 0.004642847925424576, + 0.004929996561259031, + 0.005432805046439171, + 0.0047025834210217, + 0.0063956077210605145, + 0.031756963580846786, + 0.008268882520496845, + 0.026535870507359505, + 0.010545691475272179, + 0.006976564414799213, + 0.01272814255207777, + 0.008251088671386242, + 0.02977968379855156, + 0.0601518377661705, + 0.019121859222650528, + 0.04642554372549057, + 0.03731301426887512, + 0.022351203486323357, + 0.04975823685526848, + 0.0330643430352211, + 0.04934907704591751, + 0.01318622287362814, + 0.0032172552309930325, + 0.05587944760918617, + 0.004099403973668814, + 0.0034303830470889807, + 0.010985097847878933, + 0.002562749432399869, + 0.06206327676773071, + 0.01086938101798296, + 0.006670681294053793, + 0.02118563838303089, + 0.012532941997051239, + 0.006186578422784805, + 0.014067260548472404, + 0.009553547017276287, + 0.017650403082370758, + 0.009310667403042316, + 0.008694475516676903, + 0.024718645960092545, + 0.0114534180611372, + 0.009560571983456612, + 0.010313047096133232, + 0.009151216596364975, + 0.024572793394327164, + 0.004888225346803665, + 0.0031663489062339067, + 0.00464335922151804, + 0.0027141249738633633, + 0.0023199717979878187, + 0.005179744213819504, + 0.0022642542608082294, + 0.005237017758190632, + 0.0022609964944422245, + 0.0018965716008096933, + 0.0032534017227590084, + 0.0023947760928422213, + 0.0014346912503242493, + 0.0019982117228209972, + 0.002269263379275799, + 0.0034756921231746674, + 0.0026181861758232117, + 0.0021118121221661568, + 0.004223213531076908, + 0.0035137468948960304, + 0.001520573510788381, + 0.0022938987240195274, + 0.0025107432156801224, + 0.0045914072543382645, + 0.005486513953655958, + 0.002570665907114744, + 0.025682926177978516, + 0.00348306423984468, + 0.0029585196170955896, + 0.00394396809861064, + 0.001790244597941637, + 0.0246278103441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.25211602449417114 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 face with heavy clown makeup?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera remain still throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a bright blue wig and a bright red bulbous prosthetic nose visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full crystalline tear well up from the right eye and slip down the cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a cold, focused spotlight coming from above and the side?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129535362124443 + ], + "opposite_logprob_tokens": [ + -2.2379534244537354 + ], + "logprob_reward": -0.1129535362124443, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129535362124443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06252855807542801, + 0.17393411695957184, + 0.10842923820018768, + 0.3325941860675812, + 0.021298222243785858, + 0.046846941113471985, + 0.02076108753681183, + 0.026481978595256805, + 0.016321711242198944, + 0.19080397486686707 + ], + "credit_map": [ + 0.011344127357006073, + 0.003462330438196659, + 0.01430171262472868, + 0.002974507864564657, + 0.003793688490986824, + 0.008555916137993336, + 0.0031290992628782988, + 0.0149671770632267, + 0.03964415192604065, + 0.009288802742958069, + 0.040979813784360886, + 0.009299459867179394, + 0.0073181879706680775, + 0.017561035230755806, + 0.006533714942634106, + 0.04330894351005554, + 0.020868495106697083, + 0.006703778635710478, + 0.016549700871109962, + 0.009956594556570053, + 0.007757071405649185, + 0.019005397334694862, + 0.009565350599586964, + 0.018022853881120682, + 0.019699012860655785, + 0.005038758274167776, + 0.13382397592067719, + 0.008688720874488354, + 0.006135071162134409, + 0.01214857678860426, + 0.0034391721710562706, + 0.14362090826034546, + 0.002262505004182458, + 0.0015004012966528535, + 0.005220396444201469, + 0.002285473048686981, + 0.001185172819532454, + 0.00267379661090672, + 0.0017228354699909687, + 0.004447640851140022, + 0.004081271588802338, + 0.0030362424440681934, + 0.013961954042315483, + 0.0024038469418883324, + 0.0028275020886212587, + 0.005199349485337734, + 0.0020774404983967543, + 0.013259340077638626, + 0.002500775270164013, + 0.001307860016822815, + 0.004502221941947937, + 0.0018408260075375438, + 0.0016615198692306876, + 0.0028809953946620226, + 0.0021153250709176064, + 0.003951563499867916, + 0.003867078572511673, + 0.002237857785075903, + 0.004546565003693104, + 0.002983741695061326, + 0.002098261844366789, + 0.003104340983554721, + 0.0035086108837276697, + 0.004135520197451115, + 0.002414798829704523, + 0.0017095481744036078, + 0.0027992073446512222, + 0.0009459038847126067, + 0.0014821015065535903, + 0.002753940410912037, + 0.0010462451027706265, + 0.0031699654646217823, + 0.012157256714999676, + 0.004902022425085306, + 0.0739346295595169, + 0.005713642109185457, + 0.005379670299589634, + 0.00912362802773714, + 0.003772129537537694, + 0.07582100480794907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.1129535362124443 + } + ], + "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/0000158/group_0/index_tensor([6871, 6871])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_tensor([6871, 6871])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..009845828a793e3d535f1b5cfff40671536261af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_tensor([6871, 6871])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic shot, beginning at eye level and rapidly zooming upwards to a bird's-eye view, perfectly captures a magical spectacle imbued with the mystical power of Eastern fantasy. At the outset, an ancient sword, gleaming with a cold gleam, hovers in the air, its blade engraved with intricate cloud patterns. As the camera zooms out, the sword, in a flash of dazzling light, instantly transforms into countless overlapping afterimages, ultimately solidifying into eight identical swords. With a piercing whistle, they descend from the sky with thunderous force, piercing the ground with unerring precision, surrounding the white-clad female warrior seated cross-legged at the center, forming a perfect Bagua formation. The female warrior is clad in a pure white silk robe, the sleeves embroidered with silver auspicious cloud patterns. Her cascading black hair, partly tied in a delicate bun, flutters gently in the surging currents of magic. Her face is cool and focused, her gaze fixed firmly on the horizon. Upon landing, the eight swords begin to spin synchronously at high speed, their blades slicing through the air, creating distinct vortices of white air. Immediately, dazzling blue lightning and energy streams burst forth from the blades. These energy streams rapidly intertwine and converge in mid-air, ultimately forming a vast, hemispherical, blue, transparent energy barrier. Countless tiny runes ripple across its surface, completely enveloping and protecting the female warrior. The entire scene is set on a mountaintop platform suspended above a sea of clouds, against a backdrop of swirling purple and deep blue clouds. The blue glow of the sword formation provides the sole source of light, casting sharp, dynamic shadows across the ancient stone floor. The overall scene is imbued with the rich textures of a xianxia film, with magnificent and majestic visuals, drawing heavily on the vibrant and imaginative Taoist special effects aesthetics of \"The Legend of Zu,\" creating an atmosphere of sacredness, solemnity, and power.", + "seed": 32862, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_6871_seed_32862_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": -0.0678245487390086, + "vlm_logprob": -0.0678245487390086, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.0678245487390086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.16482648253440857 + ], + "opposite_logprob_tokens": [ + -1.914826512336731 + ], + "logprob_reward": -0.16482648253440857, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16482648253440857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13326841592788696, + 0.10212131589651108, + 0.054435547441244125, + 0.10319757461547852, + 0.03540436923503876, + 0.04230402782559395, + 0.16386781632900238, + 0.10361198335886002, + 0.1451771855354309, + 0.11661174893379211 + ], + "credit_map": [ + 0.010350938886404037, + 0.031661003828048706, + 0.008379848673939705, + 0.006719618104398251, + 0.03639338165521622, + 0.01920252852141857, + 0.008055038750171661, + 0.012506049126386642, + 0.014765028841793537, + 0.010021391324698925, + 0.021113917231559753, + 0.010787855833768845, + 0.00823657400906086, + 0.016124648973345757, + 0.010366302914917469, + 0.01070560235530138, + 0.015058428049087524, + 0.0051383813843131065, + 0.009248566813766956, + 0.0033000758849084377, + 0.005822101142257452, + 0.005695777013897896, + 0.0038346133660525084, + 0.0063376049511134624, + 0.011736059561371803, + 0.013849434442818165, + 0.010347950272262096, + 0.013497963547706604, + 0.012888290919363499, + 0.014631950296461582, + 0.010419846512377262, + 0.015826081857085228, + 0.005217495374381542, + 0.005230145063251257, + 0.0047177537344396114, + 0.005516692530363798, + 0.0036262881476432085, + 0.002607033820822835, + 0.0049614147283136845, + 0.003527543507516384, + 0.005014898255467415, + 0.00452197203412652, + 0.0052141472697257996, + 0.0075788963586091995, + 0.0034044848289340734, + 0.005816272925585508, + 0.006201059091836214, + 0.004552298691123724, + 0.017988337203860283, + 0.01472525391727686, + 0.04147624969482422, + 0.021589161828160286, + 0.012619104236364365, + 0.01235898956656456, + 0.014162079431116581, + 0.028948651626706123, + 0.00848895963281393, + 0.015560699626803398, + 0.012096978724002838, + 0.011806865222752094, + 0.010489807464182377, + 0.010335204191505909, + 0.01659972220659256, + 0.01823374629020691, + 0.013245407491922379, + 0.018662000074982643, + 0.013834497891366482, + 0.018878430128097534, + 0.021199872717261314, + 0.017005939036607742, + 0.02420889399945736, + 0.018142154440283775, + 0.010013648308813572, + 0.00877318624407053, + 0.041104141622781754, + 0.008217129856348038, + 0.010443584993481636, + 0.009154370985925198, + 0.010657677426934242, + 0.018248002976179123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.16482648253440857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.02331160008907318 + ], + "opposite_logprob_tokens": [ + -3.7733116149902344 + ], + "logprob_reward": -0.02331160008907318, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331160008907318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09204793721437454, + 0.04280576482415199, + 0.07476218789815903, + 0.09503335505723953, + 0.11818847060203552, + 0.07547158002853394, + 0.36640799045562744, + 0.04153857007622719, + 0.03564625233411789, + 0.05809788033366203 + ], + "credit_map": [ + 0.007902292534708977, + 0.01651368848979473, + 0.006673275027424097, + 0.006239008624106646, + 0.022840507328510284, + 0.01456160843372345, + 0.006860451772809029, + 0.010457093827426434, + 0.0077433413825929165, + 0.003997556399554014, + 0.005942480638623238, + 0.005076277069747448, + 0.0029335536528378725, + 0.0073807076551020145, + 0.004002977162599564, + 0.0057288664393126965, + 0.0074694217182695866, + 0.006069588474929333, + 0.022733056917786598, + 0.005218425765633583, + 0.005068594589829445, + 0.00585639663040638, + 0.005094143562018871, + 0.01725255325436592, + 0.01415860652923584, + 0.009277189150452614, + 0.012085056863725185, + 0.011744847521185875, + 0.010473918169736862, + 0.013175108470022678, + 0.009118062444031239, + 0.015000561252236366, + 0.012865499593317509, + 0.015780460089445114, + 0.010349285788834095, + 0.017698336392641068, + 0.027968667447566986, + 0.008541707880795002, + 0.014958641491830349, + 0.010025868192315102, + 0.006328086834400892, + 0.00819423422217369, + 0.008617273531854153, + 0.012600325047969818, + 0.012845315970480442, + 0.006613531149923801, + 0.01245166826993227, + 0.007821138016879559, + 0.03759252652525902, + 0.015300111845135689, + 0.1463918387889862, + 0.019420605152845383, + 0.010483242571353912, + 0.018959878012537956, + 0.015030896291136742, + 0.10322887450456619, + 0.0030797638464719057, + 0.004724840167909861, + 0.0038022263906896114, + 0.004118714015930891, + 0.004745739046484232, + 0.003932653460651636, + 0.008867213502526283, + 0.0082674166187644, + 0.003572323126718402, + 0.004091966897249222, + 0.003637610701844096, + 0.004845894873142242, + 0.005267371889203787, + 0.00501192407682538, + 0.004981252830475569, + 0.00423790467903018, + 0.0054025971330702305, + 0.0041729663498699665, + 0.019171489402651787, + 0.004603315144777298, + 0.005590192042291164, + 0.005427167285233736, + 0.005085981450974941, + 0.008644167333841324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.02331160008907318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.10038213431835175 + ], + "opposite_logprob_tokens": [ + -2.350382089614868 + ], + "logprob_reward": -0.10038213431835175, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038213431835175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.058446913957595825, + 0.028648963198065758, + 0.08899784088134766, + 0.026705225929617882, + 0.06041133031249046, + 0.03149455413222313, + 0.45902326703071594, + 0.03952459245920181, + 0.03960738703608513, + 0.1671399474143982 + ], + "credit_map": [ + 0.004302343353629112, + 0.007197690196335316, + 0.0042780316434800625, + 0.005408701952546835, + 0.012405733577907085, + 0.0054671610705554485, + 0.007848750799894333, + 0.011538499966263771, + 0.004348100628703833, + 0.002902315929532051, + 0.0038289085496217012, + 0.004000820219516754, + 0.002447903621941805, + 0.0038394664879888296, + 0.004014375153928995, + 0.0032670716755092144, + 0.009620558470487595, + 0.013891751877963543, + 0.01678163930773735, + 0.006732629146426916, + 0.006378739606589079, + 0.01373292040079832, + 0.00794222392141819, + 0.01391737349331379, + 0.0031932538840919733, + 0.0035611798521131277, + 0.002911066170781851, + 0.003530145389959216, + 0.003034636378288269, + 0.004204759374260902, + 0.0027341407258063555, + 0.003536041360348463, + 0.005843378137797117, + 0.008738378994166851, + 0.0054910327307879925, + 0.008770233951508999, + 0.013469875790178776, + 0.005035913083702326, + 0.007666829973459244, + 0.005395680200308561, + 0.003630707273259759, + 0.002945652464404702, + 0.006343180779367685, + 0.003362608840689063, + 0.004751251079142094, + 0.0035330462269484997, + 0.003786278422921896, + 0.003141826018691063, + 0.0463612861931324, + 0.020722273737192154, + 0.18124975264072418, + 0.02573956549167633, + 0.014077998697757721, + 0.024418892338871956, + 0.020002370700240135, + 0.12645110487937927, + 0.0029985399451106787, + 0.004231269937008619, + 0.005261055193841457, + 0.003887808881700039, + 0.00389930815435946, + 0.003068649908527732, + 0.006648593582212925, + 0.009529363363981247, + 0.0044731274247169495, + 0.004543900024145842, + 0.005805995315313339, + 0.004433224909007549, + 0.004163697827607393, + 0.00618955260142684, + 0.004733830690383911, + 0.005264056380838156, + 0.01735357940196991, + 0.01584370620548725, + 0.03686951473355293, + 0.01562521979212761, + 0.021536607295274734, + 0.021139295771718025, + 0.020239582285284996, + 0.01853242889046669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.10038213431835175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.0019422968616709113 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019422968616709113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019422968616709113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056210801005363464, + 0.03581446409225464, + 0.059243958443403244, + 0.030784789472818375, + 0.05990312248468399, + 0.03231511637568474, + 0.46996450424194336, + 0.08452246338129044, + 0.05470096319913864, + 0.1165398359298706 + ], + "credit_map": [ + 0.006115713622421026, + 0.007348919752985239, + 0.006345504429191351, + 0.0054658860899508, + 0.008227696642279625, + 0.008222763426601887, + 0.007159042172133923, + 0.007325275801122189, + 0.006480959709733725, + 0.003105615498498082, + 0.005148856434971094, + 0.0032268960494548082, + 0.0025623012334108353, + 0.006756293121725321, + 0.0038241895381361246, + 0.0047093527391552925, + 0.008226516656577587, + 0.005625697318464518, + 0.012337582185864449, + 0.004897750448435545, + 0.0033954919781535864, + 0.009348336607217789, + 0.005301396828144789, + 0.010111186653375626, + 0.004504707641899586, + 0.0035236256662756205, + 0.0030918512493371964, + 0.003970361780375242, + 0.0028543805237859488, + 0.00461804261431098, + 0.004313797224313021, + 0.003908020909875631, + 0.01130230538547039, + 0.007290687877684832, + 0.007635368965566158, + 0.006479522213339806, + 0.008860289119184017, + 0.006275213323533535, + 0.007165821734815836, + 0.004893917124718428, + 0.00245809112675488, + 0.003508772235363722, + 0.004780617076903582, + 0.003761870553717017, + 0.003996783867478371, + 0.0037803722079843283, + 0.005725635681301355, + 0.004302973859012127, + 0.046556923538446426, + 0.022744599729776382, + 0.18218515813350677, + 0.028453698381781578, + 0.016161378473043442, + 0.02468905784189701, + 0.023603063076734543, + 0.12557058036327362, + 0.0068741352297365665, + 0.009936204180121422, + 0.009887375868856907, + 0.0069520422257483006, + 0.006952380295842886, + 0.008732295595109463, + 0.012125375680625439, + 0.023062651976943016, + 0.004422801546752453, + 0.0066491891629993916, + 0.0065474738366901875, + 0.005336443893611431, + 0.007438525557518005, + 0.009533591568470001, + 0.006577178370207548, + 0.008195758797228336, + 0.010111033916473389, + 0.009602684527635574, + 0.032407380640506744, + 0.009669320657849312, + 0.01560115721076727, + 0.012688534334301949, + 0.010740654543042183, + 0.015719063580036163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0019422968616709113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.04866022989153862 + ], + "opposite_logprob_tokens": [ + -3.0486602783203125 + ], + "logprob_reward": -0.04866022989153862, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866022989153862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22324074804782867, + 0.040734581649303436, + 0.09253624081611633, + 0.06903797388076782, + 0.09190395474433899, + 0.05416560545563698, + 0.2895597219467163, + 0.042613089084625244, + 0.05846670642495155, + 0.03774140030145645 + ], + "credit_map": [ + 0.01801987923681736, + 0.040162861347198486, + 0.013230228796601295, + 0.014936924912035465, + 0.06349389255046844, + 0.029820164665579796, + 0.01588749885559082, + 0.027689291164278984, + 0.007585939485579729, + 0.00445193238556385, + 0.005256760399788618, + 0.005309101194143295, + 0.003133480902761221, + 0.005830713082104921, + 0.0041750287637114525, + 0.004991624504327774, + 0.011412585154175758, + 0.014300445094704628, + 0.023988932371139526, + 0.005657684523612261, + 0.004899640567600727, + 0.0079036308452487, + 0.005315722431987524, + 0.01905759982764721, + 0.00805661641061306, + 0.009419652633368969, + 0.00708541227504611, + 0.008110500872135162, + 0.006578461267054081, + 0.012143032625317574, + 0.008196731097996235, + 0.009447558782994747, + 0.009729068726301193, + 0.015597451478242874, + 0.009696602821350098, + 0.012131993658840656, + 0.015197823755443096, + 0.009417099878191948, + 0.00903069507330656, + 0.011103215627372265, + 0.007020233664661646, + 0.003887827740982175, + 0.012905681505799294, + 0.006345096509903669, + 0.00567625230178237, + 0.004521702416241169, + 0.007817072793841362, + 0.005991737823933363, + 0.029240647330880165, + 0.013051540590822697, + 0.11334457993507385, + 0.016591861844062805, + 0.009762348607182503, + 0.01564715802669525, + 0.013505695387721062, + 0.0784158855676651, + 0.004277586005628109, + 0.0047235568054020405, + 0.007195339538156986, + 0.004897098522633314, + 0.004017633385956287, + 0.004509745165705681, + 0.0072840009815990925, + 0.0057081300765275955, + 0.006867971271276474, + 0.00483692204579711, + 0.01001537125557661, + 0.0069962055422365665, + 0.005970930680632591, + 0.008603321388363838, + 0.008009820245206356, + 0.007166162133216858, + 0.003929418046027422, + 0.0029041750822216272, + 0.011075242422521114, + 0.0028429138474166393, + 0.0034177424386143684, + 0.004070009104907513, + 0.004101337864995003, + 0.005400564055889845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.04866022989153862 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, beginning at eye level and rapidly zooming upwards to a bird's-eye view, perfectly captures a magical spectacle imbued with the mystical power of Eastern fantasy. At the outset, an ancient sword, gleaming with a cold gleam, hovers in the air, its blade engraved with intricate cloud patterns. As the camera zooms out, the sword, in a flash of dazzling light, instantly transforms into countless overlapping afterimages, ultimately solidifying into eight identical swords. With a piercing whistle, they descend from the sky with thunderous force, piercing the ground with unerring precision, surrounding the white-clad female warrior seated cross-legged at the center, forming a perfect Bagua formation. The female warrior is clad in a pure white silk robe, the sleeves embroidered with silver auspicious cloud patterns. Her cascading black hair, partly tied in a delicate bun, flutters gently in the surging currents of magic. Her face is cool and focused, her gaze fixed firmly on the horizon. Upon landing, the eight swords begin to spin synchronously at high speed, their blades slicing through the air, creating distinct vortices of white air. Immediately, dazzling blue lightning and energy streams burst forth from the blades. These energy streams rapidly intertwine and converge in mid-air, ultimately forming a vast, hemispherical, blue, transparent energy barrier. Countless tiny runes ripple across its surface, completely enveloping and protecting the female warrior. The entire scene is set on a mountaintop platform suspended above a sea of clouds, against a backdrop of swirling purple and deep blue clouds. The blue glow of the sword formation provides the sole source of light, casting sharp, dynamic shadows across the ancient stone floor. The overall scene is imbued with the rich textures of a xianxia film, with magnificent and majestic visuals, drawing heavily on the vibrant and imaginative Taoist special effects aesthetics of \"The Legend of Zu,\" creating an atmosphere of sacredness, solemnity, and power.", + "seed": 32863, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_6871_seed_32863_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": -0.04193173674866557, + "vlm_logprob": -0.04193173674866557, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.04193173674866557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.09096360951662064 + ], + "opposite_logprob_tokens": [ + -2.46596360206604 + ], + "logprob_reward": -0.09096360951662064, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09096360951662064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1396884173154831, + 0.1067899614572525, + 0.0626712366938591, + 0.11856190860271454, + 0.07834344357252121, + 0.12766240537166595, + 0.01310783438384533, + 0.10381808131933212, + 0.04848673194646835, + 0.20086994767189026 + ], + "credit_map": [ + 0.014156291261315346, + 0.021794887259602547, + 0.022785790264606476, + 0.012378900311887264, + 0.019148949533700943, + 0.019998345524072647, + 0.013905918225646019, + 0.015519333072006702, + 0.01340976171195507, + 0.008350533433258533, + 0.014012180268764496, + 0.007855992764234543, + 0.006668038200587034, + 0.020839160308241844, + 0.0049476632848382, + 0.03070663847029209, + 0.006806524470448494, + 0.005622204393148422, + 0.01097656786441803, + 0.005555647425353527, + 0.005549434572458267, + 0.015315811149775982, + 0.006087210029363632, + 0.006757833529263735, + 0.00757155055180192, + 0.008354701101779938, + 0.03627283498644829, + 0.011557579971849918, + 0.009279875084757805, + 0.008327525109052658, + 0.008004602044820786, + 0.029193243011832237, + 0.006337581668049097, + 0.015903521329164505, + 0.011337000876665115, + 0.011738493107259274, + 0.007577663287520409, + 0.008567625656723976, + 0.007861539721488953, + 0.009020011872053146, + 0.010611177422106266, + 0.014675924554467201, + 0.009754611179232597, + 0.01619807258248329, + 0.023573284968733788, + 0.018281053751707077, + 0.01601356267929077, + 0.018554717302322388, + 0.001546932035125792, + 0.0012239650823175907, + 0.0018274375470355153, + 0.001873369561508298, + 0.0015318053774535656, + 0.0017737976741045713, + 0.0017196530243381858, + 0.0016108739655464888, + 0.005410363897681236, + 0.0064065465703606606, + 0.010193903930485249, + 0.025238238275051117, + 0.007813415490090847, + 0.010499020107090473, + 0.027598168700933456, + 0.010658429004251957, + 0.006352740805596113, + 0.0045217303559184074, + 0.006156110670417547, + 0.006734265945851803, + 0.004527300596237183, + 0.006917829159647226, + 0.006965016946196556, + 0.0063117328099906445, + 0.014056019484996796, + 0.01251248363405466, + 0.04156949743628502, + 0.023060645908117294, + 0.018057677894830704, + 0.015595102682709694, + 0.04155007749795914, + 0.03446844220161438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.09096360951662064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.004107135813683271 + ], + "opposite_logprob_tokens": [ + -5.5041069984436035 + ], + "logprob_reward": -0.004107135813683271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107135813683271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1440448760986328, + 0.17903178930282593, + 0.11045871675014496, + 0.2024378776550293, + 0.19807441532611847, + 0.0511951670050621, + 0.024499865248799324, + 0.0398084856569767, + 0.019895588979125023, + 0.03055330365896225 + ], + "credit_map": [ + 0.014139685779809952, + 0.032420046627521515, + 0.01517639309167862, + 0.013452506624162197, + 0.012448437511920929, + 0.030972866341471672, + 0.011313538067042828, + 0.01412138156592846, + 0.01745089516043663, + 0.011333168484270573, + 0.028779154643416405, + 0.011491497047245502, + 0.011335683986544609, + 0.03356524556875229, + 0.009237238205969334, + 0.055838894098997116, + 0.008798166178166866, + 0.010766345076262951, + 0.017208926379680634, + 0.01123785600066185, + 0.011770202778279781, + 0.02232888527214527, + 0.011749360710382462, + 0.01659896783530712, + 0.014341835863888264, + 0.016362613067030907, + 0.06298786401748657, + 0.019240640103816986, + 0.015983425080776215, + 0.015234661288559437, + 0.013755598105490208, + 0.044531237334012985, + 0.023484230041503906, + 0.05141745135188103, + 0.020803112536668777, + 0.02798299491405487, + 0.023304050788283348, + 0.01360043603926897, + 0.015399744734168053, + 0.022082384675741196, + 0.0039794836193323135, + 0.01034967228770256, + 0.004061249550431967, + 0.006326661910861731, + 0.00931612029671669, + 0.007103465963155031, + 0.0058008888736367226, + 0.00425761891528964, + 0.0026840113569051027, + 0.003310754895210266, + 0.0029308015946298838, + 0.0034058985766023397, + 0.002820451743900776, + 0.0028683992568403482, + 0.003679102985188365, + 0.0028004434425383806, + 0.0029325180221349, + 0.0032988805323839188, + 0.004992532543838024, + 0.008411976508796215, + 0.0032174596562981606, + 0.003343984019011259, + 0.010783204808831215, + 0.0028279258403927088, + 0.0014446605928242207, + 0.0010772926034405828, + 0.0026589203625917435, + 0.004566725809127092, + 0.0015073486138135195, + 0.0017995131202042103, + 0.004431732464581728, + 0.0024093936663120985, + 0.0026799452025443316, + 0.002657353412359953, + 0.003953608684241772, + 0.00396333122625947, + 0.003131575183942914, + 0.003135506296530366, + 0.007745612878352404, + 0.003286367980763316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.004107135813683271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.1003418117761612 + ], + "opposite_logprob_tokens": [ + -2.350341796875 + ], + "logprob_reward": -0.1003418117761612, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003418117761612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18616048991680145, + 0.05115481838583946, + 0.0737486183643341, + 0.13795524835586548, + 0.05505078658461571, + 0.11798197031021118, + 0.021402869373559952, + 0.10809667408466339, + 0.09254631400108337, + 0.1559022068977356 + ], + "credit_map": [ + 0.019339069724082947, + 0.041755419224500656, + 0.02112758532166481, + 0.017050279304385185, + 0.016143690794706345, + 0.03831832855939865, + 0.013152260333299637, + 0.019273880869150162, + 0.005599295720458031, + 0.006182351149618626, + 0.0074624172411859035, + 0.003566259052604437, + 0.004278083797544241, + 0.008454259485006332, + 0.00386754865758121, + 0.011744610033929348, + 0.005951274186372757, + 0.006117774173617363, + 0.012297115288674831, + 0.00856330618262291, + 0.004848826676607132, + 0.016270287334918976, + 0.0070543717592954636, + 0.01264567207545042, + 0.010501077398657799, + 0.0089629627764225, + 0.03748461976647377, + 0.01778998225927353, + 0.010419357568025589, + 0.011347189545631409, + 0.010721213184297085, + 0.03072885423898697, + 0.005460740532726049, + 0.010280558839440346, + 0.006842457223683596, + 0.009764716029167175, + 0.004647029098123312, + 0.005839723628014326, + 0.006611264776438475, + 0.005604300647974014, + 0.00905379094183445, + 0.021358903497457504, + 0.008979667909443378, + 0.009870927780866623, + 0.015576498582959175, + 0.02136024832725525, + 0.00936166476458311, + 0.02242029458284378, + 0.0028012790717184544, + 0.0020727799274027348, + 0.0023053698241710663, + 0.00418970687314868, + 0.0018945303745567799, + 0.0028479392640292645, + 0.0031174211762845516, + 0.0021738449577242136, + 0.010530694387853146, + 0.02529957890510559, + 0.016875820234417915, + 0.012224960140883923, + 0.011634374037384987, + 0.009827151894569397, + 0.011773060075938702, + 0.009931052103638649, + 0.01088254339993, + 0.007741919253021479, + 0.011964462697505951, + 0.01601436361670494, + 0.01011325791478157, + 0.010605642572045326, + 0.013294843025505543, + 0.011929291300475597, + 0.012248240411281586, + 0.005862080492079258, + 0.05733625218272209, + 0.014513382688164711, + 0.010182604193687439, + 0.007865479215979576, + 0.010011606849730015, + 0.03788256645202637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.1003418117761612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.001713952049612999 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.001713952049612999, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001713952049612999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11523335427045822, + 0.10914788395166397, + 0.05973338335752487, + 0.05242722108960152, + 0.3305862545967102, + 0.04561920836567879, + 0.03870329633355141, + 0.0942368134856224, + 0.042224422097206116, + 0.11208818107843399 + ], + "credit_map": [ + 0.011007392778992653, + 0.015992889180779457, + 0.021046187728643417, + 0.01625487394630909, + 0.015776030719280243, + 0.013366039842367172, + 0.010258409194648266, + 0.011531529948115349, + 0.013993979431688786, + 0.007628221530467272, + 0.016944369301199913, + 0.006938671227544546, + 0.007399212568998337, + 0.020185034722089767, + 0.004637301899492741, + 0.03142109513282776, + 0.005214950069785118, + 0.005133130121976137, + 0.009176624938845634, + 0.007371369283646345, + 0.00688644265756011, + 0.010222241282463074, + 0.006821006070822477, + 0.008907618932425976, + 0.004239815287292004, + 0.003375062020495534, + 0.013222969137132168, + 0.006889591459184885, + 0.004873727913945913, + 0.005209425464272499, + 0.004231445956975222, + 0.010385186411440372, + 0.03992322087287903, + 0.0865933895111084, + 0.03359774500131607, + 0.04658278077840805, + 0.03630299121141434, + 0.024472780525684357, + 0.026176370680332184, + 0.03693694993853569, + 0.0036171560641378164, + 0.009655822068452835, + 0.0043998113833367825, + 0.005464338697493076, + 0.007476667873561382, + 0.004517008550465107, + 0.005412422586232424, + 0.0050759813748300076, + 0.0043698810040950775, + 0.006660377606749535, + 0.0045955548994243145, + 0.005117855034768581, + 0.006038931664079428, + 0.005050548817962408, + 0.00416257418692112, + 0.0027075731195509434, + 0.007638014853000641, + 0.011401143856346607, + 0.011758877895772457, + 0.019331950694322586, + 0.00805739313364029, + 0.008262659423053265, + 0.018859539180994034, + 0.008927236311137676, + 0.005309305153787136, + 0.003112336155027151, + 0.007591530214995146, + 0.006358711514621973, + 0.00317931454628706, + 0.004521315451711416, + 0.005924471188336611, + 0.0062274388037621975, + 0.020001724362373352, + 0.010728842578828335, + 0.016577184200286865, + 0.010706694796681404, + 0.010424160398542881, + 0.01602342538535595, + 0.016249539330601692, + 0.011376605369150639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.001713952049612999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.012532174587249756 + ], + "opposite_logprob_tokens": [ + -4.3875322341918945 + ], + "logprob_reward": -0.012532174587249756, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532174587249756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1840055137872696, + 0.09081456810235977, + 0.03273164853453636, + 0.05039166659116745, + 0.5332473516464233, + 0.02308608591556549, + 0.010052335448563099, + 0.023851018399000168, + 0.013094485737383366, + 0.03872530162334442 + ], + "credit_map": [ + 0.01926836185157299, + 0.04380205646157265, + 0.018589993938803673, + 0.017165757715702057, + 0.01391387078911066, + 0.041161924600601196, + 0.012517523020505905, + 0.017586030066013336, + 0.007350720465183258, + 0.007791428826749325, + 0.014271099120378494, + 0.006562520284205675, + 0.006962442304939032, + 0.013843715190887451, + 0.010540403425693512, + 0.023492252454161644, + 0.004355772398412228, + 0.003076893510296941, + 0.004961174912750721, + 0.003876463742926717, + 0.005308723542839289, + 0.0037823636084795, + 0.0030262451618909836, + 0.0043440149165689945, + 0.0037279233802109957, + 0.003099277149885893, + 0.014428798109292984, + 0.006127529311925173, + 0.0022932637948542833, + 0.0038262687157839537, + 0.0038387684617191553, + 0.013049843721091747, + 0.06654518842697144, + 0.1432991921901703, + 0.052726149559020996, + 0.07422894239425659, + 0.06080850586295128, + 0.03543851524591446, + 0.03940083459019661, + 0.06080006808042526, + 0.0018931942759081721, + 0.004535868298262358, + 0.002251406665891409, + 0.0028870217502117157, + 0.0036987997591495514, + 0.0028057994786649942, + 0.002906082896515727, + 0.0021079150028526783, + 0.0008092200732789934, + 0.0011102224234491587, + 0.0008586717303842306, + 0.0020713969133794308, + 0.0014418489299714565, + 0.0010883157374337316, + 0.0017786887474358082, + 0.000893972406629473, + 0.002573663368821144, + 0.002194260014221072, + 0.003484349464997649, + 0.002970978384837508, + 0.0024845879524946213, + 0.003326831851154566, + 0.0038830763660371304, + 0.002933274954557419, + 0.0016303389566019177, + 0.0010061404900625348, + 0.0028678933158516884, + 0.0019413414411246777, + 0.0009162689093500376, + 0.0014337986940518022, + 0.001511618378572166, + 0.0017870869487524033, + 0.0026736119762063026, + 0.001490618335083127, + 0.013594744727015495, + 0.0035292506217956543, + 0.0025536019820719957, + 0.0021449944470077753, + 0.003168370807543397, + 0.009570114314556122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.012532174587249756 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, beginning at eye level and rapidly zooming upwards to a bird's-eye view, perfectly captures a magical spectacle imbued with the mystical power of Eastern fantasy. At the outset, an ancient sword, gleaming with a cold gleam, hovers in the air, its blade engraved with intricate cloud patterns. As the camera zooms out, the sword, in a flash of dazzling light, instantly transforms into countless overlapping afterimages, ultimately solidifying into eight identical swords. With a piercing whistle, they descend from the sky with thunderous force, piercing the ground with unerring precision, surrounding the white-clad female warrior seated cross-legged at the center, forming a perfect Bagua formation. The female warrior is clad in a pure white silk robe, the sleeves embroidered with silver auspicious cloud patterns. Her cascading black hair, partly tied in a delicate bun, flutters gently in the surging currents of magic. Her face is cool and focused, her gaze fixed firmly on the horizon. Upon landing, the eight swords begin to spin synchronously at high speed, their blades slicing through the air, creating distinct vortices of white air. Immediately, dazzling blue lightning and energy streams burst forth from the blades. These energy streams rapidly intertwine and converge in mid-air, ultimately forming a vast, hemispherical, blue, transparent energy barrier. Countless tiny runes ripple across its surface, completely enveloping and protecting the female warrior. The entire scene is set on a mountaintop platform suspended above a sea of clouds, against a backdrop of swirling purple and deep blue clouds. The blue glow of the sword formation provides the sole source of light, casting sharp, dynamic shadows across the ancient stone floor. The overall scene is imbued with the rich textures of a xianxia film, with magnificent and majestic visuals, drawing heavily on the vibrant and imaginative Taoist special effects aesthetics of \"The Legend of Zu,\" creating an atmosphere of sacredness, solemnity, and power.", + "seed": 32864, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_6871_seed_32864_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": -0.05727672905195504, + "vlm_logprob": -0.05727672905195504, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05727672905195504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.20460286736488342 + ], + "opposite_logprob_tokens": [ + -1.704602837562561 + ], + "logprob_reward": -0.20460286736488342, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20460286736488342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2093229740858078, + 0.04614069685339928, + 0.026881730183959007, + 0.03166203200817108, + 0.03946138545870781, + 0.04587098956108093, + 0.017844384536147118, + 0.021395953372120857, + 0.537467896938324, + 0.02395196072757244 + ], + "credit_map": [ + 0.02227136120200157, + 0.032399874180555344, + 0.02214232087135315, + 0.03149116411805153, + 0.02439591847360134, + 0.014611857943236828, + 0.030580805614590645, + 0.03142966702580452, + 0.006065714173018932, + 0.0030188406817615032, + 0.014512741938233376, + 0.0030226220842450857, + 0.006092479452490807, + 0.003611590014770627, + 0.0034900337923318148, + 0.00632667588070035, + 0.00391748221591115, + 0.0025118563789874315, + 0.003316749818623066, + 0.0021476908586919308, + 0.005363707896322012, + 0.003601575968787074, + 0.002415246330201626, + 0.00360742025077343, + 0.004181183874607086, + 0.003228483023121953, + 0.003951020538806915, + 0.003352536354213953, + 0.004213033709675074, + 0.004330987576395273, + 0.005057476926594973, + 0.0033473107032477856, + 0.004770262632519007, + 0.0031128202099353075, + 0.004789154045283794, + 0.0031507357489317656, + 0.005856284871697426, + 0.006149051710963249, + 0.0063453852199018, + 0.005287692882120609, + 0.002872321056202054, + 0.0034481538459658623, + 0.011714477092027664, + 0.0034634910989552736, + 0.003831553505733609, + 0.011422660201787949, + 0.0038866656832396984, + 0.0052316635847091675, + 0.002346644178032875, + 0.0023070035967975855, + 0.0025404959451407194, + 0.0022826658096164465, + 0.002395331161096692, + 0.001798013225197792, + 0.0024586431682109833, + 0.0017155888490378857, + 0.002608915325254202, + 0.0025411357637494802, + 0.0024955926928669214, + 0.002220463939011097, + 0.0030564353801310062, + 0.0029118796810507774, + 0.002552754944190383, + 0.003008778439834714, + 0.038100212812423706, + 0.018153291195631027, + 0.18931949138641357, + 0.021272262558341026, + 0.02574928291141987, + 0.04134775698184967, + 0.02141774632036686, + 0.1821078658103943, + 0.0024476360995322466, + 0.003727065632119775, + 0.00248189689591527, + 0.0034128576517105103, + 0.0037555699236691, + 0.0022697453387081623, + 0.0033775041811168194, + 0.002479685004800558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.20460286736488342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.011088181287050247 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088181287050247, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088181287050247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27677714824676514, + 0.04830880090594292, + 0.033754345029592514, + 0.021903809159994125, + 0.024934036657214165, + 0.3859598934650421, + 0.017862459644675255, + 0.02253151871263981, + 0.13218483328819275, + 0.03578322380781174 + ], + "credit_map": [ + 0.030041353777050972, + 0.036686573177576065, + 0.030646804720163345, + 0.03813699260354042, + 0.03129465505480766, + 0.01814880408346653, + 0.04040705785155296, + 0.05141489580273628, + 0.0036802429240196943, + 0.003065620781853795, + 0.013020021840929985, + 0.00308472802862525, + 0.0037982682697474957, + 0.003038577502593398, + 0.003081818576902151, + 0.015539519488811493, + 0.005811258684843779, + 0.0031659433152526617, + 0.005527331959456205, + 0.0031643742695450783, + 0.004157013259828091, + 0.004136571194976568, + 0.0034570246934890747, + 0.004334825091063976, + 0.003638709895312786, + 0.002062414772808552, + 0.003358441637828946, + 0.0015901303850114346, + 0.0030497522093355656, + 0.0037897753063589334, + 0.002166738733649254, + 0.0022478452883660793, + 0.0028455276042222977, + 0.0025553947780281305, + 0.0029722764156758785, + 0.001998563762754202, + 0.0020076429937034845, + 0.004253049846738577, + 0.003912178333848715, + 0.004389401990920305, + 0.02151387743651867, + 0.027784179896116257, + 0.08951955288648605, + 0.03561943396925926, + 0.042181629687547684, + 0.09372925013303757, + 0.03392244130373001, + 0.04168952256441116, + 0.0019518641056492925, + 0.0017533490899950266, + 0.002599907573312521, + 0.0026946854777634144, + 0.0027071242220699787, + 0.0018191032577306032, + 0.002717292867600918, + 0.001619134098291397, + 0.0028095075394958258, + 0.003079045098274946, + 0.0023106266744434834, + 0.0025683604180812836, + 0.00449772085994482, + 0.0026971502229571342, + 0.002502061892300844, + 0.00206704530864954, + 0.009698286652565002, + 0.004733577836304903, + 0.04763927683234215, + 0.005680575501173735, + 0.006055034231394529, + 0.009593810886144638, + 0.005376072600483894, + 0.04340819641947746, + 0.00443459115922451, + 0.00552968168631196, + 0.0039382497780025005, + 0.0031052560079842806, + 0.007007240783423185, + 0.004458194132894278, + 0.004609159659594297, + 0.002700851298868656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.011088181287050247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.05498351901769638 + ], + "opposite_logprob_tokens": [ + -2.929983615875244 + ], + "logprob_reward": -0.05498351901769638, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498351901769638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.241579070687294, + 0.15719711780548096, + 0.03893108293414116, + 0.024436844512820244, + 0.019737787544727325, + 0.10141824930906296, + 0.021825090050697327, + 0.014087912626564503, + 0.1658075749874115, + 0.21497932076454163 + ], + "credit_map": [ + 0.02554236352443695, + 0.031181033700704575, + 0.029600514099001884, + 0.03239603340625763, + 0.028452275320887566, + 0.01641027070581913, + 0.032375458627939224, + 0.04562108963727951, + 0.010097113437950611, + 0.0077536627650260925, + 0.053759150207042694, + 0.008512381464242935, + 0.009616206400096416, + 0.011678516864776611, + 0.0073150829412043095, + 0.04846498742699623, + 0.008113093674182892, + 0.0032895326148718596, + 0.0054887873120605946, + 0.0026953488122671843, + 0.005076663568615913, + 0.005673600826412439, + 0.00417516753077507, + 0.004418881610035896, + 0.0033039816189557314, + 0.0023315572179853916, + 0.003397056832909584, + 0.0024736763443797827, + 0.0036096214316785336, + 0.0028189218137413263, + 0.004179868381470442, + 0.002322158543393016, + 0.0025964241940528154, + 0.001619590213522315, + 0.0024992628023028374, + 0.0019743391312658787, + 0.0023383803199976683, + 0.0030250113923102617, + 0.0029622705187648535, + 0.0027225064113736153, + 0.006792527157813311, + 0.0060091339983046055, + 0.02476351335644722, + 0.007112474646419287, + 0.00733488705009222, + 0.027986135333776474, + 0.008301623165607452, + 0.013117941096425056, + 0.0017057483782991767, + 0.0030445412266999483, + 0.0034980999771505594, + 0.0027913383673876524, + 0.0031259548850357533, + 0.0018975454149767756, + 0.003652068553492427, + 0.0021097895223647356, + 0.0011094477958977222, + 0.0019964633975178003, + 0.0017330963164567947, + 0.0014127996983006597, + 0.0025834660045802593, + 0.0015324531123042107, + 0.00224953331053257, + 0.001470650895498693, + 0.011962756514549255, + 0.006448735482990742, + 0.059329815208911896, + 0.0066872648894786835, + 0.0076824696734547615, + 0.01208331249654293, + 0.0072689601220190525, + 0.05434422940015793, + 0.027717677876353264, + 0.03202955052256584, + 0.02455393224954605, + 0.020390890538692474, + 0.04467950388789177, + 0.021744467318058014, + 0.028072720393538475, + 0.015790555626153946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.05498351901769638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.0015147175872698426 + ], + "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": [ + 0.09132111072540283, + 0.06263735145330429, + 0.04293080419301987, + 0.01805081032216549, + 0.018018392845988274, + 0.04841164872050285, + 0.017881963402032852, + 0.019596856087446213, + 0.6377672553062439, + 0.04338382929563522 + ], + "credit_map": [ + 0.01185771357268095, + 0.011146200820803642, + 0.010120760649442673, + 0.013296592980623245, + 0.011766714043915272, + 0.007692525628954172, + 0.01200564019382, + 0.01343496609479189, + 0.010202765464782715, + 0.0043252939358353615, + 0.022196605801582336, + 0.0042823320254683495, + 0.004761591553688049, + 0.003459647763520479, + 0.004025605972856283, + 0.009383506141602993, + 0.010257000103592873, + 0.003751585492864251, + 0.007411453872919083, + 0.002364081796258688, + 0.0032793390564620495, + 0.007333549205213785, + 0.0032142659183591604, + 0.005319532006978989, + 0.002774312859401107, + 0.0019347714260220528, + 0.0019956619944423437, + 0.0022504222579300404, + 0.0025571691803634167, + 0.0024661312345415354, + 0.00212083850055933, + 0.001951503800228238, + 0.002269026357680559, + 0.0018111815443262458, + 0.002065432257950306, + 0.0015680748037993908, + 0.002354980446398258, + 0.0027310356963425875, + 0.002770854625850916, + 0.002447806764394045, + 0.002913873875513673, + 0.003866871353238821, + 0.009943859651684761, + 0.005186343099921942, + 0.006333234719932079, + 0.009557453915476799, + 0.005362571682780981, + 0.005247438792139292, + 0.002092883922159672, + 0.0026263142935931683, + 0.002580632921308279, + 0.0021839123219251633, + 0.0022927094250917435, + 0.0014456952922046185, + 0.002810906386002898, + 0.0018489077920094132, + 0.0033592269755899906, + 0.0021271323785185814, + 0.0027960927691310644, + 0.0026391586288809776, + 0.0022855871357023716, + 0.0023628161288797855, + 0.0018979499582201242, + 0.0021288918796926737, + 0.04569264501333237, + 0.022163867950439453, + 0.22620807588100433, + 0.02515530027449131, + 0.029633672907948494, + 0.048549503087997437, + 0.025280218571424484, + 0.21508394181728363, + 0.007552775088697672, + 0.004471587482839823, + 0.004321279004216194, + 0.004199306480586529, + 0.006062361877411604, + 0.007672980893403292, + 0.004642592277377844, + 0.004460945725440979 + ], + "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\": \"Does the camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.014194360002875328 + ], + "opposite_logprob_tokens": [ + -4.264194488525391 + ], + "logprob_reward": -0.014194360002875328, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194360002875328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23246178030967712, + 0.2126506119966507, + 0.07715770602226257, + 0.0580558255314827, + 0.029517967253923416, + 0.05243988335132599, + 0.016440363600850105, + 0.01712905615568161, + 0.2718813717365265, + 0.032265398651361465 + ], + "credit_map": [ + 0.028047800064086914, + 0.03220510482788086, + 0.025453269481658936, + 0.03015146404504776, + 0.02822151593863964, + 0.014819013886153698, + 0.03366437554359436, + 0.039899248629808426, + 0.012681224383413792, + 0.013124017976224422, + 0.058803390711545944, + 0.01634249836206436, + 0.014447971247136593, + 0.020661992952227592, + 0.01190908532589674, + 0.06468043476343155, + 0.01018879096955061, + 0.006642752792686224, + 0.01327190175652504, + 0.008286822587251663, + 0.007671443745493889, + 0.010949775576591492, + 0.007054623682051897, + 0.013091595843434334, + 0.008350204676389694, + 0.00533511396497488, + 0.007870465517044067, + 0.00575177650898695, + 0.007285361643880606, + 0.009342283941805363, + 0.008621391840279102, + 0.005499233026057482, + 0.0048040966503322124, + 0.003333144821226597, + 0.003796687349677086, + 0.0027201776392757893, + 0.003223737934604287, + 0.004076824057847261, + 0.0037679916713386774, + 0.0037953087594360113, + 0.003495100187137723, + 0.003482808591797948, + 0.0123820835724473, + 0.004023573826998472, + 0.004557142499834299, + 0.013325901702046394, + 0.004083612002432346, + 0.007089665159583092, + 0.0014151361538097262, + 0.0019231641199439764, + 0.002087788889184594, + 0.002306127455085516, + 0.002506552031263709, + 0.0019718429539352655, + 0.002447921084240079, + 0.00178183161187917, + 0.0022077029570937157, + 0.0017195834079757333, + 0.0033196292351931334, + 0.001795964315533638, + 0.002137757372111082, + 0.0017878584330901504, + 0.0016605257987976074, + 0.0025000355672091246, + 0.019505517557263374, + 0.009046537801623344, + 0.09628937393426895, + 0.010625515133142471, + 0.013284419663250446, + 0.020626980811357498, + 0.010796765796840191, + 0.091706283390522, + 0.00239733443595469, + 0.004736271221190691, + 0.0031510749831795692, + 0.0031623526010662317, + 0.006996413227170706, + 0.003670932026579976, + 0.004610370844602585, + 0.0035406507086008787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.014194360002875328 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot, beginning at eye level and rapidly zooming upwards to a bird's-eye view, perfectly captures a magical spectacle imbued with the mystical power of Eastern fantasy. At the outset, an ancient sword, gleaming with a cold gleam, hovers in the air, its blade engraved with intricate cloud patterns. As the camera zooms out, the sword, in a flash of dazzling light, instantly transforms into countless overlapping afterimages, ultimately solidifying into eight identical swords. With a piercing whistle, they descend from the sky with thunderous force, piercing the ground with unerring precision, surrounding the white-clad female warrior seated cross-legged at the center, forming a perfect Bagua formation. The female warrior is clad in a pure white silk robe, the sleeves embroidered with silver auspicious cloud patterns. Her cascading black hair, partly tied in a delicate bun, flutters gently in the surging currents of magic. Her face is cool and focused, her gaze fixed firmly on the horizon. Upon landing, the eight swords begin to spin synchronously at high speed, their blades slicing through the air, creating distinct vortices of white air. Immediately, dazzling blue lightning and energy streams burst forth from the blades. These energy streams rapidly intertwine and converge in mid-air, ultimately forming a vast, hemispherical, blue, transparent energy barrier. Countless tiny runes ripple across its surface, completely enveloping and protecting the female warrior. The entire scene is set on a mountaintop platform suspended above a sea of clouds, against a backdrop of swirling purple and deep blue clouds. The blue glow of the sword formation provides the sole source of light, casting sharp, dynamic shadows across the ancient stone floor. The overall scene is imbued with the rich textures of a xianxia film, with magnificent and majestic visuals, drawing heavily on the vibrant and imaginative Taoist special effects aesthetics of \"The Legend of Zu,\" creating an atmosphere of sacredness, solemnity, and power.", + "seed": 32865, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_0/index_6871_seed_32865_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": -0.14919885904528202, + "vlm_logprob": -0.14919885904528202, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.14919885904528202, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.6346833109855652 + ], + "opposite_logprob_tokens": [ + -0.7596833109855652 + ], + "logprob_reward": -0.6346833109855652, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6346833109855652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15434470772743225, + 0.13160157203674316, + 0.04239732399582863, + 0.091018445789814, + 0.13647736608982086, + 0.06321579217910767, + 0.049039531499147415, + 0.10628875344991684, + 0.06682492047548294, + 0.15879163146018982 + ], + "credit_map": [ + 0.023622192442417145, + 0.018947694450616837, + 0.021069010719656944, + 0.01853281818330288, + 0.015994420275092125, + 0.018079834058880806, + 0.01763998158276081, + 0.020458750426769257, + 0.011303972452878952, + 0.02443910948932171, + 0.015337993390858173, + 0.01159521285444498, + 0.029793022200465202, + 0.013197685591876507, + 0.017096543684601784, + 0.008838020265102386, + 0.00631522573530674, + 0.0041208514012396336, + 0.003848278895020485, + 0.004849026910960674, + 0.0075154672376811504, + 0.007038555573672056, + 0.003971129655838013, + 0.004738785792142153, + 0.011511617340147495, + 0.011233129538595676, + 0.011265097185969353, + 0.006695217918604612, + 0.010375161655247211, + 0.01869991607964039, + 0.008510243147611618, + 0.012728051282465458, + 0.023517563939094543, + 0.010867837816476822, + 0.022860253229737282, + 0.013453160412609577, + 0.012415580451488495, + 0.017914747819304466, + 0.0180906243622303, + 0.017357593402266502, + 0.010348757728934288, + 0.0065880222246050835, + 0.007995531894266605, + 0.0060181221924722195, + 0.006940930616110563, + 0.0077580176293849945, + 0.006167820189148188, + 0.011398582719266415, + 0.004956371150910854, + 0.00488917576149106, + 0.0055018882267177105, + 0.011385305784642696, + 0.004976646974682808, + 0.005393995903432369, + 0.007565724663436413, + 0.004370423499494791, + 0.008350646123290062, + 0.01057753711938858, + 0.012759207747876644, + 0.0226536113768816, + 0.013853784650564194, + 0.008071284741163254, + 0.017250146716833115, + 0.012772521935403347, + 0.005690449383109808, + 0.00711088627576828, + 0.01363642793148756, + 0.006133608054369688, + 0.008263015188276768, + 0.005369937513023615, + 0.006294669583439827, + 0.014325923286378384, + 0.013257949613034725, + 0.016350461170077324, + 0.03623885288834572, + 0.009340186603367329, + 0.020065123215317726, + 0.013149594888091087, + 0.013801933266222477, + 0.036587510257959366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.6346833109855652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.014209638349711895 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209638349711895, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209638349711895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11942324787378311, + 0.2914460003376007, + 0.2004222869873047, + 0.04613364487886429, + 0.1138237863779068, + 0.07062507420778275, + 0.0401734858751297, + 0.038521118462085724, + 0.029032152146100998, + 0.05039920657873154 + ], + "credit_map": [ + 0.010507493279874325, + 0.008342714048922062, + 0.012426048517227173, + 0.01825416088104248, + 0.008963807485997677, + 0.015162208117544651, + 0.01926872879266739, + 0.026498086750507355, + 0.025054147467017174, + 0.05327128991484642, + 0.03190983831882477, + 0.025562960654497147, + 0.06617286056280136, + 0.03583577275276184, + 0.033992789685726166, + 0.019646331667900085, + 0.024058185517787933, + 0.02477792091667652, + 0.02067650854587555, + 0.02668006718158722, + 0.03304430842399597, + 0.031155651435256004, + 0.018401728942990303, + 0.02162790670990944, + 0.0050303093157708645, + 0.005198772996664047, + 0.00461063114926219, + 0.0033744587562978268, + 0.004290188197046518, + 0.012691136449575424, + 0.004387791268527508, + 0.006550355814397335, + 0.011887529864907265, + 0.013450146652758121, + 0.017306195572018623, + 0.009296988137066364, + 0.014068445190787315, + 0.018159683793783188, + 0.007628359831869602, + 0.02202642895281315, + 0.011153784580528736, + 0.011037365533411503, + 0.007572336588054895, + 0.006639537401497364, + 0.011388672515749931, + 0.008052827790379524, + 0.007048820145428181, + 0.007731731981039047, + 0.0035659822169691324, + 0.004444315563887358, + 0.005026187282055616, + 0.010205459780991077, + 0.004959693644195795, + 0.003547328058630228, + 0.0054732817225158215, + 0.0029512399341911077, + 0.0028818794526159763, + 0.005347439553588629, + 0.004400753881782293, + 0.0067780655808746815, + 0.005105603486299515, + 0.002931612078100443, + 0.005484157707542181, + 0.005591605789959431, + 0.0017268520314246416, + 0.0032673559617251158, + 0.004911045543849468, + 0.005316469818353653, + 0.0037642912939190865, + 0.0014982367865741253, + 0.004016819875687361, + 0.004531082697212696, + 0.006147358100861311, + 0.007279878947883844, + 0.005005403887480497, + 0.003227112116292119, + 0.01077941618859768, + 0.008584036491811275, + 0.005191382020711899, + 0.0041846162639558315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.014209638349711895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.062068451195955276 + ], + "opposite_logprob_tokens": [ + -2.812068462371826 + ], + "logprob_reward": -0.062068451195955276, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062068451195955276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07226552069187164, + 0.26386550068855286, + 0.15206529200077057, + 0.03329763188958168, + 0.039403967559337616, + 0.028422439470887184, + 0.0990636870265007, + 0.046916663646698, + 0.08684305101633072, + 0.17785629630088806 + ], + "credit_map": [ + 0.008807990700006485, + 0.008213229477405548, + 0.008009339682757854, + 0.0089336596429348, + 0.007304469123482704, + 0.009251849725842476, + 0.012037011794745922, + 0.009707962162792683, + 0.022905156016349792, + 0.049403585493564606, + 0.02909111976623535, + 0.023820457980036736, + 0.06044617295265198, + 0.03088206984102726, + 0.03040737845003605, + 0.0169095229357481, + 0.01454134937375784, + 0.016564276069402695, + 0.01412411779165268, + 0.016284102573990822, + 0.0301993228495121, + 0.029013078659772873, + 0.013883329927921295, + 0.017455706372857094, + 0.0031704141292721033, + 0.003469779621809721, + 0.002785861724987626, + 0.0036281589418649673, + 0.0036802352406084538, + 0.008054102770984173, + 0.0038951083552092314, + 0.004613970406353474, + 0.00653922650963068, + 0.0029856960754841566, + 0.006331616081297398, + 0.0030369821470230818, + 0.004364839754998684, + 0.007162732537835836, + 0.003420448862016201, + 0.005562420003116131, + 0.004098043777048588, + 0.0041807894594967365, + 0.003496238263323903, + 0.002970689907670021, + 0.003850300330668688, + 0.00348104489967227, + 0.00304012605920434, + 0.00330520560964942, + 0.009356317110359669, + 0.010685937479138374, + 0.012936225160956383, + 0.022106800228357315, + 0.011451894417405128, + 0.009394953027367592, + 0.015742983669042587, + 0.007388575933873653, + 0.00358033599331975, + 0.006199385039508343, + 0.006239274051040411, + 0.009253044612705708, + 0.007436245214194059, + 0.003304386045783758, + 0.007694314233958721, + 0.0032096768263727427, + 0.006866013631224632, + 0.01152582373470068, + 0.011274445801973343, + 0.013472208753228188, + 0.013650745153427124, + 0.006284923292696476, + 0.011006001383066177, + 0.012762884609401226, + 0.02126084826886654, + 0.023418651893734932, + 0.0204978808760643, + 0.01549145020544529, + 0.03306804224848747, + 0.028321441262960434, + 0.018681928515434265, + 0.017116056755185127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 2 + ], + "selected_reward": -0.062068451195955276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.0013367536012083292 + ], + "opposite_logprob_tokens": [ + -6.626336574554443 + ], + "logprob_reward": -0.0013367536012083292, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013367536012083292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0991092324256897, + 0.04099570959806442, + 0.1190863624215126, + 0.031046126037836075, + 0.054728493094444275, + 0.07843814045190811, + 0.07396632432937622, + 0.060700494796037674, + 0.13138709962368011, + 0.3105420172214508 + ], + "credit_map": [ + 0.011926714330911636, + 0.01058595534414053, + 0.005291075445711613, + 0.013992867432534695, + 0.011883108876645565, + 0.008263668045401573, + 0.015974460169672966, + 0.02119137905538082, + 0.004882107954472303, + 0.006464036647230387, + 0.006398254539817572, + 0.00478212907910347, + 0.006268003955483437, + 0.003222005208954215, + 0.004373794887214899, + 0.004605374298989773, + 0.01064345147460699, + 0.011202694848179817, + 0.010135404765605927, + 0.010081478394567966, + 0.026700332760810852, + 0.025923000648617744, + 0.010627624578773975, + 0.01377237681299448, + 0.004371491260826588, + 0.003124201437458396, + 0.0036150028463453054, + 0.002013513818383217, + 0.0029116421937942505, + 0.00918311532586813, + 0.001872656517662108, + 0.0039545041508972645, + 0.008833302184939384, + 0.0057702199555933475, + 0.008054670877754688, + 0.0049178521148860455, + 0.006888008210808039, + 0.006144402548670769, + 0.005891335196793079, + 0.008228704333305359, + 0.010671772994101048, + 0.012272228486835957, + 0.009103036485612392, + 0.006763341370970011, + 0.011147960089147091, + 0.012243352830410004, + 0.006677606608718634, + 0.00955884251743555, + 0.00507090101018548, + 0.009840825572609901, + 0.008730282075703144, + 0.01682625710964203, + 0.010297001339495182, + 0.006752327084541321, + 0.00980737991631031, + 0.006641349755227566, + 0.005802837200462818, + 0.008840806782245636, + 0.005516435019671917, + 0.009931778535246849, + 0.009026161395013332, + 0.0065391031093895435, + 0.007720188237726688, + 0.0073231845162808895, + 0.011882429011166096, + 0.019086845219135284, + 0.015452575869858265, + 0.020489167422056198, + 0.018364282324910164, + 0.010515823028981686, + 0.014897041954100132, + 0.020698942244052887, + 0.04033777490258217, + 0.04190090298652649, + 0.0308819729834795, + 0.021950021386146545, + 0.06021158769726753, + 0.055853281170129776, + 0.03111928701400757, + 0.028287222608923912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.0013367536012083292 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start at eye level and rapidly zoom upwards to a bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a white-clad female warrior seated cross-legged at the center of the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do eight identical swords descend from the sky and form a perfect Bagua formation around the warrior?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the swords spin synchronously to create a vast, hemispherical, blue, transparent energy barrier with tiny runes on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a mountaintop platform suspended above a sea of clouds with a backdrop of swirling purple and deep blue 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": [ + -0.033696141093969345 + ], + "opposite_logprob_tokens": [ + -3.408696174621582 + ], + "logprob_reward": -0.033696141093969345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033696141093969345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18779751658439636, + 0.1975676566362381, + 0.11277291178703308, + 0.06878233700990677, + 0.260615736246109, + 0.04747863858938217, + 0.02753375470638275, + 0.03344248607754707, + 0.020503295585513115, + 0.04350561648607254 + ], + "credit_map": [ + 0.017771292477846146, + 0.014936894178390503, + 0.014056971296668053, + 0.029437843710184097, + 0.020922604948282242, + 0.02119772881269455, + 0.03161722049117088, + 0.037856969982385635, + 0.016876304522156715, + 0.03158007189631462, + 0.028017116710543633, + 0.01831982098519802, + 0.0372869037091732, + 0.017789287492632866, + 0.02791484445333481, + 0.019783305004239082, + 0.009601738303899765, + 0.010398728772997856, + 0.008446735329926014, + 0.00953556690365076, + 0.030067531391978264, + 0.022450705990195274, + 0.01112122181802988, + 0.011150682345032692, + 0.007224287372082472, + 0.01031156349927187, + 0.007840811274945736, + 0.005971216131001711, + 0.0074987211264669895, + 0.013294212520122528, + 0.008289339952170849, + 0.008352180011570454, + 0.039322156459093094, + 0.01758585311472416, + 0.046858057379722595, + 0.021299397572875023, + 0.023111341521143913, + 0.05572457239031792, + 0.016788285225629807, + 0.03992606699466705, + 0.00601651705801487, + 0.007546464912593365, + 0.00565178832039237, + 0.005235596094280481, + 0.006685774773359299, + 0.005826197098940611, + 0.004302614368498325, + 0.006213688291609287, + 0.0030915907118469477, + 0.003513770643621683, + 0.003546283347532153, + 0.004772344138473272, + 0.003431010525673628, + 0.0027610263787209988, + 0.0038172812201082706, + 0.00260044913738966, + 0.0025449953973293304, + 0.002790951170027256, + 0.0049428194761276245, + 0.007356889545917511, + 0.0059663099236786366, + 0.0022683071438223124, + 0.005134957376867533, + 0.002437256509438157, + 0.001415569568052888, + 0.002537249820306897, + 0.002876816550269723, + 0.003215678734704852, + 0.0030930598732084036, + 0.001664337469264865, + 0.0027423198334872723, + 0.0029582649003714323, + 0.004634810611605644, + 0.005465488415211439, + 0.00847337394952774, + 0.003751105396077037, + 0.004821880254894495, + 0.005184749141335487, + 0.004678856581449509, + 0.006495353765785694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.033696141093969345 + } + ], + "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/0000158/group_1/index_tensor([1491, 1491])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_tensor([1491, 1491])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d89d33b72e774fa976ddeaedbe42492a0d807ab --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_tensor([1491, 1491])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, telephoto lens, shot from above, focuses on the vast battlefield, where thick, blood-red clouds of smoke swirl and surge violently. As if imbued with life, the smoke begins to converge and swirl in a preternatural manner. From this amorphous chaos, a towering silhouette slowly sculpted by an invisible hand takes shape. The first thing to emerge is a general with a face as red as a date and a majestic expression. His signature narrow, phoenix-like eyes, half-closed, reveal a penetrating sharpness. His thick, ornate black beard flows gracefully through the air, a breathtaking display of grace. He clads in emerald green armor formed from condensed smoke, its intricate dragon patterns faintly visible in the dim light. In his muscular right hand, he clutches the legendary Blue Dragon Crescent Blade. The massive blade emanates a cold azure light, creating a stark contrast to the crimson clouds surrounding it. The compressed depth of field of the lens makes the outline of the battlefield below blurred, and only broken flags and sporadic beacons can be vaguely seen. The whole picture is full of epic solemnity and sacredness, as if a legendary god of war has descended at this moment, creating an awe-inspiring, almost static fantasy atmosphere.", + "seed": 4143, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4143_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": -0.17742331665940583, + "vlm_logprob": -0.17742331665940583, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.17742331665940583, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8426038026809692 + ], + "opposite_logprob_tokens": [ + -0.5926038026809692 + ], + "logprob_reward": -0.8426038026809692, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8426038026809692, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07403189688920975, + 0.07155417650938034, + 0.035226769745349884, + 0.5772002935409546, + 0.014029394835233688, + 0.037952128797769547, + 0.04804462939500809, + 0.10131010413169861, + 0.01732873171567917, + 0.023321911692619324 + ], + "credit_map": [ + 0.01063497643917799, + 0.0064222845248878, + 0.011159664019942284, + 0.010180437937378883, + 0.005658186972141266, + 0.009626084938645363, + 0.007984395138919353, + 0.012365859933197498, + 0.020473752170801163, + 0.004325107671320438, + 0.004431457258760929, + 0.0046342951245605946, + 0.004238734487444162, + 0.025268707424402237, + 0.0042246682569384575, + 0.003957449458539486, + 0.0028474444989115, + 0.00297466735355556, + 0.007263713050633669, + 0.00378434918820858, + 0.0031773673836141825, + 0.004805405158549547, + 0.0046399058774113655, + 0.005733917001634836, + 0.018350277096033096, + 0.01874527335166931, + 0.021003903821110725, + 0.14188848435878754, + 0.019259138032794, + 0.23178648948669434, + 0.03495359793305397, + 0.09121309220790863, + 0.0017614172538742423, + 0.0015417929971590638, + 0.0014114546356722713, + 0.0016250786138698459, + 0.001525898347608745, + 0.001769443741068244, + 0.002285819035023451, + 0.0021084898617118597, + 0.0022798317950218916, + 0.0030908004846423864, + 0.005251388996839523, + 0.002882112516090274, + 0.0045879860408604145, + 0.007726989686489105, + 0.004457813687622547, + 0.007675205357372761, + 0.0037424489855766296, + 0.004826598335057497, + 0.007194479927420616, + 0.004574111197143793, + 0.009208264760673046, + 0.00578452181071043, + 0.006521511822938919, + 0.006192689761519432, + 0.006363117601722479, + 0.0048791225999593735, + 0.007420318201184273, + 0.005099055822938681, + 0.004593866877257824, + 0.03622422367334366, + 0.003652423620223999, + 0.03307797759771347, + 0.0015763217816129327, + 0.0017147057224065065, + 0.00251941685564816, + 0.0020255637355148792, + 0.0028524240478873253, + 0.002093702321872115, + 0.0021432414650917053, + 0.0024033566005527973, + 0.0016678298125043511, + 0.0037458986043930054, + 0.0034719647374004126, + 0.0020111382473260164, + 0.0035201937425881624, + 0.003235262120142579, + 0.0021918355487287045, + 0.0034777887631207705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.8426038026809692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.0032431883737444878 + ], + "opposite_logprob_tokens": [ + -5.7532429695129395 + ], + "logprob_reward": -0.0032431883737444878, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032431883737444878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10660437494516373, + 0.2685783803462982, + 0.1094735711812973, + 0.19309309124946594, + 0.2059783637523651, + 0.020407257601618767, + 0.029679233208298683, + 0.03611064329743385, + 0.013122421689331532, + 0.016952620819211006 + ], + "credit_map": [ + 0.015299546532332897, + 0.009038847871124744, + 0.014527801424264908, + 0.01534397341310978, + 0.010521492920815945, + 0.014238090254366398, + 0.016798969358205795, + 0.010835663415491581, + 0.07984659075737, + 0.015637734904885292, + 0.011157304979860783, + 0.014408783987164497, + 0.014176344498991966, + 0.10679059475660324, + 0.011555246077477932, + 0.015005807392299175, + 0.00833026971668005, + 0.011972601525485516, + 0.011331859044730663, + 0.012314001098275185, + 0.01476500928401947, + 0.01532247569411993, + 0.01679133251309395, + 0.018646037206053734, + 0.006685454864054918, + 0.005911652464419603, + 0.00828396063297987, + 0.03359869122505188, + 0.007215296383947134, + 0.0825309306383133, + 0.011068386025726795, + 0.03779873996973038, + 0.006972884759306908, + 0.004541377071291208, + 0.007739196997135878, + 0.015589005313813686, + 0.007743268273770809, + 0.0873321145772934, + 0.007584117352962494, + 0.06847642362117767, + 0.0022526297252625227, + 0.0031823248136788607, + 0.0030344976112246513, + 0.0021383981220424175, + 0.0030953404493629932, + 0.0022490888368338346, + 0.0022708121687173843, + 0.0021841684356331825, + 0.0020336436573415995, + 0.0028644015546888113, + 0.004771266132593155, + 0.0029558155220001936, + 0.005608129780739546, + 0.00401298888027668, + 0.0035934168845415115, + 0.0038395735900849104, + 0.002091142348945141, + 0.0023985581938177347, + 0.0029596572276204824, + 0.0027059942949563265, + 0.0022544118110090494, + 0.011410538107156754, + 0.0021326744463294744, + 0.010157668963074684, + 0.0015249979915097356, + 0.0013804149348288774, + 0.0015948747750371695, + 0.0016815022099763155, + 0.0015708078863099217, + 0.0016963817179203033, + 0.0017977304523810744, + 0.001875713118351996, + 0.0016031792620196939, + 0.0019705011509358883, + 0.0020911938045173883, + 0.0014199920697137713, + 0.0036265903618186712, + 0.0022984931711107492, + 0.002162798075005412, + 0.00177987280767411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.0032431883737444878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.008643477223813534 + ], + "opposite_logprob_tokens": [ + -4.758643627166748 + ], + "logprob_reward": -0.008643477223813534, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643477223813534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060669511556625366, + 0.07603325694799423, + 0.10118671506643295, + 0.4223860204219818, + 0.023047039285302162, + 0.023886503651738167, + 0.07238272577524185, + 0.15043120086193085, + 0.024902019649744034, + 0.04507499933242798 + ], + "credit_map": [ + 0.006576926447451115, + 0.005367093253880739, + 0.012071930803358555, + 0.010045868344604969, + 0.005345585756003857, + 0.006020069122314453, + 0.008031059987843037, + 0.0072109755128622055, + 0.02245183102786541, + 0.00436674989759922, + 0.004201487638056278, + 0.004210276063531637, + 0.0048177484422922134, + 0.028022227808833122, + 0.0038901588413864374, + 0.004072778858244419, + 0.00822457391768694, + 0.010265376418828964, + 0.010311642661690712, + 0.013484022580087185, + 0.011637602001428604, + 0.013600521720945835, + 0.016708752140402794, + 0.01695423200726509, + 0.01826426014304161, + 0.016263028606772423, + 0.017954740673303604, + 0.08087189495563507, + 0.01859925501048565, + 0.171209454536438, + 0.023120438680052757, + 0.07610297203063965, + 0.0024057775735855103, + 0.0017375668976455927, + 0.0025117136538028717, + 0.002342511899769306, + 0.003130453871563077, + 0.004845916293561459, + 0.0034338715486228466, + 0.00263922824524343, + 0.0019904959481209517, + 0.003161810338497162, + 0.0033242919016629457, + 0.0021190717816352844, + 0.004342782776802778, + 0.002745002508163452, + 0.002899906598031521, + 0.0033031427301466465, + 0.0054816617630422115, + 0.009851313196122646, + 0.011614286340773106, + 0.008151683025062084, + 0.014692481607198715, + 0.006632007658481598, + 0.007734659127891064, + 0.008224633522331715, + 0.008773209527134895, + 0.006999137345701456, + 0.0119563527405262, + 0.005166022107005119, + 0.009159001521766186, + 0.05286269262433052, + 0.004772713407874107, + 0.05074208602309227, + 0.0027129692025482655, + 0.0038653439842164516, + 0.0025491181295365095, + 0.0025839272420853376, + 0.004129443317651749, + 0.003018685383722186, + 0.0025008737575262785, + 0.003541659563779831, + 0.0037596055772155523, + 0.006198191083967686, + 0.003911841660737991, + 0.003583890153095126, + 0.012443963438272476, + 0.006143752485513687, + 0.0046576461754739285, + 0.0043761092238128185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.008643477223813534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.002822702983394265 + ], + "opposite_logprob_tokens": [ + -5.8778228759765625 + ], + "logprob_reward": -0.002822702983394265, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822702983394265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27900898456573486, + 0.024392297491431236, + 0.0885925218462944, + 0.09778668731451035, + 0.20982272922992706, + 0.158850759267807, + 0.03061792068183422, + 0.05071468651294708, + 0.025368209928274155, + 0.03484520688652992 + ], + "credit_map": [ + 0.03471626713871956, + 0.0261104553937912, + 0.0557071715593338, + 0.03046468459069729, + 0.022705813869833946, + 0.030915454030036926, + 0.03755226731300354, + 0.040836866945028305, + 0.00438223872333765, + 0.0021414561197161674, + 0.0025836965069174767, + 0.00422772066667676, + 0.0029540362302213907, + 0.003247118555009365, + 0.002811945741996169, + 0.002044086577370763, + 0.007046678103506565, + 0.007654760032892227, + 0.009694868698716164, + 0.021560801193118095, + 0.00777563638985157, + 0.007942495867609978, + 0.017248610034585, + 0.009668667800724506, + 0.006159254349768162, + 0.005918291863054037, + 0.005012417212128639, + 0.018385598435997963, + 0.008389593102037907, + 0.019737396389245987, + 0.01348381582647562, + 0.020700326189398766, + 0.008175505325198174, + 0.005527090281248093, + 0.00764156598597765, + 0.016655858606100082, + 0.009981770068407059, + 0.0881037563085556, + 0.008058332838118076, + 0.06567884236574173, + 0.012463420629501343, + 0.014641215093433857, + 0.02359739877283573, + 0.014240318909287453, + 0.016859080642461777, + 0.03028896637260914, + 0.018213503062725067, + 0.028546854853630066, + 0.002206874080002308, + 0.0029534597415477037, + 0.004686080850660801, + 0.005854317452758551, + 0.0048477137461304665, + 0.004300546832382679, + 0.0030431835912168026, + 0.0027257439214736223, + 0.0033163700718432665, + 0.0045306989923119545, + 0.004584293346852064, + 0.0054651424288749695, + 0.0045567168854177, + 0.011117617599666119, + 0.006168280728161335, + 0.01097556296736002, + 0.0033672794234007597, + 0.004334812052547932, + 0.0028313950169831514, + 0.0026930219028145075, + 0.0041086734272539616, + 0.0030835706274956465, + 0.002543659182265401, + 0.0024057975970208645, + 0.0028378991410136223, + 0.004595894366502762, + 0.003654411993920803, + 0.0023782344069331884, + 0.006270235870033503, + 0.00424911268055439, + 0.006919081788510084, + 0.0039403378032147884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.002822702983394265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.029803412035107613 + ], + "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": [ + 0.2588297724723816, + 0.04365614429116249, + 0.04576408118009567, + 0.2953847050666809, + 0.046942051500082016, + 0.20172221958637238, + 0.034621741622686386, + 0.042891789227724075, + 0.014147660695016384, + 0.016039833426475525 + ], + "credit_map": [ + 0.0367979072034359, + 0.022609863430261612, + 0.03819285333156586, + 0.031562332063913345, + 0.01932644285261631, + 0.03439999371767044, + 0.03465551137924194, + 0.04128486290574074, + 0.009853626601397991, + 0.0028253551572561264, + 0.003509971546009183, + 0.005585220642387867, + 0.00402596453204751, + 0.009875421412289143, + 0.004021967761218548, + 0.003958615940064192, + 0.004147856496274471, + 0.004462387412786484, + 0.0070319934748113155, + 0.007852663286030293, + 0.004579359199851751, + 0.004377030301839113, + 0.00660310685634613, + 0.006709683686494827, + 0.010638708248734474, + 0.010791641660034657, + 0.010404697619378567, + 0.08317511528730392, + 0.010635203681886196, + 0.10742802172899246, + 0.024598773568868637, + 0.037712570279836655, + 0.002060752362012863, + 0.0015803721034899354, + 0.002236683154478669, + 0.004050895571708679, + 0.0021990586537867785, + 0.018617231398820877, + 0.0023894875776022673, + 0.013807572424411774, + 0.011331855319440365, + 0.013977042399346828, + 0.02361748367547989, + 0.013781345449388027, + 0.017308799549937248, + 0.054625049233436584, + 0.018004601821303368, + 0.049076057970523834, + 0.0032471283338963985, + 0.003393938997760415, + 0.005209305323660374, + 0.0035795315634459257, + 0.006392453797161579, + 0.0038319083396345377, + 0.0045701987110078335, + 0.004397277720272541, + 0.003164883702993393, + 0.004285797011107206, + 0.004508784506469965, + 0.00444264430552721, + 0.0036636115983128548, + 0.010529871098697186, + 0.0028758032713085413, + 0.009420396760106087, + 0.0014072778867557645, + 0.001455428428016603, + 0.002050390699878335, + 0.001663584727793932, + 0.0019075173186138272, + 0.0017919634701684117, + 0.0014417603379115462, + 0.0024297386407852173, + 0.001851141918450594, + 0.002280249260365963, + 0.0016357216518372297, + 0.0015752507606521249, + 0.002786438213661313, + 0.002081911778077483, + 0.0017110659973695874, + 0.002118053613230586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.029803412035107613 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, telephoto lens, shot from above, focuses on the vast battlefield, where thick, blood-red clouds of smoke swirl and surge violently. As if imbued with life, the smoke begins to converge and swirl in a preternatural manner. From this amorphous chaos, a towering silhouette slowly sculpted by an invisible hand takes shape. The first thing to emerge is a general with a face as red as a date and a majestic expression. His signature narrow, phoenix-like eyes, half-closed, reveal a penetrating sharpness. His thick, ornate black beard flows gracefully through the air, a breathtaking display of grace. He clads in emerald green armor formed from condensed smoke, its intricate dragon patterns faintly visible in the dim light. In his muscular right hand, he clutches the legendary Blue Dragon Crescent Blade. The massive blade emanates a cold azure light, creating a stark contrast to the crimson clouds surrounding it. The compressed depth of field of the lens makes the outline of the battlefield below blurred, and only broken flags and sporadic beacons can be vaguely seen. The whole picture is full of epic solemnity and sacredness, as if a legendary god of war has descended at this moment, creating an awe-inspiring, almost static fantasy atmosphere.", + "seed": 4144, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4144_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": -0.10727166063152253, + "vlm_logprob": -0.10727166063152253, + "vlm_yes_no_margin": 4.174999994039536, + "vlm_token_credit": -0.10727166063152253, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.49185290932655334 + ], + "opposite_logprob_tokens": [ + -0.991852879524231 + ], + "logprob_reward": -0.49185290932655334, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.49185290932655334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11135537177324295, + 0.04458592087030411, + 0.20001186430454254, + 0.07042605429887772, + 0.059574928134679794, + 0.17689847946166992, + 0.136013925075531, + 0.06899074465036392, + 0.06449375301599503, + 0.06764889508485794 + ], + "credit_map": [ + 0.013237737119197845, + 0.01090281829237938, + 0.022408686578273773, + 0.012420126236975193, + 0.009228886105120182, + 0.015675876289606094, + 0.01171526312828064, + 0.01576598733663559, + 0.004725384991616011, + 0.006501138210296631, + 0.005307769402861595, + 0.006179102696478367, + 0.005860972683876753, + 0.008076447062194347, + 0.003683660877868533, + 0.004251448903232813, + 0.01084316335618496, + 0.01108647882938385, + 0.010037768632173538, + 0.0697992742061615, + 0.013404777273535728, + 0.01146648358553648, + 0.06405569612979889, + 0.009318232536315918, + 0.005232122261077166, + 0.007381500210613012, + 0.009698683395981789, + 0.01359617430716753, + 0.008595743216574192, + 0.010421430692076683, + 0.007930781692266464, + 0.007569627836346626, + 0.005822010338306427, + 0.01019179169088602, + 0.006198389455676079, + 0.008900085464119911, + 0.007503098342567682, + 0.007583956699818373, + 0.006813484244048595, + 0.006562117021530867, + 0.0115558672696352, + 0.01618964597582817, + 0.022842757403850555, + 0.024116355925798416, + 0.02166525088250637, + 0.03827960044145584, + 0.0283783208578825, + 0.013870698399841785, + 0.01617428846657276, + 0.007788694929331541, + 0.011372871696949005, + 0.009255629032850266, + 0.012163609266281128, + 0.035206906497478485, + 0.033679451793432236, + 0.010372485034167767, + 0.005603961646556854, + 0.005768211558461189, + 0.005391004029661417, + 0.009576083160936832, + 0.01567724533379078, + 0.007290431763976812, + 0.012963858433067799, + 0.006719949655234814, + 0.0072326562367379665, + 0.010813814587891102, + 0.009037693031132221, + 0.009968547150492668, + 0.007922164164483547, + 0.00528684863820672, + 0.0089559992775321, + 0.0052760266698896885, + 0.00573463412001729, + 0.01088033989071846, + 0.008858045563101768, + 0.011364268139004707, + 0.007763697300106287, + 0.005651819985359907, + 0.010452687740325928, + 0.00694340281188488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.49185290932655334 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.002538674511015415 + ], + "opposite_logprob_tokens": [ + -6.002538681030273 + ], + "logprob_reward": -0.002538674511015415, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002538674511015415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16651777923107147, + 0.09308528155088425, + 0.20656277239322662, + 0.09210711717605591, + 0.03870011121034622, + 0.12929923832416534, + 0.13295011222362518, + 0.04621570557355881, + 0.0572182796895504, + 0.03734366595745087 + ], + "credit_map": [ + 0.01943506859242916, + 0.02172507904469967, + 0.021009471267461777, + 0.019751960411667824, + 0.018884077668190002, + 0.0240035317838192, + 0.019812708720564842, + 0.02189585566520691, + 0.008747000247240067, + 0.014880964532494545, + 0.0062775807455182076, + 0.01899670623242855, + 0.01175482477992773, + 0.01568167842924595, + 0.00924394465982914, + 0.007502574939280748, + 0.020174695178866386, + 0.03627417981624603, + 0.02137700468301773, + 0.022035421803593636, + 0.02763650193810463, + 0.037642862647771835, + 0.022302037104964256, + 0.019120050594210625, + 0.009603512473404408, + 0.009330840781331062, + 0.013560140505433083, + 0.011994907632470131, + 0.010036846622824669, + 0.010737068019807339, + 0.011505275964736938, + 0.015338510274887085, + 0.003801270853728056, + 0.0046222819946706295, + 0.005181759130209684, + 0.00486385403200984, + 0.004469023551791906, + 0.004289734642952681, + 0.0070362333208322525, + 0.004435947630554438, + 0.010945357382297516, + 0.012572684325277805, + 0.01716489903628826, + 0.01689334399998188, + 0.017908375710248947, + 0.018190009519457817, + 0.019394200295209885, + 0.01623033918440342, + 0.012784837745130062, + 0.007162000983953476, + 0.009274362586438656, + 0.007908841595053673, + 0.015257298946380615, + 0.03600266948342323, + 0.03564465790987015, + 0.008915427140891552, + 0.005235025659203529, + 0.003922665491700172, + 0.0044644670560956, + 0.006424626801162958, + 0.00721079483628273, + 0.005029892083257437, + 0.008971449919044971, + 0.004956779535859823, + 0.007509020622819662, + 0.007724577561020851, + 0.010506832972168922, + 0.007673847954720259, + 0.005364063195884228, + 0.005006636027246714, + 0.007914870977401733, + 0.0055184271186590195, + 0.003633833257481456, + 0.005057104863226414, + 0.0054705459624528885, + 0.004469194449484348, + 0.0047913771122694016, + 0.004104527644813061, + 0.00543761532753706, + 0.00437946105375886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.002538674511015415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.003624063218012452 + ], + "opposite_logprob_tokens": [ + -5.628623962402344 + ], + "logprob_reward": -0.003624063218012452, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624063218012452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11270423978567123, + 0.026693182066082954, + 0.15137338638305664, + 0.06380072981119156, + 0.04030516371130943, + 0.14615529775619507, + 0.17310915887355804, + 0.14538007974624634, + 0.07911078631877899, + 0.06136798486113548 + ], + "credit_map": [ + 0.00995676964521408, + 0.009988874197006226, + 0.017191505059599876, + 0.01799853891134262, + 0.009297746233642101, + 0.012772466987371445, + 0.013455343432724476, + 0.02204299159348011, + 0.0022177009377628565, + 0.0041241818107664585, + 0.0024123333860188723, + 0.005081631243228912, + 0.0038540572859346867, + 0.0032449786085635424, + 0.002804941264912486, + 0.002953357994556427, + 0.007694859057664871, + 0.00930960662662983, + 0.008503896184265614, + 0.050590697675943375, + 0.012915787287056446, + 0.008565644733607769, + 0.04664871096611023, + 0.007144175004214048, + 0.005968800280243158, + 0.01140141673386097, + 0.00961027666926384, + 0.00821045320481062, + 0.007433454506099224, + 0.008131137117743492, + 0.006420627236366272, + 0.00662456126883626, + 0.0038830963894724846, + 0.006951413583010435, + 0.004483250435441732, + 0.0044242823496460915, + 0.005061606876552105, + 0.004270660690963268, + 0.005677177105098963, + 0.0055536748841404915, + 0.00702203344553709, + 0.017018279060721397, + 0.015658486634492874, + 0.01626482978463173, + 0.03096611052751541, + 0.01771925948560238, + 0.025629518553614616, + 0.015876781195402145, + 0.009419619105756283, + 0.006393096875399351, + 0.009141326881945133, + 0.010382777079939842, + 0.012610931880772114, + 0.05573264881968498, + 0.05664342641830444, + 0.012785341590642929, + 0.006410977803170681, + 0.01044374518096447, + 0.009829030372202396, + 0.02070925384759903, + 0.026916736736893654, + 0.01733335293829441, + 0.037748731672763824, + 0.01598825305700302, + 0.006335557904094458, + 0.008941021747887135, + 0.014601824805140495, + 0.011665573343634605, + 0.009357727132737637, + 0.00717986561357975, + 0.012705042958259583, + 0.008324173279106617, + 0.0057511599734425545, + 0.008970093913376331, + 0.00943136028945446, + 0.007127663120627403, + 0.008776534348726273, + 0.005342811346054077, + 0.01040616724640131, + 0.005562194623053074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.003624063218012452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.004640405531972647 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.004640405531972647, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640405531972647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11871400475502014, + 0.062028467655181885, + 0.15363799035549164, + 0.0674687996506691, + 0.02842343971133232, + 0.22112910449504852, + 0.07357488572597504, + 0.08586706221103668, + 0.09909665584564209, + 0.0900595411658287 + ], + "credit_map": [ + 0.010259930975735188, + 0.010229542851448059, + 0.01820128597319126, + 0.019099624827504158, + 0.01158470194786787, + 0.013694868423044682, + 0.012309908866882324, + 0.023334138095378876, + 0.006683690939098597, + 0.009676346555352211, + 0.006015004124492407, + 0.010166482999920845, + 0.007685997989028692, + 0.010959913954138756, + 0.004980400670319796, + 0.005860632751137018, + 0.006932710763067007, + 0.008677748963236809, + 0.007381344214081764, + 0.05633161589503288, + 0.008216147311031818, + 0.007063701748847961, + 0.052534885704517365, + 0.006499835290014744, + 0.007589779328554869, + 0.012933596037328243, + 0.01020771637558937, + 0.008695525117218494, + 0.006738949101418257, + 0.007447156123816967, + 0.006560889538377523, + 0.007295187562704086, + 0.0029048281721770763, + 0.004382400307804346, + 0.002901127329096198, + 0.003103993833065033, + 0.003171122632920742, + 0.0035707317292690277, + 0.004320347681641579, + 0.004068886861205101, + 0.025666577741503716, + 0.024443652480840683, + 0.03115847520530224, + 0.021855760365724564, + 0.027104506269097328, + 0.02907690592110157, + 0.03621629998087883, + 0.02560693398118019, + 0.005932242143899202, + 0.010440708138048649, + 0.004664173815399408, + 0.011502890847623348, + 0.007824383676052094, + 0.01070583425462246, + 0.014303471893072128, + 0.008201180957257748, + 0.006757412105798721, + 0.009044984355568886, + 0.007992839440703392, + 0.01397472620010376, + 0.014589413069188595, + 0.009404600597918034, + 0.016239918768405914, + 0.007863170467317104, + 0.011243894696235657, + 0.011137640103697777, + 0.020268257707357407, + 0.014652170240879059, + 0.008562402799725533, + 0.009019018150866032, + 0.014104406349360943, + 0.010108868591487408, + 0.010548258200287819, + 0.009961416944861412, + 0.012409771792590618, + 0.01612883247435093, + 0.007143196649849415, + 0.008281704969704151, + 0.015419946983456612, + 0.010166417807340622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.004640405531972647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.03370225057005882 + ], + "opposite_logprob_tokens": [ + -3.4087021350860596 + ], + "logprob_reward": -0.03370225057005882, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370225057005882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17940033972263336, + 0.1498095989227295, + 0.15302059054374695, + 0.047015316784381866, + 0.06516045331954956, + 0.14162801206111908, + 0.12378383427858353, + 0.06654119491577148, + 0.03940913826227188, + 0.034231606870889664 + ], + "credit_map": [ + 0.018238918855786324, + 0.01611482910811901, + 0.028086859732866287, + 0.028824737295508385, + 0.01398861687630415, + 0.0169349517673254, + 0.020286669954657555, + 0.03692474961280823, + 0.01636841893196106, + 0.02210835926234722, + 0.013822103850543499, + 0.02302897348999977, + 0.018263274803757668, + 0.030440842732787132, + 0.013323396444320679, + 0.01245423685759306, + 0.007440150715410709, + 0.008118662983179092, + 0.007643554825335741, + 0.05658348277211189, + 0.007309570908546448, + 0.007735465187579393, + 0.051540836691856384, + 0.0066488622687757015, + 0.004953014198690653, + 0.004675065167248249, + 0.00852894876152277, + 0.006702776998281479, + 0.005496036261320114, + 0.0070111071690917015, + 0.004878724459558725, + 0.004769641440361738, + 0.0070642465725541115, + 0.007808991242200136, + 0.007853485643863678, + 0.011269468814134598, + 0.0066717686131596565, + 0.008670372888445854, + 0.006427712272852659, + 0.009394405409693718, + 0.010060352273285389, + 0.013158229179680347, + 0.021628960967063904, + 0.019756248220801353, + 0.02043888345360756, + 0.01943085715174675, + 0.021842237561941147, + 0.015312233939766884, + 0.009763820096850395, + 0.008200511336326599, + 0.007166313473135233, + 0.008147251792252064, + 0.01596628874540329, + 0.033270757645368576, + 0.03193254768848419, + 0.009336334653198719, + 0.004779146984219551, + 0.007242267951369286, + 0.006348737049847841, + 0.007157129235565662, + 0.01440821960568428, + 0.007525681518018246, + 0.011159156449139118, + 0.007920855656266212, + 0.003894353983923793, + 0.0044884332455694675, + 0.006443181540817022, + 0.005092641804367304, + 0.005500332452356815, + 0.003742798464372754, + 0.0053085871040821075, + 0.004938810598105192, + 0.003288182197138667, + 0.0032322106417268515, + 0.005237626377493143, + 0.005722117144614458, + 0.0031302948482334614, + 0.0032337228767573833, + 0.005342756398022175, + 0.005044695921242237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03370225057005882 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, telephoto lens, shot from above, focuses on the vast battlefield, where thick, blood-red clouds of smoke swirl and surge violently. As if imbued with life, the smoke begins to converge and swirl in a preternatural manner. From this amorphous chaos, a towering silhouette slowly sculpted by an invisible hand takes shape. The first thing to emerge is a general with a face as red as a date and a majestic expression. His signature narrow, phoenix-like eyes, half-closed, reveal a penetrating sharpness. His thick, ornate black beard flows gracefully through the air, a breathtaking display of grace. He clads in emerald green armor formed from condensed smoke, its intricate dragon patterns faintly visible in the dim light. In his muscular right hand, he clutches the legendary Blue Dragon Crescent Blade. The massive blade emanates a cold azure light, creating a stark contrast to the crimson clouds surrounding it. The compressed depth of field of the lens makes the outline of the battlefield below blurred, and only broken flags and sporadic beacons can be vaguely seen. The whole picture is full of epic solemnity and sacredness, as if a legendary god of war has descended at this moment, creating an awe-inspiring, almost static fantasy atmosphere.", + "seed": 4145, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4145_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": -0.16874540839344263, + "vlm_logprob": -0.16874540839344263, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.16874540839344263, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.779387891292572 + ], + "opposite_logprob_tokens": [ + -0.654387891292572 + ], + "logprob_reward": -0.779387891292572, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.779387891292572, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.38751232624053955, + 0.02936401404440403, + 0.021029654890298843, + 0.04512546584010124, + 0.04135618731379509, + 0.032105494290590286, + 0.36504125595092773, + 0.033390216529369354, + 0.023622974753379822, + 0.021452367305755615 + ], + "credit_map": [ + 0.02773321606218815, + 0.013946730643510818, + 0.014071978628635406, + 0.15387390553951263, + 0.012440932914614677, + 0.011911834590137005, + 0.13823363184928894, + 0.015300130471587181, + 0.0026123139541596174, + 0.0035311344545334578, + 0.0030753372702747583, + 0.006279537919908762, + 0.002020767889916897, + 0.0031895977444946766, + 0.005548399873077869, + 0.0031069281976670027, + 0.0017587722977623343, + 0.0022210380993783474, + 0.0023763212375342846, + 0.0028215390630066395, + 0.0022610239684581757, + 0.003092107828706503, + 0.0028780673164874315, + 0.0036207884550094604, + 0.004386475775390863, + 0.00841804314404726, + 0.0042161415331065655, + 0.013792171142995358, + 0.0025890578981488943, + 0.0022097760811448097, + 0.0036814184859395027, + 0.005832384806126356, + 0.009987697936594486, + 0.005491756368428469, + 0.005615188740193844, + 0.006243588402867317, + 0.0032384353689849377, + 0.0032474168110638857, + 0.0030216784216463566, + 0.0045104301534593105, + 0.0028956797905266285, + 0.0038057423662394285, + 0.0029101918917149305, + 0.0044503542594611645, + 0.005582892335951328, + 0.00449116388335824, + 0.003217985387891531, + 0.004751486703753471, + 0.010691708885133266, + 0.011823748238384724, + 0.009519031271338463, + 0.014582626521587372, + 0.021224893629550934, + 0.14259131252765656, + 0.023563124239444733, + 0.13104486465454102, + 0.003534313989803195, + 0.0028260827530175447, + 0.00310243247076869, + 0.0032864019740372896, + 0.004432070534676313, + 0.006275640334933996, + 0.003460356267169118, + 0.00647292286157608, + 0.002640146529302001, + 0.0017562236171215773, + 0.002911811927333474, + 0.002272950019687414, + 0.002539545064792037, + 0.003653774270787835, + 0.004149663727730513, + 0.003698861226439476, + 0.00234571797773242, + 0.002371582668274641, + 0.001812012167647481, + 0.002181390533223748, + 0.0036708731204271317, + 0.0033043574076145887, + 0.002927660010755062, + 0.0028387748170644045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.779387891292572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.0059962146915495396 + ], + "opposite_logprob_tokens": [ + -5.130996227264404 + ], + "logprob_reward": -0.0059962146915495396, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059962146915495396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.435428261756897, + 0.09118930250406265, + 0.038374800235033035, + 0.039048805832862854, + 0.11144756525754929, + 0.1164865717291832, + 0.09797458350658417, + 0.019901616498827934, + 0.014277799986302853, + 0.03587069734930992 + ], + "credit_map": [ + 0.05829813331365585, + 0.03420752286911011, + 0.028046607971191406, + 0.08543886989355087, + 0.027442550286650658, + 0.03845452517271042, + 0.13369134068489075, + 0.029848720878362656, + 0.006980510428547859, + 0.006681753322482109, + 0.008245659992098808, + 0.021737638860940933, + 0.0060391854494810104, + 0.010307862423360348, + 0.02343892492353916, + 0.007757765706628561, + 0.003547340165823698, + 0.0045346156693995, + 0.0047068786807358265, + 0.0061622546054422855, + 0.004785609431564808, + 0.004812501836568117, + 0.004326401278376579, + 0.005499198101460934, + 0.0019928719848394394, + 0.006388010922819376, + 0.0024663745425641537, + 0.013685226440429688, + 0.0022108410485088825, + 0.0018720610532909632, + 0.0028673706110566854, + 0.00756604690104723, + 0.01370043121278286, + 0.011974127031862736, + 0.010452616028487682, + 0.014188679866492748, + 0.012255039997398853, + 0.015104536898434162, + 0.013877700082957745, + 0.0198944341391325, + 0.007883337326347828, + 0.009338737465441227, + 0.01097086165100336, + 0.014048819430172443, + 0.01822200044989586, + 0.023113276809453964, + 0.01227912213653326, + 0.02063041180372238, + 0.004059903789311647, + 0.003453850280493498, + 0.002017449354752898, + 0.0036829945165663958, + 0.0060927304439246655, + 0.03926162049174309, + 0.006259409710764885, + 0.03314661607146263, + 0.0018136901780962944, + 0.0014302899362519383, + 0.0011793465819209814, + 0.002411717316135764, + 0.002384594874456525, + 0.003995003644376993, + 0.002307521877810359, + 0.004379451274871826, + 0.00160456623416394, + 0.001583981211297214, + 0.001262792618945241, + 0.0016510117566213012, + 0.0020157117396593094, + 0.0023538500536233187, + 0.002022780478000641, + 0.0017831061268225312, + 0.002392415190115571, + 0.0036959401331841946, + 0.0027427824679762125, + 0.003233254887163639, + 0.006209618877619505, + 0.006065437104552984, + 0.008027936331927776, + 0.0035033123567700386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0059962146915495396 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.0076352194882929325 + ], + "opposite_logprob_tokens": [ + -4.882635116577148 + ], + "logprob_reward": -0.0076352194882929325, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076352194882929325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7801269292831421, + 0.013827778398990631, + 0.01054424699395895, + 0.021251700818538666, + 0.015228803269565105, + 0.02152658998966217, + 0.056276969611644745, + 0.025902004912495613, + 0.023402702063322067, + 0.031912315636873245 + ], + "credit_map": [ + 0.09908678382635117, + 0.03101121261715889, + 0.029014883562922478, + 0.3078417479991913, + 0.02145388536155224, + 0.027274277061223984, + 0.231188103556633, + 0.033256035298109055, + 0.00134855427313596, + 0.0012453628005459905, + 0.001170662697404623, + 0.0017152288928627968, + 0.0019581413362175226, + 0.0019544819369912148, + 0.0029551335610449314, + 0.0014802125515416265, + 0.0008302964270114899, + 0.0010880541522055864, + 0.0008321324130520225, + 0.002484857337549329, + 0.0010478697950020432, + 0.0014658552827313542, + 0.0014129910850897431, + 0.0013821899192407727, + 0.001448728027753532, + 0.00542990630492568, + 0.001361436559818685, + 0.006331083830446005, + 0.0019832102116197348, + 0.0013618429657071829, + 0.0018637795001268387, + 0.0014717127196490765, + 0.002982670906931162, + 0.0025746196042746305, + 0.0016162439715117216, + 0.0022563228849321604, + 0.0017476421780884266, + 0.0011117758695036173, + 0.0013511108700186014, + 0.001588417449966073, + 0.002618319820612669, + 0.00293134362436831, + 0.0020667987409979105, + 0.0031481445766985416, + 0.003219782840460539, + 0.00214821333065629, + 0.0021117376163601875, + 0.003282249439507723, + 0.00210898881778121, + 0.0022060235496610403, + 0.0014751814305782318, + 0.0025912022683769464, + 0.004192420281469822, + 0.020274804905056953, + 0.003898054361343384, + 0.019530296325683594, + 0.001833000686019659, + 0.002102714031934738, + 0.0025170326698571444, + 0.0019860246684402227, + 0.0031586436089128256, + 0.004962089471518993, + 0.003142937086522579, + 0.006199561059474945, + 0.0013625986175611615, + 0.002673941198736429, + 0.0016705634770914912, + 0.0024475972168147564, + 0.004777206107974052, + 0.004026448354125023, + 0.003601900301873684, + 0.0028424463234841824, + 0.001991111785173416, + 0.0032466405536979437, + 0.0023242018651217222, + 0.0027516805566847324, + 0.005856014788150787, + 0.005876941606402397, + 0.006720549426972866, + 0.003145175287500024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.0076352194882929325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.007637467235326767 + ], + "opposite_logprob_tokens": [ + -4.8826375007629395 + ], + "logprob_reward": -0.007637467235326767, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007637467235326767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7700346112251282, + 0.018117636442184448, + 0.007256020791828632, + 0.007800500374287367, + 0.01052937749773264, + 0.015928134322166443, + 0.047061894088983536, + 0.019003575667738914, + 0.018047092482447624, + 0.08622117340564728 + ], + "credit_map": [ + 0.09017302840948105, + 0.0291859433054924, + 0.02782558649778366, + 0.30567601323127747, + 0.021008983254432678, + 0.026063866913318634, + 0.2378406524658203, + 0.032260503619909286, + 0.0018285327823832631, + 0.0019350466318428516, + 0.0017861024243757129, + 0.00249978294596076, + 0.0025979464408010244, + 0.002139601157978177, + 0.0035515963099896908, + 0.001779027166776359, + 0.0009218620834872127, + 0.0009044582839123905, + 0.000832411169540137, + 0.001278811483643949, + 0.0006645178073085845, + 0.0008468175074085593, + 0.0008782604127191007, + 0.000928881810978055, + 0.0005823977407999337, + 0.0020181720610708, + 0.0005718945758417249, + 0.0013776045525446534, + 0.0006416917312890291, + 0.0006954980199225247, + 0.0006547404918819666, + 0.0012585013173520565, + 0.0025007904041558504, + 0.001525978441350162, + 0.0012538896407932043, + 0.0013465860392898321, + 0.0009390206541866064, + 0.0007953567546792328, + 0.000925432366784662, + 0.0012423230800777674, + 0.00227085268124938, + 0.0019305185414850712, + 0.0015083058970049024, + 0.0021945489570498466, + 0.0025520978961139917, + 0.0016348612261936069, + 0.0015878572594374418, + 0.0022490916308015585, + 0.0016368833603337407, + 0.00181464827619493, + 0.001117619569413364, + 0.002081887563690543, + 0.0031104804947972298, + 0.018656527623534203, + 0.0028841288294643164, + 0.015759719535708427, + 0.0023624678142368793, + 0.0019583841785788536, + 0.0015606082743033767, + 0.002700171433389187, + 0.002503918716683984, + 0.0024888066109269857, + 0.0029703134205192327, + 0.0024589039385318756, + 0.0013737203553318977, + 0.002097642282024026, + 0.0014121872372925282, + 0.002229204634204507, + 0.0032075364142656326, + 0.0030169817619025707, + 0.0023267334327101707, + 0.0023830861318856478, + 0.004808929283171892, + 0.009008381515741348, + 0.006588731426745653, + 0.00717539805918932, + 0.015522327274084091, + 0.01575489342212677, + 0.01844927668571472, + 0.00891323946416378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.007637467235326767 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.04307024925947189 + ], + "opposite_logprob_tokens": [ + -3.1680703163146973 + ], + "logprob_reward": -0.04307024925947189, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307024925947189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32114946842193604, + 0.04716223105788231, + 0.034925173968076706, + 0.07869629561901093, + 0.05524623766541481, + 0.11086099594831467, + 0.2733844518661499, + 0.040364909917116165, + 0.019742343574762344, + 0.018467893823981285 + ], + "credit_map": [ + 0.021338896825909615, + 0.012729537673294544, + 0.012771527282893658, + 0.12253405898809433, + 0.010293884202837944, + 0.011807522736489773, + 0.11354219168424606, + 0.016131790354847908, + 0.004115878138691187, + 0.004660322330892086, + 0.004211730323731899, + 0.009883302263915539, + 0.0034459137823432684, + 0.005016250070184469, + 0.01170242577791214, + 0.004126403480768204, + 0.003106695367023349, + 0.0045703393407166, + 0.004750433377921581, + 0.0061027430929243565, + 0.004074994008988142, + 0.0032457420602440834, + 0.004426589235663414, + 0.004647633992135525, + 0.0038220181595534086, + 0.00954432599246502, + 0.004080420359969139, + 0.029857221990823746, + 0.0034314384683966637, + 0.003085355507209897, + 0.005867989268153906, + 0.019007517024874687, + 0.005014144349843264, + 0.0094945402815938, + 0.006796061061322689, + 0.004410519264638424, + 0.009455354884266853, + 0.0052232202142477036, + 0.00595690356567502, + 0.008895485661923885, + 0.005481795873492956, + 0.008781171403825283, + 0.009570092894136906, + 0.012828192673623562, + 0.017640244215726852, + 0.022064516320824623, + 0.011171714402735233, + 0.023323260247707367, + 0.006993155926465988, + 0.007505779154598713, + 0.004640224855393171, + 0.010206579230725765, + 0.014401504769921303, + 0.11139196157455444, + 0.015972843393683434, + 0.1022723838686943, + 0.0020315335132181644, + 0.0024908853229135275, + 0.0034926978405565023, + 0.002381891245022416, + 0.004421048332005739, + 0.009286709129810333, + 0.004433462396264076, + 0.01182667538523674, + 0.0019365495536476374, + 0.0017052178736776114, + 0.0019356518751010299, + 0.001873555826023221, + 0.0023391630966216326, + 0.0035646879114210606, + 0.003430455457419157, + 0.0029570593032985926, + 0.001965626608580351, + 0.002067388268187642, + 0.0016750454669818282, + 0.002160202246159315, + 0.0031109019182622433, + 0.0026439367793500423, + 0.0028533125296235085, + 0.001991475932300091 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.04307024925947189 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, telephoto lens, shot from above, focuses on the vast battlefield, where thick, blood-red clouds of smoke swirl and surge violently. As if imbued with life, the smoke begins to converge and swirl in a preternatural manner. From this amorphous chaos, a towering silhouette slowly sculpted by an invisible hand takes shape. The first thing to emerge is a general with a face as red as a date and a majestic expression. His signature narrow, phoenix-like eyes, half-closed, reveal a penetrating sharpness. His thick, ornate black beard flows gracefully through the air, a breathtaking display of grace. He clads in emerald green armor formed from condensed smoke, its intricate dragon patterns faintly visible in the dim light. In his muscular right hand, he clutches the legendary Blue Dragon Crescent Blade. The massive blade emanates a cold azure light, creating a stark contrast to the crimson clouds surrounding it. The compressed depth of field of the lens makes the outline of the battlefield below blurred, and only broken flags and sporadic beacons can be vaguely seen. The whole picture is full of epic solemnity and sacredness, as if a legendary god of war has descended at this moment, creating an awe-inspiring, almost static fantasy atmosphere.", + "seed": 4146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_1/index_1491_seed_4146_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": -0.27427639458328484, + "vlm_logprob": -0.27427639458328484, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.27427639458328484, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely visible?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2358227968215942 + ], + "opposite_logprob_tokens": [ + -0.3608228266239166 + ], + "logprob_reward": -1.2358227968215942, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2358227968215942, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11881344020366669, + 0.049735311418771744, + 0.09742949903011322, + 0.0691298395395279, + 0.08239820599555969, + 0.1516691893339157, + 0.04175994545221329, + 0.09215754270553589, + 0.21286644041538239, + 0.08404060453176498 + ], + "credit_map": [ + 0.010050253011286259, + 0.014529460109770298, + 0.017895350232720375, + 0.011016817763447762, + 0.022503016516566277, + 0.014780260622501373, + 0.012054595164954662, + 0.01598367653787136, + 0.005578306969255209, + 0.00560977216809988, + 0.009707252494990826, + 0.005771552678197622, + 0.0059011648409068584, + 0.004858642350882292, + 0.008524507284164429, + 0.00378410960547626, + 0.008829090744256973, + 0.01679312251508236, + 0.012399549596011639, + 0.012413921765983105, + 0.012366698123514652, + 0.014609378762543201, + 0.012598915956914425, + 0.007418815512210131, + 0.006004509050399065, + 0.004103883169591427, + 0.005226681940257549, + 0.01911340467631817, + 0.0037233643233776093, + 0.018109403550624847, + 0.007264761719852686, + 0.005583827383816242, + 0.007781995460391045, + 0.011756857857108116, + 0.011973700486123562, + 0.0080385347828269, + 0.007081026677042246, + 0.010573231615126133, + 0.014205435290932655, + 0.0109874177724123, + 0.004404430277645588, + 0.0037806793116033077, + 0.006218322552740574, + 0.00627093156799674, + 0.0064695170149207115, + 0.061754509806632996, + 0.0550837367773056, + 0.007687048986554146, + 0.004365146160125732, + 0.0035336546134203672, + 0.004248537588864565, + 0.005002887453883886, + 0.004261929076164961, + 0.005952070001512766, + 0.00896518025547266, + 0.005430539604276419, + 0.005582526326179504, + 0.005320761818438768, + 0.009149816818535328, + 0.00807999074459076, + 0.012094304896891117, + 0.020295394584536552, + 0.011018792167305946, + 0.020615942776203156, + 0.047992561012506485, + 0.020420869812369347, + 0.035802897065877914, + 0.01857098937034607, + 0.01785140484571457, + 0.03255873918533325, + 0.0184647049754858, + 0.021204277873039246, + 0.013491950929164886, + 0.005154330283403397, + 0.016534391790628433, + 0.00692782923579216, + 0.005766388960182667, + 0.010360277257859707, + 0.015022186562418938, + 0.01078324019908905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -1.2358227968215942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.00982921663671732 + ], + "opposite_logprob_tokens": [ + -4.634829044342041 + ], + "logprob_reward": -0.00982921663671732, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00982921663671732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08486629277467728, + 0.2167799323797226, + 0.07774924486875534, + 0.07190439850091934, + 0.05369645729660988, + 0.10389672219753265, + 0.028367456048727036, + 0.046915844082832336, + 0.23684464395046234, + 0.07897895574569702 + ], + "credit_map": [ + 0.008270569145679474, + 0.014065496623516083, + 0.011835300363600254, + 0.009843978099524975, + 0.010563928633928299, + 0.009606038220226765, + 0.011102436110377312, + 0.009578538127243519, + 0.014743207953870296, + 0.024385903030633926, + 0.03578275442123413, + 0.020742641761898994, + 0.026599649339914322, + 0.03345726802945137, + 0.037699148058891296, + 0.023369356989860535, + 0.009804756380617619, + 0.010431729257106781, + 0.009964209981262684, + 0.008920599706470966, + 0.010338102467358112, + 0.008845259435474873, + 0.01076480746269226, + 0.008679783903062344, + 0.008405770175158978, + 0.005607105791568756, + 0.007726064883172512, + 0.016657358035445213, + 0.004418227821588516, + 0.014492361806333065, + 0.008034159429371357, + 0.0065633500926196575, + 0.004394455812871456, + 0.00731577118858695, + 0.006423620972782373, + 0.005773720797151327, + 0.004924086853861809, + 0.0075338310562074184, + 0.012022284790873528, + 0.005308685824275017, + 0.00671142153441906, + 0.006289301440119743, + 0.006637962535023689, + 0.00711184972897172, + 0.007216960657387972, + 0.03326268121600151, + 0.028203153982758522, + 0.008463386446237564, + 0.002295911544933915, + 0.003293163375928998, + 0.002857714891433716, + 0.0026522709522396326, + 0.0035557267256081104, + 0.004787804558873177, + 0.004513977561146021, + 0.004410887137055397, + 0.008217946626245975, + 0.003459468251094222, + 0.0054784538224339485, + 0.00465978030115366, + 0.004565231967717409, + 0.006530021317303181, + 0.0062661366537213326, + 0.007738802582025528, + 0.052027393132448196, + 0.02231772430241108, + 0.040506940335035324, + 0.020734582096338272, + 0.019533831626176834, + 0.03803498297929764, + 0.020165788009762764, + 0.023523399606347084, + 0.011359138414263725, + 0.0043940008617937565, + 0.01270347274839878, + 0.0055468264035880566, + 0.006348209455609322, + 0.012481696903705597, + 0.013804170303046703, + 0.012341447174549103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 5 + ], + "selected_reward": -0.00982921663671732 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.04306728392839432 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.04306728392839432, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306728392839432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040706388652324677, + 0.02926746942102909, + 0.04941185936331749, + 0.12266451865434647, + 0.0531931146979332, + 0.1598200500011444, + 0.03566323593258858, + 0.13928699493408203, + 0.2718673348426819, + 0.09811904281377792 + ], + "credit_map": [ + 0.002629287773743272, + 0.00734941428527236, + 0.004507740493863821, + 0.0038313071709126234, + 0.006675004493445158, + 0.005180938169360161, + 0.0044644782319664955, + 0.006068218499422073, + 0.0034570195712149143, + 0.003921716473996639, + 0.004360205959528685, + 0.0031917663291096687, + 0.0029763549100607634, + 0.0026613285299390554, + 0.006249179597944021, + 0.002449898747727275, + 0.00577126070857048, + 0.00544553529471159, + 0.007026239763945341, + 0.005590892396867275, + 0.006592608988285065, + 0.009211858734488487, + 0.0052087390795350075, + 0.00456472672522068, + 0.005595708731561899, + 0.006125816144049168, + 0.007719659712165594, + 0.04529949650168419, + 0.005068774335086346, + 0.037652503699064255, + 0.008885407820343971, + 0.006317144259810448, + 0.006797181908041239, + 0.007106497418135405, + 0.00853743590414524, + 0.005721048451960087, + 0.007423426955938339, + 0.005143349524587393, + 0.005297200754284859, + 0.007166973780840635, + 0.01330321840941906, + 0.01922471635043621, + 0.016980024054646492, + 0.0141900684684515, + 0.015400421805679798, + 0.030658699572086334, + 0.032210104167461395, + 0.01785280928015709, + 0.004180374089628458, + 0.004201920237392187, + 0.003963171970099211, + 0.003124556504189968, + 0.00369453732855618, + 0.006520665716379881, + 0.005584686994552612, + 0.0043933214619755745, + 0.010818690061569214, + 0.005214724224060774, + 0.00865850504487753, + 0.0044185142032802105, + 0.009806222282350063, + 0.05014790594577789, + 0.008941940031945705, + 0.041280489414930344, + 0.05323479697108269, + 0.02922651916742325, + 0.040673960000276566, + 0.021979384124279022, + 0.03191889077425003, + 0.041750673204660416, + 0.026122499257326126, + 0.026960594579577446, + 0.011668361723423004, + 0.0045143174938857555, + 0.02047242410480976, + 0.006007709074765444, + 0.008082490414381027, + 0.015618919394910336, + 0.014661493711173534, + 0.017093323171138763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.04306728392839432 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.020594336092472076 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.020594336092472076, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594336092472076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11307790875434875, + 0.04824574291706085, + 0.07954427599906921, + 0.0657748281955719, + 0.06362941116094589, + 0.05013415217399597, + 0.07246816158294678, + 0.0676460862159729, + 0.22269806265830994, + 0.216781347990036 + ], + "credit_map": [ + 0.006465943064540625, + 0.030963437631726265, + 0.009509570896625519, + 0.012677719816565514, + 0.01286392007023096, + 0.012938245199620724, + 0.009015248157083988, + 0.018643826246261597, + 0.005819239187985659, + 0.00703092198818922, + 0.00615520216524601, + 0.005235216114670038, + 0.005566298495978117, + 0.004160745069384575, + 0.00986888911575079, + 0.0044092293828725815, + 0.015506397932767868, + 0.014447728171944618, + 0.010599871166050434, + 0.008827880024909973, + 0.008492814376950264, + 0.006958590354770422, + 0.006305531598627567, + 0.008405465632677078, + 0.007269201800227165, + 0.0034602463711053133, + 0.004387752152979374, + 0.02115206979215145, + 0.003833207068964839, + 0.0184328630566597, + 0.004144466016441584, + 0.003095025895163417, + 0.007677793502807617, + 0.007824258878827095, + 0.010497872717678547, + 0.007406511344015598, + 0.008799411356449127, + 0.00806245394051075, + 0.00636138254776597, + 0.0069997310638427734, + 0.003353750566020608, + 0.003613816574215889, + 0.005038300063461065, + 0.004212864208966494, + 0.0046158148907125, + 0.011605058796703815, + 0.01043127290904522, + 0.0072632753290236, + 0.009280548430979252, + 0.007603470701724291, + 0.007833976298570633, + 0.010258123278617859, + 0.008329367265105247, + 0.008775308728218079, + 0.013371890410780907, + 0.007015479262918234, + 0.007667952217161655, + 0.004557964392006397, + 0.007750098127871752, + 0.00594769511371851, + 0.008826098404824734, + 0.011308019980788231, + 0.008578813634812832, + 0.01300943922251463, + 0.04229919612407684, + 0.02127140201628208, + 0.03829246014356613, + 0.025165431201457977, + 0.023785801604390144, + 0.027948331087827682, + 0.023530350998044014, + 0.020405100658535957, + 0.031334131956100464, + 0.009710870683193207, + 0.034718167036771774, + 0.015020481310784817, + 0.016360092908143997, + 0.03521013632416725, + 0.03735829144716263, + 0.0370691753923893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020594336092472076 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high angle looking down at a battlefield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thick blood-red clouds of smoke swirling and surging in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a general with a face as red as a date and a thick ornate black beard visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the general wearing emerald green armor with faint dragon patterns and holding a blue glowing blade?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the depth of field compressed such that the battlefield background is blurred with only broken flags and beacons vaguely 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": [ + -0.06206833943724632 + ], + "opposite_logprob_tokens": [ + -2.812068223953247 + ], + "logprob_reward": -0.06206833943724632, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206833943724632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30792364478111267, + 0.0829671174287796, + 0.08716694265604019, + 0.09244424104690552, + 0.03779203072190285, + 0.1297728270292282, + 0.03769408538937569, + 0.11465272307395935, + 0.051754556596279144, + 0.057831887155771255 + ], + "credit_map": [ + 0.022710494697093964, + 0.04669308662414551, + 0.03567637875676155, + 0.032757822424173355, + 0.05618434399366379, + 0.037500761449337006, + 0.037556927651166916, + 0.03884381428360939, + 0.00863366574048996, + 0.010271008126437664, + 0.01687386818230152, + 0.0087843406945467, + 0.009639570489525795, + 0.007782856933772564, + 0.015017377212643623, + 0.005964427720755339, + 0.015686213970184326, + 0.011098779737949371, + 0.010380340740084648, + 0.009426907636225224, + 0.012057534419000149, + 0.011382394470274448, + 0.008538680151104927, + 0.008596089668571949, + 0.009228522889316082, + 0.004431731998920441, + 0.006998010445386171, + 0.03134584426879883, + 0.0037490185350179672, + 0.026988061144948006, + 0.005690282676368952, + 0.004012768622487783, + 0.003652398707345128, + 0.004609223920851946, + 0.007499137427657843, + 0.004514326341450214, + 0.004836194217205048, + 0.004476879723370075, + 0.004361772909760475, + 0.0038420981727540493, + 0.009211748838424683, + 0.008775364607572556, + 0.010583518072962761, + 0.0073594991117715836, + 0.007267961278557777, + 0.04147755727171898, + 0.038810137659311295, + 0.006287037394940853, + 0.003041634103283286, + 0.00376335671171546, + 0.00363400811329484, + 0.004278489388525486, + 0.00375588471069932, + 0.005072880070656538, + 0.008517514914274216, + 0.005630319472402334, + 0.006814835127443075, + 0.0038819594774395227, + 0.0078128632158041, + 0.0051109264604747295, + 0.007222117390483618, + 0.039104726165533066, + 0.011076606810092926, + 0.0336286798119545, + 0.009926700964570045, + 0.005912933964282274, + 0.00756757752969861, + 0.004493033513426781, + 0.005294439848512411, + 0.007750684395432472, + 0.004864560440182686, + 0.005944626405835152, + 0.007642028387635946, + 0.0033361969981342554, + 0.01827891729772091, + 0.004063061438500881, + 0.0033998070284724236, + 0.010978640988469124, + 0.005385951139032841, + 0.004747283644974232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.06206833943724632 + } + ], + "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/0000158/group_10/index_tensor([7997, 7997])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_tensor([7997, 7997])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f8d45849a0f65a818a303a0b000ee1e9f72146cc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_tensor([7997, 7997])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-definition, eye-level still shot captures the scene in a minimalist ink-wash style, subtly perceiving the delicate fibrous texture of the underlying rice paper. A long, serpentine, catfish-like creature rests motionless on the pale, sandy riverbed, its body conforming to the soft contours of the ground. The shot's intense detail renders subtle gradations of gray and deeply saturated black across its smooth skin with artistic precision, while its long, thin whiskers, delicately barbed, drift almost imperceptibly. Nearby, a sinuous seagrass reed, resembling bold calligraphy strokes, sways gently in an invisible current. A subtle, intermittent stream of small bubbles escapes the creature's gills, drifting slowly and deliberately upwards to disappear into the pale, empty water above, creating an atmosphere of tranquil and contemplative solitude.", + "seed": 12889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12889_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": -0.053125934582203625, + "vlm_logprob": -0.053125934582203625, + "vlm_yes_no_margin": 3.7749999523162843, + "vlm_token_credit": -0.053125934582203625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.02139057032763958 + ], + "opposite_logprob_tokens": [ + -4.021390438079834 + ], + "logprob_reward": -0.02139057032763958, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02139057032763958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27845874428749084, + 0.07497981190681458, + 0.0706755742430687, + 0.08867892622947693, + 0.06836865097284317, + 0.035171665251255035, + 0.11541321128606796, + 0.057840343564748764, + 0.07175297290086746, + 0.13866005837917328 + ], + "credit_map": [ + 0.039481550455093384, + 0.04620083048939705, + 0.03659016266465187, + 0.04096828028559685, + 0.030537039041519165, + 0.03853493183851242, + 0.020337769761681557, + 0.0258081816136837, + 0.007478099782019854, + 0.010173625312745571, + 0.010911442339420319, + 0.010535555891692638, + 0.009702709503471851, + 0.01124341320246458, + 0.007716277614235878, + 0.007218692917376757, + 0.007097989786416292, + 0.008767735213041306, + 0.009695328772068024, + 0.012631996534764767, + 0.006957375910133123, + 0.008469813503324986, + 0.009849216789007187, + 0.007206122390925884, + 0.0107876630499959, + 0.014077674597501755, + 0.01308042835444212, + 0.013255739584565163, + 0.008006956428289413, + 0.01074058748781681, + 0.008272356353700161, + 0.010457523167133331, + 0.007314767688512802, + 0.00798220094293356, + 0.009653187356889248, + 0.011317581869661808, + 0.00910734012722969, + 0.010138452984392643, + 0.00637248856946826, + 0.0064826300367712975, + 0.005009506829082966, + 0.0047009713016450405, + 0.005023236386477947, + 0.006965160369873047, + 0.0031625607516616583, + 0.005306772422045469, + 0.0026091698091477156, + 0.0023942883126437664, + 0.01021077111363411, + 0.014126274734735489, + 0.018659353256225586, + 0.0252713430672884, + 0.012728485278785229, + 0.015312676317989826, + 0.01033351756632328, + 0.008770786225795746, + 0.006198075599968433, + 0.006638259161263704, + 0.009619513526558876, + 0.015695970505475998, + 0.0051752678118646145, + 0.006318180356174707, + 0.005018887110054493, + 0.0031761883292347193, + 0.008407843299210072, + 0.009583814069628716, + 0.010775960981845856, + 0.01260028500109911, + 0.0074074845761060715, + 0.009287755005061626, + 0.008244147524237633, + 0.005445684306323528, + 0.014273149892687798, + 0.021193740889430046, + 0.024402529001235962, + 0.04025702178478241, + 0.008752903901040554, + 0.010115769691765308, + 0.012691337615251541, + 0.00697360560297966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.02139057032763958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.005261146929115057 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005261146929115057, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261146929115057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23456574976444244, + 0.07425929605960846, + 0.07372091710567474, + 0.02469535358250141, + 0.2829894721508026, + 0.06280339509248734, + 0.0458805151283741, + 0.09427610039710999, + 0.04007592052221298, + 0.06673325598239899 + ], + "credit_map": [ + 0.027893932536244392, + 0.02510269545018673, + 0.03341295197606087, + 0.026763148605823517, + 0.03371885418891907, + 0.04552725329995155, + 0.016079293563961983, + 0.026067616418004036, + 0.008296464569866657, + 0.008082639425992966, + 0.011110669933259487, + 0.011387554928660393, + 0.00794330332428217, + 0.012253385968506336, + 0.007558414246886969, + 0.007626862730830908, + 0.011214185506105423, + 0.011717118322849274, + 0.010122108273208141, + 0.015225829556584358, + 0.005300131160765886, + 0.003970233723521233, + 0.011470685712993145, + 0.0047006201930344105, + 0.0027622755151242018, + 0.002221978036686778, + 0.0037821857258677483, + 0.002936414908617735, + 0.003173244884237647, + 0.0041164979338645935, + 0.0030867173336446285, + 0.0026160392444580793, + 0.02711445465683937, + 0.039429858326911926, + 0.0435885414481163, + 0.059849753975868225, + 0.028562093153595924, + 0.038752514868974686, + 0.0238124281167984, + 0.02187984809279442, + 0.0067153070122003555, + 0.008777997456490993, + 0.009111734107136726, + 0.010171923786401749, + 0.007767008151859045, + 0.008721844293177128, + 0.006108145695179701, + 0.005429434590041637, + 0.003831341629847884, + 0.006132896989583969, + 0.007690648082643747, + 0.010329757817089558, + 0.0047989049926400185, + 0.005415759515017271, + 0.004101385362446308, + 0.0035798191092908382, + 0.008807741105556488, + 0.012061198242008686, + 0.014927400276064873, + 0.018976470455527306, + 0.010319017805159092, + 0.014547602273523808, + 0.007531024515628815, + 0.007105647586286068, + 0.005773664452135563, + 0.004841125104576349, + 0.005994696635752916, + 0.007370581850409508, + 0.004157351329922676, + 0.005131702870130539, + 0.00350048765540123, + 0.0033063089940696955, + 0.0067286863923072815, + 0.010184992104768753, + 0.01197043340653181, + 0.01979818008840084, + 0.00447951490059495, + 0.00465375604107976, + 0.005705286283046007, + 0.0032124125864356756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.005261146929115057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.02635815367102623 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.02635815367102623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635815367102623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4307595193386078, + 0.1069660559296608, + 0.08773103356361389, + 0.03702760487794876, + 0.06295277923345566, + 0.05200504884123802, + 0.06993724405765533, + 0.07205520570278168, + 0.04882461205124855, + 0.0317409485578537 + ], + "credit_map": [ + 0.04909832775592804, + 0.04910809546709061, + 0.05826283618807793, + 0.05169899761676788, + 0.06030963733792305, + 0.07346704602241516, + 0.03620302304625511, + 0.05261155217885971, + 0.009443088434636593, + 0.01914805918931961, + 0.01756175421178341, + 0.019849400967359543, + 0.00951068289577961, + 0.008666902780532837, + 0.01401832140982151, + 0.008767838589847088, + 0.008939926512539387, + 0.01298418827354908, + 0.012486157938838005, + 0.011642182245850563, + 0.01120823621749878, + 0.012473815120756626, + 0.009958595968782902, + 0.008037930354475975, + 0.0039926012977957726, + 0.00586584210395813, + 0.005680452566593885, + 0.008405345492064953, + 0.0030646335799247026, + 0.002829998964443803, + 0.0041344789788126945, + 0.0030542544554919004, + 0.005118292290717363, + 0.007455962244421244, + 0.009624127298593521, + 0.013144337572157383, + 0.007025447208434343, + 0.009785093367099762, + 0.005808358546346426, + 0.004991158843040466, + 0.006195568945258856, + 0.006454971618950367, + 0.007132094353437424, + 0.006866664160043001, + 0.007092359941452742, + 0.01002285536378622, + 0.003989935852587223, + 0.004250602331012487, + 0.004915573168545961, + 0.006658819038420916, + 0.011661502532660961, + 0.02104339189827442, + 0.007635432295501232, + 0.006527311634272337, + 0.007435874547809362, + 0.00405934127047658, + 0.007480186875909567, + 0.009825631976127625, + 0.010697461664676666, + 0.009187517687678337, + 0.010569480247795582, + 0.012251128442585468, + 0.005579841323196888, + 0.006463954690843821, + 0.005839582998305559, + 0.005359484814107418, + 0.007880997844040394, + 0.012116852216422558, + 0.004958581645041704, + 0.004549648147076368, + 0.004602003376930952, + 0.0035174586810171604, + 0.003854840761050582, + 0.007288060151040554, + 0.004215830471366644, + 0.004670530091971159, + 0.003508480731397867, + 0.0033496334217488766, + 0.0028812186792492867, + 0.0019723554141819477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.02635815367102623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.01107061468064785 + ], + "opposite_logprob_tokens": [ + -4.511070728302002 + ], + "logprob_reward": -0.01107061468064785, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107061468064785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14442548155784607, + 0.21354897320270538, + 0.11711280047893524, + 0.09473568946123123, + 0.2768554091453552, + 0.030433356761932373, + 0.046371523290872574, + 0.030624788254499435, + 0.020785272121429443, + 0.025106636807322502 + ], + "credit_map": [ + 0.01218760572373867, + 0.02108152024447918, + 0.021273406222462654, + 0.018666643649339676, + 0.021168451756238937, + 0.02531559020280838, + 0.009613989852368832, + 0.015118293464183807, + 0.02038257196545601, + 0.027145573869347572, + 0.030818624421954155, + 0.02836010791361332, + 0.02742334082722664, + 0.0382453091442585, + 0.022005777806043625, + 0.019167693331837654, + 0.011005229316651821, + 0.010985607281327248, + 0.01696578599512577, + 0.014809254556894302, + 0.015449000522494316, + 0.018573828041553497, + 0.015067080035805702, + 0.01425702404230833, + 0.007717239670455456, + 0.01028518844395876, + 0.014538763090968132, + 0.009275069460272789, + 0.015365122817456722, + 0.017929820343852043, + 0.007167655974626541, + 0.012456840835511684, + 0.02730604074895382, + 0.02906700037419796, + 0.03939233720302582, + 0.03129393234848976, + 0.04133756831288338, + 0.056545767933130264, + 0.022171027958393097, + 0.02974173240363598, + 0.0032230739016085863, + 0.004231558181345463, + 0.004772745538502932, + 0.007299291901290417, + 0.0026822062209248543, + 0.003227589186280966, + 0.0033085516188293695, + 0.001688341610133648, + 0.004680763464421034, + 0.006263040006160736, + 0.007048433646559715, + 0.006210057996213436, + 0.006857403088361025, + 0.006875758990645409, + 0.003760704305022955, + 0.0046753655187785625, + 0.0036143753677606583, + 0.004221510607749224, + 0.004636878147721291, + 0.005207052454352379, + 0.0036599093582481146, + 0.004308559466153383, + 0.002539978828281164, + 0.002436527982354164, + 0.003203128231689334, + 0.003191847587004304, + 0.0030073784291744232, + 0.0030973851680755615, + 0.0023379006888717413, + 0.0021781499963253736, + 0.0020454733166843653, + 0.0017240092856809497, + 0.0022751467768102884, + 0.0037107039242982864, + 0.0047562094405293465, + 0.005915466230362654, + 0.002400087658315897, + 0.0022697432432323694, + 0.00227755936793983, + 0.001501721329987049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.01107061468064785 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.20154918730258942 + ], + "opposite_logprob_tokens": [ + -1.7015491724014282 + ], + "logprob_reward": -0.20154918730258942, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154918730258942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24976249039173126, + 0.10185113549232483, + 0.06023469194769859, + 0.029991138726472855, + 0.14299747347831726, + 0.06011536344885826, + 0.032395653426647186, + 0.04132357984781265, + 0.13225366175174713, + 0.1490747630596161 + ], + "credit_map": [ + 0.02634471282362938, + 0.027315830811858177, + 0.03536548465490341, + 0.028917046263813972, + 0.03786636143922806, + 0.04369724169373512, + 0.019818739965558052, + 0.03043709136545658, + 0.007575551513582468, + 0.010180693119764328, + 0.016300665214657784, + 0.014945040456950665, + 0.012749004177749157, + 0.017689155414700508, + 0.01113278977572918, + 0.011278239078819752, + 0.006853925064206123, + 0.008690888993442059, + 0.00841852743178606, + 0.010315208695828915, + 0.004828830249607563, + 0.005352536682039499, + 0.010247980244457722, + 0.0055268011055886745, + 0.002876811660826206, + 0.004677247256040573, + 0.004339020699262619, + 0.005106503609567881, + 0.0031853222753852606, + 0.0030975330155342817, + 0.0031882256735116243, + 0.003520475933328271, + 0.01114485040307045, + 0.018823349848389626, + 0.022531835362315178, + 0.032852061092853546, + 0.01464072149246931, + 0.019863620400428772, + 0.01328370813280344, + 0.009857334196567535, + 0.0065980833023786545, + 0.007520043756812811, + 0.00846074614673853, + 0.010306029580533504, + 0.007311667315661907, + 0.011167144402861595, + 0.004439564887434244, + 0.004312086384743452, + 0.0028960793279111385, + 0.0029522418044507504, + 0.004899896681308746, + 0.008192338049411774, + 0.004197821021080017, + 0.003976654261350632, + 0.003276646137237549, + 0.002003979403525591, + 0.0045607504434883595, + 0.005507179535925388, + 0.006277202628552914, + 0.007107652258127928, + 0.005089389160275459, + 0.006291084457188845, + 0.0031416364945471287, + 0.003348683472722769, + 0.015691697597503662, + 0.015805989503860474, + 0.020455632358789444, + 0.031206276267766953, + 0.011279123835265636, + 0.015308507718145847, + 0.013438143767416477, + 0.009068291634321213, + 0.013851319439709187, + 0.019067171961069107, + 0.02687348611652851, + 0.046600982546806335, + 0.008685325272381306, + 0.01138647273182869, + 0.015111175365746021, + 0.0074988435953855515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.20154918730258942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition, eye-level still shot captures the scene in a minimalist ink-wash style, subtly perceiving the delicate fibrous texture of the underlying rice paper. A long, serpentine, catfish-like creature rests motionless on the pale, sandy riverbed, its body conforming to the soft contours of the ground. The shot's intense detail renders subtle gradations of gray and deeply saturated black across its smooth skin with artistic precision, while its long, thin whiskers, delicately barbed, drift almost imperceptibly. Nearby, a sinuous seagrass reed, resembling bold calligraphy strokes, sways gently in an invisible current. A subtle, intermittent stream of small bubbles escapes the creature's gills, drifting slowly and deliberately upwards to disappear into the pale, empty water above, creating an atmosphere of tranquil and contemplative solitude.", + "seed": 12890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12890_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": -0.08121448326855898, + "vlm_logprob": -0.08121448326855898, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.08121448326855898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.046223487704992294 + ], + "opposite_logprob_tokens": [ + -3.1712234020233154 + ], + "logprob_reward": -0.046223487704992294, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.046223487704992294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5085097551345825, + 0.14028309285640717, + 0.0656684935092926, + 0.04536554217338562, + 0.03168158605694771, + 0.035926319658756256, + 0.02893568016588688, + 0.06859704107046127, + 0.021474536508321762, + 0.053558047860860825 + ], + "credit_map": [ + 0.028378628194332123, + 0.06518729776144028, + 0.0955342948436737, + 0.043696291744709015, + 0.04331782087683678, + 0.07898495346307755, + 0.06364238262176514, + 0.08976805955171585, + 0.007857874035835266, + 0.019478436559438705, + 0.022676199674606323, + 0.01520301029086113, + 0.015032880939543247, + 0.015181472525000572, + 0.021507907658815384, + 0.023345312103629112, + 0.004275696352124214, + 0.013348027132451534, + 0.01110045425593853, + 0.006247248966246843, + 0.005552131216973066, + 0.006915309000760317, + 0.007558012846857309, + 0.010671613737940788, + 0.0026730394456535578, + 0.004984351806342602, + 0.009553592652082443, + 0.0033228520769625902, + 0.0051757837645709515, + 0.005211420822888613, + 0.004217560403048992, + 0.010226939804852009, + 0.002737994771450758, + 0.003803773084655404, + 0.00597739452496171, + 0.0033189367968589067, + 0.0028170363511890173, + 0.0026894614566117525, + 0.0035597647074609995, + 0.006777224130928516, + 0.0032444149255752563, + 0.004612891469150782, + 0.005730446893721819, + 0.003691945457831025, + 0.003864259459078312, + 0.005053723696619272, + 0.0047088488936424255, + 0.00501979049295187, + 0.002514240564778447, + 0.0073524359613657, + 0.004702301695942879, + 0.0025484818033874035, + 0.0022610698360949755, + 0.0028863882180303335, + 0.0024679817724972963, + 0.004202777985483408, + 0.007710350211709738, + 0.015443289652466774, + 0.01208518072962761, + 0.005175286438316107, + 0.0056379251182079315, + 0.008475679904222488, + 0.00470466073602438, + 0.009364670142531395, + 0.0021055606193840504, + 0.003057036316022277, + 0.003292362205684185, + 0.002709850901737809, + 0.0022861675824970007, + 0.002345481188967824, + 0.002273287158459425, + 0.0034047907683998346, + 0.004739087074995041, + 0.007626102771610022, + 0.01073276624083519, + 0.0042929030023515224, + 0.0031517352908849716, + 0.005720483139157295, + 0.0034058024175465107, + 0.013889167457818985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.046223487704992294 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.005957583896815777 + ], + "opposite_logprob_tokens": [ + -5.13095760345459 + ], + "logprob_reward": -0.005957583896815777, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957583896815777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18094277381896973, + 0.22478751838207245, + 0.045502450317144394, + 0.05187993496656418, + 0.03652488440275192, + 0.03192280977964401, + 0.052081286907196045, + 0.22314172983169556, + 0.08254130184650421, + 0.0706753358244896 + ], + "credit_map": [ + 0.017973642796278, + 0.03342089802026749, + 0.031287118792533875, + 0.013912972062826157, + 0.015970412641763687, + 0.023436153307557106, + 0.018973706290125847, + 0.025967862457036972, + 0.04233931750059128, + 0.03831807151436806, + 0.03489309176802635, + 0.019283996894955635, + 0.01838407851755619, + 0.02253376506268978, + 0.021686207503080368, + 0.027348993346095085, + 0.004077313933521509, + 0.00965050794184208, + 0.008814374916255474, + 0.005178938154131174, + 0.0037068137899041176, + 0.004346870351582766, + 0.004414548631757498, + 0.005313082132488489, + 0.006029230542480946, + 0.007263297215104103, + 0.012018725275993347, + 0.003848293097689748, + 0.003610273590311408, + 0.004804490599781275, + 0.0031229795422405005, + 0.011182649061083794, + 0.005802962463349104, + 0.00477256765589118, + 0.008291440084576607, + 0.002329444047063589, + 0.002237604698166251, + 0.0038439719937741756, + 0.0028080216143280268, + 0.006438871379941702, + 0.004752189852297306, + 0.004948013927787542, + 0.005176853388547897, + 0.002720079617574811, + 0.0027451387140899897, + 0.004153044894337654, + 0.003237507538869977, + 0.004189982078969479, + 0.005970472004264593, + 0.01020501833409071, + 0.00975002534687519, + 0.003800696926191449, + 0.003773914184421301, + 0.006440322380512953, + 0.005129373632371426, + 0.007011460606008768, + 0.016349272802472115, + 0.053733840584754944, + 0.04448631778359413, + 0.014133625663816929, + 0.019267505034804344, + 0.02450421452522278, + 0.016878172755241394, + 0.0337887704372406, + 0.01115802675485611, + 0.014459236524999142, + 0.016622217372059822, + 0.006314159836620092, + 0.005180452950298786, + 0.011512565426528454, + 0.0058564129285514355, + 0.011438229121267796, + 0.007227989379316568, + 0.007932274602353573, + 0.011601842939853668, + 0.005700814072042704, + 0.008522236719727516, + 0.010908819735050201, + 0.0068235574290156364, + 0.011957800015807152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.005957583896815777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.03371411934494972 + ], + "opposite_logprob_tokens": [ + -3.4087140560150146 + ], + "logprob_reward": -0.03371411934494972, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371411934494972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39734429121017456, + 0.13888487219810486, + 0.09084011614322662, + 0.04093386232852936, + 0.0479695163667202, + 0.024763813242316246, + 0.01996094547212124, + 0.11194301396608353, + 0.030978573486208916, + 0.09638101607561111 + ], + "credit_map": [ + 0.028119219467043877, + 0.07037465274333954, + 0.07582642883062363, + 0.029071148484945297, + 0.03470362722873688, + 0.05792618915438652, + 0.03788399696350098, + 0.06343898177146912, + 0.008706590160727501, + 0.018189596012234688, + 0.02174931764602661, + 0.010452632792294025, + 0.020329786464571953, + 0.029635664075613022, + 0.01347708236426115, + 0.01634419895708561, + 0.008479917421936989, + 0.012245719321072102, + 0.013999136164784431, + 0.01363708358258009, + 0.00939906481653452, + 0.010542221367359161, + 0.010532879270613194, + 0.012004096992313862, + 0.0073862867429852486, + 0.005489813629537821, + 0.007247183471918106, + 0.0038819522596895695, + 0.004323981702327728, + 0.004109691362828016, + 0.0035371165722608566, + 0.004957834258675575, + 0.005060572177171707, + 0.005525085609406233, + 0.01085664238780737, + 0.0044535682536661625, + 0.003282589139416814, + 0.005234439857304096, + 0.0037360566202551126, + 0.009820563718676567, + 0.0023526358418166637, + 0.0037744916044175625, + 0.0043206592090427876, + 0.0028977319598197937, + 0.0028850766830146313, + 0.0031061710324138403, + 0.00271555595099926, + 0.002711489563807845, + 0.002251379657536745, + 0.0034811843652278185, + 0.0030072294175624847, + 0.0022310959175229073, + 0.00186321756336838, + 0.002587274182587862, + 0.0018852222710847855, + 0.0026543422136455774, + 0.01221064105629921, + 0.023610271513462067, + 0.019740324467420578, + 0.011223914101719856, + 0.00892342533916235, + 0.011772162280976772, + 0.008479183539748192, + 0.01598309352993965, + 0.004979295190423727, + 0.004748424980789423, + 0.004624555353075266, + 0.0028060348704457283, + 0.002711772220209241, + 0.004196968395262957, + 0.0025293442886322737, + 0.004382177721709013, + 0.009293684735894203, + 0.012840881012380123, + 0.015338716097176075, + 0.010722263716161251, + 0.008879967033863068, + 0.01738174632191658, + 0.008902900852262974, + 0.01302085630595684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.03371411934494972 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.006732044741511345 + ], + "opposite_logprob_tokens": [ + -5.006731986999512 + ], + "logprob_reward": -0.006732044741511345, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732044741511345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15544727444648743, + 0.20908060669898987, + 0.11215568333864212, + 0.11283224076032639, + 0.053774770349264145, + 0.04411329701542854, + 0.057399068027734756, + 0.10628464072942734, + 0.03473881632089615, + 0.11417359858751297 + ], + "credit_map": [ + 0.009595124050974846, + 0.021851204335689545, + 0.028974492102861404, + 0.010745092295110226, + 0.014825155027210712, + 0.027495404705405235, + 0.019988687708973885, + 0.021972112357616425, + 0.051109883934259415, + 0.029098667204380035, + 0.039072178304195404, + 0.013673419132828712, + 0.012898681685328484, + 0.022380640730261803, + 0.013677526265382767, + 0.0271696075797081, + 0.023171180859208107, + 0.016597174108028412, + 0.018838023766875267, + 0.00825346540659666, + 0.009798482991755009, + 0.013270207680761814, + 0.00944153405725956, + 0.012785609811544418, + 0.016738377511501312, + 0.01373807992786169, + 0.027459846809506416, + 0.006360836327075958, + 0.0075621879659593105, + 0.010803662240505219, + 0.006593276746571064, + 0.02357597090303898, + 0.005684759002178907, + 0.007725834846496582, + 0.012463563121855259, + 0.0035857539623975754, + 0.004560616798698902, + 0.005927164573222399, + 0.004353728611022234, + 0.009473348036408424, + 0.0055561070330441, + 0.00816062930971384, + 0.008004647679626942, + 0.0034085651859641075, + 0.0036565344780683517, + 0.005864829756319523, + 0.0037568495608866215, + 0.005705134477466345, + 0.006593332625925541, + 0.008337296545505524, + 0.008786618709564209, + 0.005983917973935604, + 0.006245245225727558, + 0.007916503585875034, + 0.005286260508000851, + 0.00824989378452301, + 0.012229754589498043, + 0.020403249189257622, + 0.01962030678987503, + 0.011825633235275745, + 0.008953370153903961, + 0.008886964060366154, + 0.009524182416498661, + 0.014841184951364994, + 0.004403550177812576, + 0.006311499979346991, + 0.00614359974861145, + 0.0028113771695643663, + 0.0024592126719653606, + 0.004910854157060385, + 0.0026601527351886034, + 0.00503856735303998, + 0.020946374163031578, + 0.017329320311546326, + 0.019008567556738853, + 0.009332134388387203, + 0.006134672090411186, + 0.01754799671471119, + 0.005911869462579489, + 0.017962664365768433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.006732044741511345 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.31344518065452576 + ], + "opposite_logprob_tokens": [ + -1.3134452104568481 + ], + "logprob_reward": -0.31344518065452576, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31344518065452576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4682895243167877, + 0.1558198481798172, + 0.07616005837917328, + 0.04957978427410126, + 0.02153753489255905, + 0.03687877953052521, + 0.046652473509311676, + 0.05324995890259743, + 0.03694210574030876, + 0.054889943450689316 + ], + "credit_map": [ + 0.03170555830001831, + 0.054361436516046524, + 0.07736559212207794, + 0.0463155172765255, + 0.050089072436094284, + 0.07253177464008331, + 0.06099972128868103, + 0.07492084056138992, + 0.016496604308485985, + 0.016532624140381813, + 0.021708302199840546, + 0.018412228673696518, + 0.015146243385970592, + 0.027908945456147194, + 0.01898990198969841, + 0.020624998956918716, + 0.005301810335367918, + 0.009959930554032326, + 0.01182902604341507, + 0.01090407557785511, + 0.008275779895484447, + 0.01062594261020422, + 0.009172790683805943, + 0.010090700350701809, + 0.004324944689869881, + 0.007491639815270901, + 0.010027624666690826, + 0.005194069817662239, + 0.0037778159603476524, + 0.004567169118672609, + 0.00386875425465405, + 0.010327767580747604, + 0.002474461914971471, + 0.003833241993561387, + 0.00327169056981802, + 0.002053244039416313, + 0.00204627588391304, + 0.002642832463607192, + 0.0020842356607317924, + 0.0031315519008785486, + 0.0029129297472536564, + 0.004901973996311426, + 0.006767468061298132, + 0.003421943401917815, + 0.003857237519696355, + 0.005179268773645163, + 0.004427279345691204, + 0.005410678684711456, + 0.003344964701682329, + 0.010626494884490967, + 0.0077304174192249775, + 0.004109344910830259, + 0.00447920523583889, + 0.00598652008920908, + 0.0038973181508481503, + 0.006478206720203161, + 0.003011946566402912, + 0.008616209961473942, + 0.010921156965196133, + 0.005439544562250376, + 0.004347455222159624, + 0.007267822977155447, + 0.005972795654088259, + 0.007673026993870735, + 0.003431282239034772, + 0.0060274298302829266, + 0.007084263954311609, + 0.004364424385130405, + 0.0029096698854118586, + 0.005432422738522291, + 0.0027343672700226307, + 0.004958245903253555, + 0.007085033692419529, + 0.00821444857865572, + 0.007677555084228516, + 0.0063829622231423855, + 0.00460832379758358, + 0.007310246117413044, + 0.004478867165744305, + 0.009132507257163525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.31344518065452576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition, eye-level still shot captures the scene in a minimalist ink-wash style, subtly perceiving the delicate fibrous texture of the underlying rice paper. A long, serpentine, catfish-like creature rests motionless on the pale, sandy riverbed, its body conforming to the soft contours of the ground. The shot's intense detail renders subtle gradations of gray and deeply saturated black across its smooth skin with artistic precision, while its long, thin whiskers, delicately barbed, drift almost imperceptibly. Nearby, a sinuous seagrass reed, resembling bold calligraphy strokes, sways gently in an invisible current. A subtle, intermittent stream of small bubbles escapes the creature's gills, drifting slowly and deliberately upwards to disappear into the pale, empty water above, creating an atmosphere of tranquil and contemplative solitude.", + "seed": 12891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12891_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": -0.054609728418290614, + "vlm_logprob": -0.054609728418290614, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.054609728418290614, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.01046267244964838 + ], + "opposite_logprob_tokens": [ + -4.885462760925293 + ], + "logprob_reward": -0.01046267244964838, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01046267244964838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3890426456928253, + 0.07890905439853668, + 0.12237722426652908, + 0.07378647476434708, + 0.061078738421201706, + 0.08433173596858978, + 0.059747468680143356, + 0.06487486511468887, + 0.026042619720101357, + 0.039809174835681915 + ], + "credit_map": [ + 0.12243711203336716, + 0.024836493656039238, + 0.01260228082537651, + 0.017386315390467644, + 0.12374341487884521, + 0.024114400148391724, + 0.04362506791949272, + 0.02029760554432869, + 0.012079847976565361, + 0.009289288893342018, + 0.006896484177559614, + 0.005982293747365475, + 0.02022506482899189, + 0.006277861073613167, + 0.012413942255079746, + 0.005744272843003273, + 0.01679321750998497, + 0.013733108527958393, + 0.008421335369348526, + 0.009427565149962902, + 0.01629473827779293, + 0.020973896607756615, + 0.013935767114162445, + 0.0227976031601429, + 0.011832754127681255, + 0.008915909565985203, + 0.005532681476324797, + 0.004421889781951904, + 0.011889423243701458, + 0.015261086635291576, + 0.008003287948668003, + 0.007929438725113869, + 0.008582072332501411, + 0.009084425866603851, + 0.00469674589112401, + 0.006857512053102255, + 0.0061492216773331165, + 0.010562320239841938, + 0.006388024426996708, + 0.008758415468037128, + 0.011971273459494114, + 0.011827274225652218, + 0.008680266328155994, + 0.009683645330369473, + 0.01562902145087719, + 0.00813713762909174, + 0.009333969093859196, + 0.009069150313735008, + 0.009104469791054726, + 0.011452767066657543, + 0.004549081437289715, + 0.006951699499040842, + 0.006337113678455353, + 0.007142973132431507, + 0.008242027834057808, + 0.005967338569462299, + 0.009884982369840145, + 0.011020113714039326, + 0.007168847136199474, + 0.006718512158840895, + 0.009052809327840805, + 0.007957673631608486, + 0.006376700010150671, + 0.00669523561373353, + 0.0034595762845128775, + 0.0033218434546142817, + 0.0029691059608012438, + 0.003065563039854169, + 0.004234643187373877, + 0.0027800307143479586, + 0.002492228988558054, + 0.0037196308840066195, + 0.004809923470020294, + 0.004479904659092426, + 0.004635154735296965, + 0.005435402970761061, + 0.005178939551115036, + 0.004307988565415144, + 0.005119881127029657, + 0.0058419834822416306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01046267244964838 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.006740570068359375 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740570068359375, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740570068359375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35595157742500305, + 0.11550801247358322, + 0.06149570271372795, + 0.03738374635577202, + 0.04333491250872612, + 0.0164396483451128, + 0.02504798024892807, + 0.2711534798145294, + 0.02581842988729477, + 0.04786648973822594 + ], + "credit_map": [ + 0.11690983176231384, + 0.017672281712293625, + 0.011530105024576187, + 0.016437096521258354, + 0.11187361180782318, + 0.027994146570563316, + 0.023988433182239532, + 0.02954607643187046, + 0.016082730144262314, + 0.01327220443636179, + 0.012102736160159111, + 0.010789629071950912, + 0.026557888835668564, + 0.008609052747488022, + 0.018762515857815742, + 0.009331257082521915, + 0.008026125840842724, + 0.009830975905060768, + 0.004786009434610605, + 0.006436490919440985, + 0.011966479010879993, + 0.006640227977186441, + 0.007602103520184755, + 0.006207289639860392, + 0.004999489523470402, + 0.004194706212729216, + 0.004678212106227875, + 0.005221480969339609, + 0.005020579788833857, + 0.004402216523885727, + 0.0042783347889781, + 0.004588725045323372, + 0.005816672462970018, + 0.008360330015420914, + 0.0036343843676149845, + 0.005356901325285435, + 0.008526839315891266, + 0.0038719058502465487, + 0.00402502715587616, + 0.0037428527139127254, + 0.0022123854141682386, + 0.0021945976186543703, + 0.0018998433370143175, + 0.0019607795402407646, + 0.0018128666561096907, + 0.0022233747877180576, + 0.002004687674343586, + 0.002131113549694419, + 0.0034785238094627857, + 0.003330804640427232, + 0.0024334087502211332, + 0.003743324661627412, + 0.0038402832578867674, + 0.0027407535817474127, + 0.003062897128984332, + 0.0024179851170629263, + 0.0399906188249588, + 0.04827776923775673, + 0.02568376623094082, + 0.029272383078932762, + 0.03900810703635216, + 0.03491067513823509, + 0.02859809808433056, + 0.025412065908312798, + 0.003970613703131676, + 0.003378107910975814, + 0.002872998593375087, + 0.003911019768565893, + 0.0048006256110966206, + 0.002416437491774559, + 0.0025833642575889826, + 0.0018852625507861376, + 0.0067913890816271305, + 0.005925927311182022, + 0.007193003315478563, + 0.007146353833377361, + 0.0052274554036557674, + 0.00565671781077981, + 0.005241758190095425, + 0.004683884792029858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.006740570068359375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.03372311219573021 + ], + "opposite_logprob_tokens": [ + -3.4087231159210205 + ], + "logprob_reward": -0.03372311219573021, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372311219573021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41961297392845154, + 0.07036804407835007, + 0.09492503106594086, + 0.03241392597556114, + 0.10112737119197845, + 0.046290215104818344, + 0.029928581789135933, + 0.11921069025993347, + 0.028343306854367256, + 0.05777982994914055 + ], + "credit_map": [ + 0.1335924118757248, + 0.034080006182193756, + 0.017870668321847916, + 0.02517174743115902, + 0.13211312890052795, + 0.021118834614753723, + 0.03895873576402664, + 0.016707422211766243, + 0.01043244544416666, + 0.007002434693276882, + 0.0061961072497069836, + 0.006524601485580206, + 0.015508344396948814, + 0.007100289687514305, + 0.010542088188230991, + 0.0070617301389575005, + 0.011632587760686874, + 0.01505363080650568, + 0.008837558329105377, + 0.011679808609187603, + 0.01125331036746502, + 0.013057054951786995, + 0.010107289999723434, + 0.013303785584867, + 0.0044608511961996555, + 0.005912903230637312, + 0.003383234841749072, + 0.0036713453009724617, + 0.003013032954186201, + 0.004364720080047846, + 0.0039731590077281, + 0.0036346784327179193, + 0.016652965918183327, + 0.01912730187177658, + 0.009034601971507072, + 0.010581265203654766, + 0.01884113997220993, + 0.009476786479353905, + 0.008573411963880062, + 0.008839893154799938, + 0.007849134504795074, + 0.006229248363524675, + 0.0042531960643827915, + 0.005190291907638311, + 0.008806059136986732, + 0.004671729169785976, + 0.005080242175608873, + 0.0042103128507733345, + 0.004969452507793903, + 0.003936248365789652, + 0.0022842567414045334, + 0.0035335482098162174, + 0.005559946410357952, + 0.002937830053269863, + 0.0033333012834191322, + 0.0033739996142685413, + 0.01838425174355507, + 0.010940096341073513, + 0.010040263645350933, + 0.013861759565770626, + 0.03162505850195885, + 0.01014629378914833, + 0.015354519709944725, + 0.008858450688421726, + 0.0038727722130715847, + 0.003168179653584957, + 0.002630152041092515, + 0.0029780028853565454, + 0.006133875343948603, + 0.0029848935082554817, + 0.0029248553328216076, + 0.003650578437373042, + 0.007050617597997189, + 0.006402913480997086, + 0.0071564302779734135, + 0.0073354304768145084, + 0.010105744004249573, + 0.0060692280530929565, + 0.0063695465214550495, + 0.007289920933544636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.03372311219573021 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.020571446046233177 + ], + "opposite_logprob_tokens": [ + -3.89557147026062 + ], + "logprob_reward": -0.020571446046233177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571446046233177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4916650950908661, + 0.09735064953565598, + 0.10804622620344162, + 0.022985368967056274, + 0.022926324978470802, + 0.0645281970500946, + 0.032408930361270905, + 0.07912381738424301, + 0.030307674780488014, + 0.0506577268242836 + ], + "credit_map": [ + 0.18146483600139618, + 0.027396628633141518, + 0.01713661476969719, + 0.015640994533896446, + 0.17410334944725037, + 0.020780857652425766, + 0.03990014269948006, + 0.015241701155900955, + 0.012788982130587101, + 0.010472816415131092, + 0.010023060254752636, + 0.009217729791998863, + 0.020163187757134438, + 0.008506176993250847, + 0.015364793129265308, + 0.010813901200890541, + 0.015033978037536144, + 0.009159802459180355, + 0.010526662692427635, + 0.013006635941565037, + 0.020306915044784546, + 0.016303984448313713, + 0.01134865265339613, + 0.012359588406980038, + 0.004099337384104729, + 0.0031000343151390553, + 0.002069612266495824, + 0.0020226638298481703, + 0.003610044252127409, + 0.002179231960326433, + 0.003060533432289958, + 0.00284391175955534, + 0.003810987574979663, + 0.0038905821274966, + 0.0019008155213668942, + 0.0017460978124290705, + 0.004151309374719858, + 0.0028380807489156723, + 0.0021388453897088766, + 0.0024496077094227076, + 0.00861657876521349, + 0.008064900524914265, + 0.007791354786604643, + 0.010755280032753944, + 0.010141515173017979, + 0.005068621598184109, + 0.006968831177800894, + 0.007121114060282707, + 0.004362485837191343, + 0.004584202077239752, + 0.003014859976246953, + 0.004974160809069872, + 0.00415925495326519, + 0.003062603296712041, + 0.0038154313806444407, + 0.004435933195054531, + 0.011498554609715939, + 0.011928795836865902, + 0.010718721896409988, + 0.009434903040528297, + 0.009528824128210545, + 0.010127222165465355, + 0.008861019276082516, + 0.007025776896625757, + 0.004449052270501852, + 0.004784174729138613, + 0.002762516960501671, + 0.003395438427105546, + 0.003462559776380658, + 0.0037847114726901054, + 0.0030014747753739357, + 0.004667744971811771, + 0.006813135463744402, + 0.008920606225728989, + 0.004026159178465605, + 0.006902998313307762, + 0.005936419125646353, + 0.005290171597152948, + 0.007599886041134596, + 0.0051683527417480946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.020571446046233177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.20155084133148193 + ], + "opposite_logprob_tokens": [ + -1.701550841331482 + ], + "logprob_reward": -0.20155084133148193, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20155084133148193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46021977066993713, + 0.03978473320603371, + 0.05165155231952667, + 0.04656313732266426, + 0.0352632962167263, + 0.03169548138976097, + 0.03338485583662987, + 0.1545507162809372, + 0.036015044897794724, + 0.11087147146463394 + ], + "credit_map": [ + 0.1561250388622284, + 0.029806142672896385, + 0.01796470209956169, + 0.016637835651636124, + 0.15471555292606354, + 0.026365358382463455, + 0.03809845447540283, + 0.020506639033555984, + 0.005394272040575743, + 0.003986682742834091, + 0.0038069712463766336, + 0.004562194459140301, + 0.006521480157971382, + 0.005085048731416464, + 0.0056021204218268394, + 0.004825963173061609, + 0.006683844141662121, + 0.007442235480993986, + 0.0040291245095431805, + 0.005084939766675234, + 0.008629309013485909, + 0.0059913406148552895, + 0.007672721520066261, + 0.0061180368065834045, + 0.006261132191866636, + 0.006468754727393389, + 0.006001265719532967, + 0.006353961769491434, + 0.00647720554843545, + 0.005532555282115936, + 0.004067554138600826, + 0.005400711204856634, + 0.006377377547323704, + 0.004487167112529278, + 0.003268985776230693, + 0.0038345549255609512, + 0.006038264837116003, + 0.004960708320140839, + 0.0031593504827469587, + 0.0031368876807391644, + 0.0050115021876990795, + 0.004901457112282515, + 0.0035964210983365774, + 0.003931398969143629, + 0.0036965366452932358, + 0.004373244941234589, + 0.0032041328959167004, + 0.00298078591004014, + 0.006012713070958853, + 0.006231832318007946, + 0.002957077929750085, + 0.0026623979210853577, + 0.003481500083580613, + 0.004510891158133745, + 0.003802960505709052, + 0.0037254823837429285, + 0.02190590091049671, + 0.02415643259882927, + 0.01762562431395054, + 0.01638520509004593, + 0.02141273021697998, + 0.021384045481681824, + 0.018733832985162735, + 0.012946927919983864, + 0.0049666459672153, + 0.0048048351891338825, + 0.003707133000716567, + 0.0036571745295077562, + 0.006172615103423595, + 0.0036033636424690485, + 0.003977471496909857, + 0.005125805269926786, + 0.014758610166609287, + 0.019652653485536575, + 0.007832484319806099, + 0.013162638992071152, + 0.010341391898691654, + 0.014727937988936901, + 0.017865432426333427, + 0.012530319392681122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.20155084133148193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-definition, eye-level still shot captures the scene in a minimalist ink-wash style, subtly perceiving the delicate fibrous texture of the underlying rice paper. A long, serpentine, catfish-like creature rests motionless on the pale, sandy riverbed, its body conforming to the soft contours of the ground. The shot's intense detail renders subtle gradations of gray and deeply saturated black across its smooth skin with artistic precision, while its long, thin whiskers, delicately barbed, drift almost imperceptibly. Nearby, a sinuous seagrass reed, resembling bold calligraphy strokes, sways gently in an invisible current. A subtle, intermittent stream of small bubbles escapes the creature's gills, drifting slowly and deliberately upwards to disappear into the pale, empty water above, creating an atmosphere of tranquil and contemplative solitude.", + "seed": 12892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_10/index_7997_seed_12892_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": -0.06556731974706054, + "vlm_logprob": -0.06556731974706054, + "vlm_yes_no_margin": 3.674999952316284, + "vlm_token_credit": -0.06556731974706054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.0153641477227211 + ], + "opposite_logprob_tokens": [ + -4.390364170074463 + ], + "logprob_reward": -0.0153641477227211, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0153641477227211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6295230388641357, + 0.0800987258553505, + 0.03505121171474457, + 0.015559850260615349, + 0.018340814858675003, + 0.03013823926448822, + 0.0583631694316864, + 0.012418593280017376, + 0.03227684646844864, + 0.08822955936193466 + ], + "credit_map": [ + 0.24195004999637604, + 0.018842916935682297, + 0.033522944897413254, + 0.015549728646874428, + 0.01805427297949791, + 0.27483317255973816, + 0.013288816437125206, + 0.01348115038126707, + 0.017210165038704872, + 0.01365124061703682, + 0.009455372579395771, + 0.007387990597635508, + 0.007130485028028488, + 0.014006953686475754, + 0.005304152611643076, + 0.0059523615054786205, + 0.0076836044900119305, + 0.0060965088196098804, + 0.003804450388997793, + 0.003369824727997184, + 0.003995850216597319, + 0.0048406063579022884, + 0.0022738485131412745, + 0.0029865128453820944, + 0.0027008692268282175, + 0.002975016599521041, + 0.0022492639254778624, + 0.001834800816141069, + 0.0017675607232376933, + 0.0018934148829430342, + 0.0009320217068307102, + 0.0012069015065208077, + 0.0028501825872808695, + 0.002971579786390066, + 0.0020904801785945892, + 0.0024224359076470137, + 0.0022166993003338575, + 0.002240185160189867, + 0.0016510796267539263, + 0.001898171380162239, + 0.004937927704304457, + 0.004009296651929617, + 0.0045166113413870335, + 0.00590381259098649, + 0.0026096163783222437, + 0.0039423503912985325, + 0.002220523776486516, + 0.0019980978686362505, + 0.011320424266159534, + 0.009545804001390934, + 0.007854466326534748, + 0.009949946776032448, + 0.004581957124173641, + 0.007579849101603031, + 0.003488184418529272, + 0.004042534157633781, + 0.0017621315782889724, + 0.0015920181758701801, + 0.0019489765400066972, + 0.0018667408730834723, + 0.00126650498714298, + 0.0014731524279341102, + 0.0011632763780653477, + 0.0013457913883030415, + 0.0062461132183671, + 0.005696630105376244, + 0.00591156305745244, + 0.0031658431980758905, + 0.0028153513558208942, + 0.00461532361805439, + 0.0018857072573155165, + 0.0019403137266635895, + 0.012756568379700184, + 0.00895627774298191, + 0.014585761353373528, + 0.01873430795967579, + 0.005605940707027912, + 0.012185359373688698, + 0.008007255382835865, + 0.007398083806037903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0153641477227211 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.583602249622345, + 0.0576377809047699, + 0.057288479059934616, + 0.036794230341911316, + 0.03675829619169235, + 0.07474105805158615, + 0.045811865478754044, + 0.04002295434474945, + 0.02801625430583954, + 0.03932679817080498 + ], + "credit_map": [ + 0.19388915598392487, + 0.0411389097571373, + 0.04394315555691719, + 0.02351776883006096, + 0.02820284105837345, + 0.21608005464076996, + 0.019483154639601707, + 0.01734718680381775, + 0.009212908335030079, + 0.00869798008352518, + 0.007468073628842831, + 0.006515327841043472, + 0.007524306420236826, + 0.008142326958477497, + 0.004965789616107941, + 0.005111071281135082, + 0.010437238030135632, + 0.008526092395186424, + 0.006808008998632431, + 0.008414272218942642, + 0.006242269184440374, + 0.0070028165355324745, + 0.004032602999359369, + 0.005825181491672993, + 0.006477581337094307, + 0.004263092298060656, + 0.0060351453721523285, + 0.004477438051253557, + 0.004409404471516609, + 0.004464824218302965, + 0.0032510655000805855, + 0.003415678394958377, + 0.0075368573889136314, + 0.007604571059346199, + 0.003968002740293741, + 0.002440810902044177, + 0.004354554228484631, + 0.005073212552815676, + 0.002703365171328187, + 0.0030769214499741793, + 0.01075526513159275, + 0.010632914490997791, + 0.011028433218598366, + 0.014421341009438038, + 0.0075942291878163815, + 0.009264044463634491, + 0.005770863965153694, + 0.005273961462080479, + 0.005495685152709484, + 0.005637264810502529, + 0.00614930409938097, + 0.009280547499656677, + 0.00539801362901926, + 0.005146455019712448, + 0.004552037920802832, + 0.004152555484324694, + 0.007193719036877155, + 0.009168296121060848, + 0.0032424121163785458, + 0.0028476433362811804, + 0.004681540187448263, + 0.005246142856776714, + 0.0034443992190063, + 0.004198805894702673, + 0.004866653587669134, + 0.00435621477663517, + 0.0032632285729050636, + 0.003976304084062576, + 0.0024092495441436768, + 0.004088209476321936, + 0.0024756949860602617, + 0.002580697415396571, + 0.006428660359233618, + 0.006645404268056154, + 0.00492745591327548, + 0.006652803625911474, + 0.0027348811272531748, + 0.0061321621760725975, + 0.0032926613930612803, + 0.0025127704720944166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.03811635076999664 + ], + "opposite_logprob_tokens": [ + -3.288116455078125 + ], + "logprob_reward": -0.03811635076999664, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811635076999664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6813519597053528, + 0.05431710556149483, + 0.025236701592803, + 0.01692913845181465, + 0.04587089270353317, + 0.0784124881029129, + 0.022030919790267944, + 0.02771010622382164, + 0.019657252356410027, + 0.028483394533395767 + ], + "credit_map": [ + 0.23251745104789734, + 0.054003432393074036, + 0.03133287653326988, + 0.01948343589901924, + 0.029078694060444832, + 0.2640089988708496, + 0.02975318394601345, + 0.021173907443881035, + 0.01000868808478117, + 0.010223763063549995, + 0.0065687247551977634, + 0.0045813885517418385, + 0.005002669990062714, + 0.00863178912550211, + 0.0041533890180289745, + 0.0051466915756464005, + 0.0043609868735075, + 0.004103874322026968, + 0.002957709599286318, + 0.0022686158772557974, + 0.004238397814333439, + 0.0027275909669697285, + 0.0023343446664512157, + 0.0022451819386333227, + 0.003199075348675251, + 0.002370839472860098, + 0.002557760803028941, + 0.001758734812028706, + 0.0016589869046583772, + 0.0022398075088858604, + 0.0013961291406303644, + 0.0017478048102930188, + 0.007849017158150673, + 0.008585573174059391, + 0.0049605704843997955, + 0.006280680187046528, + 0.006220924202352762, + 0.005288875661790371, + 0.003437632229179144, + 0.003247619839385152, + 0.01258684229105711, + 0.010865791700780392, + 0.011704095639288425, + 0.015876419842243195, + 0.006441518664360046, + 0.010278631001710892, + 0.005895402282476425, + 0.004763783887028694, + 0.0035554259084165096, + 0.0026960154063999653, + 0.0043127816170454025, + 0.002664626110345125, + 0.0022967555560171604, + 0.002789811696857214, + 0.0015687041450291872, + 0.002146800048649311, + 0.003913520835340023, + 0.0042571816593408585, + 0.0032247169874608517, + 0.0036921738646924496, + 0.003565695136785507, + 0.003441124688833952, + 0.0030376482754945755, + 0.002578045940026641, + 0.003861576085910201, + 0.002295894781127572, + 0.002449083374813199, + 0.001967791235074401, + 0.0031704544089734554, + 0.002550715347751975, + 0.002026170492172241, + 0.0013355660485103726, + 0.005252298898994923, + 0.0034567073453217745, + 0.0041196634992957115, + 0.004690028727054596, + 0.002658510347828269, + 0.003830807050690055, + 0.0022048011887818575, + 0.002270578173920512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03811635076999664 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.018179109320044518 + ], + "opposite_logprob_tokens": [ + -4.018178939819336 + ], + "logprob_reward": -0.018179109320044518, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179109320044518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46903228759765625, + 0.03845268860459328, + 0.0662953108549118, + 0.03763674944639206, + 0.11536125838756561, + 0.19478340446949005, + 0.009728735312819481, + 0.02240092121064663, + 0.019817104563117027, + 0.026491528376936913 + ], + "credit_map": [ + 0.1609078049659729, + 0.022239843383431435, + 0.0332954078912735, + 0.013094449415802956, + 0.026961905881762505, + 0.1854199767112732, + 0.015353234484791756, + 0.011759662069380283, + 0.008570917882025242, + 0.00494063924998045, + 0.004415604285895824, + 0.003522993065416813, + 0.003933250438421965, + 0.007248698268085718, + 0.003439044114202261, + 0.0023815419990569353, + 0.014579799957573414, + 0.012486490420997143, + 0.008108293637633324, + 0.006002366542816162, + 0.00719072250649333, + 0.00903808232396841, + 0.0040168315172195435, + 0.004872726276516914, + 0.0073750559240579605, + 0.0036155960988253355, + 0.006546509452164173, + 0.005780537147074938, + 0.0037500103935599327, + 0.004736282862722874, + 0.0028260902035981417, + 0.00300667155534029, + 0.019396917894482613, + 0.0179852657020092, + 0.015897659584879875, + 0.01837204210460186, + 0.012033873237669468, + 0.01418239250779152, + 0.007681521121412516, + 0.009811598807573318, + 0.02552451379597187, + 0.027674023061990738, + 0.029668742790818214, + 0.03684729337692261, + 0.02288863994181156, + 0.02229064516723156, + 0.015133187174797058, + 0.014756366610527039, + 0.001831564586609602, + 0.0011480869725346565, + 0.0018636068562045693, + 0.0010158902732655406, + 0.0010371507378295064, + 0.0012974523706361651, + 0.0008851392194628716, + 0.0006498450529761612, + 0.0031984015367925167, + 0.0030701628420501947, + 0.002966134110465646, + 0.003351053223013878, + 0.002734209643676877, + 0.0023847618140280247, + 0.002398357493802905, + 0.0022978410124778748, + 0.002642458537593484, + 0.0024984232150018215, + 0.002280957531183958, + 0.002885267371311784, + 0.0030660596676170826, + 0.0027282291557639837, + 0.0025306676980108023, + 0.0011850440641865134, + 0.004047044087201357, + 0.0033032000064849854, + 0.0036950083449482918, + 0.003808171022683382, + 0.003761985572054982, + 0.003648428013548255, + 0.0026220595464110374, + 0.0016056335298344493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.018179109320044518 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a minimalist ink-wash style with a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a long, serpentine, catfish-like creature visible resting on a pale, sandy riverbed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the creature have long, thin, delicately barbed whiskers that drift almost imperceptibly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a sinuous seagrass reed resembling bold calligraphy strokes swaying gently nearby?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are small bubbles escaping the creature's gills and drifting slowly upwards into the 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": [ + -0.25207579135894775 + ], + "opposite_logprob_tokens": [ + -1.5020757913589478 + ], + "logprob_reward": -0.25207579135894775, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207579135894775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8577138185501099, + 0.027144184336066246, + 0.010327299125492573, + 0.009698000736534595, + 0.020885780453681946, + 0.01814495399594307, + 0.01215605903416872, + 0.013484467752277851, + 0.016048679128289223, + 0.01439672987908125 + ], + "credit_map": [ + 0.31598958373069763, + 0.032875869423151016, + 0.04663524031639099, + 0.026049500331282616, + 0.034209027886390686, + 0.35585105419158936, + 0.02440526708960533, + 0.021698348224163055, + 0.004986566491425037, + 0.0054217055439949036, + 0.0028177695348858833, + 0.002378688892349601, + 0.0030542665626853704, + 0.004576222971081734, + 0.0016428291564807296, + 0.002266138093546033, + 0.002085144631564617, + 0.002040242310613394, + 0.0012423978187143803, + 0.0008579965797252953, + 0.0010931766591966152, + 0.0014562319265678525, + 0.0007468278054147959, + 0.0008052836055867374, + 0.0018734857439994812, + 0.0019932836294174194, + 0.0011287984671071172, + 0.0009045213810168207, + 0.0008507244056090713, + 0.0013010252732783556, + 0.0008368901908397675, + 0.0008092722855508327, + 0.0032777662854641676, + 0.003353097476065159, + 0.002696842886507511, + 0.001738835941068828, + 0.0026619115378707647, + 0.0025717192329466343, + 0.001804847619496286, + 0.0027807613369077444, + 0.004047074355185032, + 0.0016776176635175943, + 0.002400260651484132, + 0.002229362726211548, + 0.0023022317327558994, + 0.002634827047586441, + 0.001649970654398203, + 0.0012036093976348639, + 0.002009085612371564, + 0.0025317256804555655, + 0.0014555169036611915, + 0.0014323227806016803, + 0.0010243243305012584, + 0.0014411702286452055, + 0.0011348556727170944, + 0.001127059105783701, + 0.002195229521021247, + 0.002170866820961237, + 0.0015848982147872448, + 0.0018397116800770164, + 0.0017799681518226862, + 0.0015314907068386674, + 0.0013514363672584295, + 0.001030866289511323, + 0.0027812174521386623, + 0.002484179800376296, + 0.0016438732855021954, + 0.0016016244189813733, + 0.00245343754068017, + 0.0024863393045961857, + 0.0012508986983448267, + 0.001347110839560628, + 0.003347863210365176, + 0.001302366959862411, + 0.002354410942643881, + 0.0013532928423956037, + 0.0017372521106153727, + 0.0020651277154684067, + 0.0010399633320048451, + 0.0011964542791247368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.25207579135894775 + } + ], + "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/0000158/group_11/index_tensor([6353, 6353])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_tensor([6353, 6353])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6212e0e5083f8a5a28e3282df0b52f4ffa94f98 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_tensor([6353, 6353])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, smooth, 360-degree orbiting shot places the audience in an absolutely pitch-black, boundless void. An intense, concentrated, dramatic overhead light, like a stage spotlight, pierces this darkness vertically, creating a distinct, volumetric column of light. Within this column, countless tiny golden particles slowly drift. Rather than simply plummeting, they gracefully swirl and float through the air, as if drawn by an invisible breeze, creating a silent, stardust-like ballet. Each particle resembles a miniature diamond, varying in size and brightness. As they pass through the beam, their faceted structures capture and refract the dazzling light, creating a fleeting, dazzling starburst effect and subtle rainbow dispersions before the camera. As the camera rotates, the relationship between light and shadow shifts, with some particles illuminating momentarily upon entering the column and fading quietly into darkness upon exiting, creating a perpetually dynamic aesthetic. The main colors of the entire picture are deep black and warm and dazzling gold, creating a magical atmosphere that is mysterious, grand, peaceful and dreamy, perfectly capturing this cosmic dance carefully woven by light and shadow.", + "seed": 8899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_6353_seed_8899_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": -0.5337566248141229, + "vlm_logprob": -0.5337566248141229, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.5337566248141229, + "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 background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6020499467849731 + ], + "opposite_logprob_tokens": [ + -0.22704997658729553 + ], + "logprob_reward": -1.6020499467849731, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6020499467849731, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08103229850530624, + 0.022862814366817474, + 0.02165524661540985, + 0.01219909731298685, + 0.28155288100242615, + 0.2721921503543854, + 0.16553442180156708, + 0.079984150826931, + 0.03510797768831253, + 0.027879001572728157 + ], + "credit_map": [ + 0.011305205523967743, + 0.004873615689575672, + 0.013518470339477062, + 0.013797655701637268, + 0.00422037485986948, + 0.014365503564476967, + 0.008963762782514095, + 0.009987716563045979, + 0.0025998486671596766, + 0.0019711535423994064, + 0.003781010629609227, + 0.004725033417344093, + 0.0009751545148901641, + 0.0029108943417668343, + 0.0025211500469595194, + 0.0033785689156502485, + 0.003221627790480852, + 0.002846434013918042, + 0.0023441817611455917, + 0.0020938485395163298, + 0.001835961127653718, + 0.004497549496591091, + 0.001910999184474349, + 0.0029046442359685898, + 0.0017643090104684234, + 0.0011660262243822217, + 0.001452298485673964, + 0.0014183413004502654, + 0.0013444935902953148, + 0.0021845847368240356, + 0.0013332183007150888, + 0.0015358254313468933, + 0.04930676519870758, + 0.013763359747827053, + 0.04087730124592781, + 0.010876897722482681, + 0.01972297392785549, + 0.09568209946155548, + 0.011620823293924332, + 0.03970266506075859, + 0.020321901887655258, + 0.029498951509594917, + 0.029933864250779152, + 0.013501162640750408, + 0.06432449817657471, + 0.06307221204042435, + 0.024956034496426582, + 0.026583537459373474, + 0.013325514271855354, + 0.023898029699921608, + 0.011538468301296234, + 0.016689743846654892, + 0.050382018089294434, + 0.017606038600206375, + 0.02152053639292717, + 0.010574069805443287, + 0.00839766301214695, + 0.02259909175336361, + 0.005689363460987806, + 0.00976466853171587, + 0.009899710305035114, + 0.00881931558251381, + 0.004750196821987629, + 0.010064137168228626, + 0.005280673503875732, + 0.006033188197761774, + 0.002542458940297365, + 0.0025986102409660816, + 0.0052591897547245026, + 0.007862669415771961, + 0.002160844625905156, + 0.0033703441731631756, + 0.0023256775457412004, + 0.005738923791795969, + 0.004258531145751476, + 0.0026086692232638597, + 0.005137206055223942, + 0.0026460897643119097, + 0.002797988010570407, + 0.002365916036069393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -1.6020499467849731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.004707207437604666 + ], + "opposite_logprob_tokens": [ + -5.379707336425781 + ], + "logprob_reward": -0.004707207437604666, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004707207437604666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08781834691762924, + 0.02199835143983364, + 0.018851833418011665, + 0.013097938150167465, + 0.2664000689983368, + 0.2703052759170532, + 0.1932985633611679, + 0.07176892459392548, + 0.029514659196138382, + 0.0269460566341877 + ], + "credit_map": [ + 0.011087378486990929, + 0.006235894747078419, + 0.013866886496543884, + 0.016538528725504875, + 0.004869276657700539, + 0.015145701356232166, + 0.009863855317234993, + 0.010210830718278885, + 0.003088525030761957, + 0.001754092052578926, + 0.0033623126801103354, + 0.004081460181623697, + 0.0009906451450660825, + 0.002955536125227809, + 0.0021192056592553854, + 0.0036465737503021955, + 0.002179177012294531, + 0.0028483353089541197, + 0.002423780271783471, + 0.002666839864104986, + 0.001376339583657682, + 0.0028577300254255533, + 0.0019094564486294985, + 0.00259017338976264, + 0.002063395921140909, + 0.001111767371185124, + 0.001503043924458325, + 0.001735805650241673, + 0.0012454179814085364, + 0.0020638699643313885, + 0.0014636617852374911, + 0.0019109754357486963, + 0.0473618246614933, + 0.01588192582130432, + 0.037133313715457916, + 0.010308677330613136, + 0.020141778513789177, + 0.08775131404399872, + 0.011187034659087658, + 0.036634184420108795, + 0.017406487837433815, + 0.03198035806417465, + 0.028713645413517952, + 0.014389882795512676, + 0.0693570226430893, + 0.0574093833565712, + 0.027251558378338814, + 0.023796936497092247, + 0.015093689784407616, + 0.027732590213418007, + 0.012770824134349823, + 0.01886844076216221, + 0.06036459282040596, + 0.021301213651895523, + 0.025469612330198288, + 0.011697585694491863, + 0.007467396557331085, + 0.01942490041255951, + 0.005098291207104921, + 0.00912028644233942, + 0.00924796424806118, + 0.008047047071158886, + 0.004449672996997833, + 0.008913364261388779, + 0.004308556672185659, + 0.005606019403785467, + 0.002242787042632699, + 0.002168891951441765, + 0.004506455734372139, + 0.006020167842507362, + 0.0019859338644891977, + 0.002675845054909587, + 0.0023251818493008614, + 0.0038077873177826405, + 0.005843762774020433, + 0.0023071870673447847, + 0.0043196543119847775, + 0.0028166044503450394, + 0.0028495595324784517, + 0.0026763163041323423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.004707207437604666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.00865730457007885 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.00865730457007885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865730457007885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0899462103843689, + 0.023485524579882622, + 0.020660819485783577, + 0.015609260648488998, + 0.24894382059574127, + 0.3336229622364044, + 0.1133900061249733, + 0.06561257690191269, + 0.04032738879323006, + 0.048401445150375366 + ], + "credit_map": [ + 0.011331883259117603, + 0.006799281574785709, + 0.013283665291965008, + 0.017164794728159904, + 0.005521972198039293, + 0.014721936546266079, + 0.01009690947830677, + 0.011025762185454369, + 0.0037009590305387974, + 0.001911296509206295, + 0.0034111500717699528, + 0.004415623378008604, + 0.0011077109957113862, + 0.0035779301542788744, + 0.0019497084431350231, + 0.003411145880818367, + 0.0020698413718491793, + 0.002562647918239236, + 0.0029610241763293743, + 0.003082084935158491, + 0.0014272663975134492, + 0.0030645986553281546, + 0.002178759314119816, + 0.003314596600830555, + 0.0024322376120835543, + 0.0012255863985046744, + 0.0018498171120882034, + 0.002548338146880269, + 0.0009352919296361506, + 0.0024001188576221466, + 0.0020224673207849264, + 0.0021954034455120564, + 0.04431794956326485, + 0.014555453322827816, + 0.03482721000909805, + 0.009022173471748829, + 0.018938040360808372, + 0.08259370177984238, + 0.010416064411401749, + 0.03427322208881378, + 0.021241014823317528, + 0.0400480255484581, + 0.034037575125694275, + 0.01689596101641655, + 0.08794689178466797, + 0.0700201764702797, + 0.03471076488494873, + 0.028722578659653664, + 0.010150332935154438, + 0.01534692570567131, + 0.009615642949938774, + 0.012517056427896023, + 0.03033098392188549, + 0.011495872400701046, + 0.015143156051635742, + 0.008790039457380772, + 0.006703963968902826, + 0.017929136753082275, + 0.0047484515234827995, + 0.007980921305716038, + 0.008041528053581715, + 0.007152360398322344, + 0.004101566970348358, + 0.00895464327186346, + 0.005124122370034456, + 0.004220050293952227, + 0.005954490043222904, + 0.0030493552330881357, + 0.008819730021059513, + 0.005584948230534792, + 0.003952871076762676, + 0.0036218208260834217, + 0.0039847735315561295, + 0.007014057599008083, + 0.009318524971604347, + 0.0037739318795502186, + 0.00914364866912365, + 0.004883932415395975, + 0.005495313089340925, + 0.004787263460457325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.00865730457007885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.07901433855295181 + ], + "opposite_logprob_tokens": [ + -2.579014301300049 + ], + "logprob_reward": -0.07901433855295181, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901433855295181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07073637843132019, + 0.023498516529798508, + 0.02843528240919113, + 0.027138058096170425, + 0.1783403605222702, + 0.34963205456733704, + 0.10775306820869446, + 0.08196120709180832, + 0.044761497527360916, + 0.0877436026930809 + ], + "credit_map": [ + 0.008822597563266754, + 0.00599905988201499, + 0.01215912215411663, + 0.014171145856380463, + 0.003969215322285891, + 0.011327052488923073, + 0.007730749901384115, + 0.006557441782206297, + 0.0024955656845122576, + 0.0017731317784637213, + 0.0032790207769721746, + 0.00424186373129487, + 0.001635564025491476, + 0.003022333374246955, + 0.0034764937590807676, + 0.0035745452623814344, + 0.0034940638579428196, + 0.002831347519531846, + 0.003706946736201644, + 0.0036941394209861755, + 0.002724664518609643, + 0.006077294237911701, + 0.002975320443511009, + 0.002931507071480155, + 0.0036911596544086933, + 0.0018281495431438088, + 0.003378797322511673, + 0.003751524956896901, + 0.002109544351696968, + 0.006087727379053831, + 0.0028555518947541714, + 0.0034356042742729187, + 0.030778322368860245, + 0.023116396740078926, + 0.019958650693297386, + 0.01285959780216217, + 0.017227312549948692, + 0.041167788207530975, + 0.010383199900388718, + 0.022849109023809433, + 0.032653696835041046, + 0.051029022783041, + 0.03135215491056442, + 0.021920790895819664, + 0.10870592296123505, + 0.039971984922885895, + 0.04480665177106857, + 0.019191816449165344, + 0.006886438466608524, + 0.013010957278311253, + 0.007642765995115042, + 0.009565392509102821, + 0.03228865563869476, + 0.016916826367378235, + 0.013572446070611477, + 0.00786958821117878, + 0.008597367443144321, + 0.012147968634963036, + 0.007068089209496975, + 0.010963856242597103, + 0.01595533825457096, + 0.009738565422594547, + 0.006730944383889437, + 0.010759077034890652, + 0.005767523776739836, + 0.0050356145948171616, + 0.008501045405864716, + 0.005085875280201435, + 0.006163419224321842, + 0.006257832050323486, + 0.004033568315207958, + 0.003916618414223194, + 0.008779429830610752, + 0.008602026849985123, + 0.023856699466705322, + 0.007361449766904116, + 0.011310006491839886, + 0.010155989788472652, + 0.007815766148269176, + 0.009862231090664864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.07901433855295181 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743543267250061 + ], + "opposite_logprob_tokens": [ + -0.4743543267250061 + ], + "logprob_reward": -0.9743543267250061, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743543267250061, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08125337958335876, + 0.025412162765860558, + 0.0353836864233017, + 0.04596147686243057, + 0.07008559256792068, + 0.2613774538040161, + 0.1531294733285904, + 0.1486724317073822, + 0.05739973112940788, + 0.12132465094327927 + ], + "credit_map": [ + 0.00882573053240776, + 0.005958918947726488, + 0.012964446097612381, + 0.01312550064176321, + 0.005367964506149292, + 0.012713910080492496, + 0.009524857625365257, + 0.01277204230427742, + 0.002869910793378949, + 0.0019542279187589884, + 0.003447926603257656, + 0.004952446091920137, + 0.0014958851970732212, + 0.003383166156709194, + 0.003980747889727354, + 0.0033278479240834713, + 0.005674130283296108, + 0.002118840115144849, + 0.00437391409650445, + 0.004563615657389164, + 0.002449535531923175, + 0.006705698557198048, + 0.003714370308443904, + 0.005783580243587494, + 0.0068685635924339294, + 0.002559077460318804, + 0.006592940539121628, + 0.006017325446009636, + 0.002046598121523857, + 0.011971520259976387, + 0.0041237324476242065, + 0.005781712010502815, + 0.0059852683916687965, + 0.013002884574234486, + 0.006320709362626076, + 0.008688859641551971, + 0.01173081062734127, + 0.008243842981755733, + 0.007757529616355896, + 0.008355684578418732, + 0.017121128737926483, + 0.035507120192050934, + 0.02477446012198925, + 0.015355352312326431, + 0.07735620439052582, + 0.04311313480138779, + 0.03088521957397461, + 0.017264803871512413, + 0.011715604923665524, + 0.022213902324438095, + 0.011936713941395283, + 0.017972908914089203, + 0.04306580871343613, + 0.015397332608699799, + 0.020229125395417213, + 0.010598060674965382, + 0.011929989792406559, + 0.028852812945842743, + 0.011835935525596142, + 0.014223555102944374, + 0.03780524805188179, + 0.017656825482845306, + 0.011892943643033504, + 0.014475107192993164, + 0.007397687062621117, + 0.006409807596355677, + 0.009714467450976372, + 0.006354645825922489, + 0.007622342091053724, + 0.008955605328083038, + 0.005013319663703442, + 0.0059318519197404385, + 0.013176136650145054, + 0.010189861059188843, + 0.02977997250854969, + 0.011592532508075237, + 0.013175871223211288, + 0.013230050913989544, + 0.016336360946297646, + 0.013843865133821964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.9743543267250061 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, smooth, 360-degree orbiting shot places the audience in an absolutely pitch-black, boundless void. An intense, concentrated, dramatic overhead light, like a stage spotlight, pierces this darkness vertically, creating a distinct, volumetric column of light. Within this column, countless tiny golden particles slowly drift. Rather than simply plummeting, they gracefully swirl and float through the air, as if drawn by an invisible breeze, creating a silent, stardust-like ballet. Each particle resembles a miniature diamond, varying in size and brightness. As they pass through the beam, their faceted structures capture and refract the dazzling light, creating a fleeting, dazzling starburst effect and subtle rainbow dispersions before the camera. As the camera rotates, the relationship between light and shadow shifts, with some particles illuminating momentarily upon entering the column and fading quietly into darkness upon exiting, creating a perpetually dynamic aesthetic. The main colors of the entire picture are deep black and warm and dazzling gold, creating a magical atmosphere that is mysterious, grand, peaceful and dreamy, perfectly capturing this cosmic dance carefully woven by light and shadow.", + "seed": 8900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_6353_seed_8900_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": -0.3093191974563524, + "vlm_logprob": -0.3093191974563524, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.3093191974563524, + "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 background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.3497401177883148 + ], + "opposite_logprob_tokens": [ + -1.2247401475906372 + ], + "logprob_reward": -0.3497401177883148, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3497401177883148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30457526445388794, + 0.09498655050992966, + 0.05437522754073143, + 0.06312653422355652, + 0.07147694379091263, + 0.044054508209228516, + 0.025287576019763947, + 0.04540539160370827, + 0.15802446007728577, + 0.13868755102157593 + ], + "credit_map": [ + 0.02964743599295616, + 0.022397002205252647, + 0.05363890901207924, + 0.026854807510972023, + 0.026585189625620842, + 0.06644625961780548, + 0.03862043470144272, + 0.04038524627685547, + 0.007521791849285364, + 0.008551239967346191, + 0.019206009805202484, + 0.008553560823202133, + 0.008858758956193924, + 0.020628931000828743, + 0.008022979833185673, + 0.013643277809023857, + 0.005011401139199734, + 0.0062776231206953526, + 0.007998712360858917, + 0.00504678301513195, + 0.007831871509552002, + 0.009570527821779251, + 0.005476821679621935, + 0.0071614813059568405, + 0.006713405717164278, + 0.008161665871739388, + 0.007029609754681587, + 0.006020962726324797, + 0.0074271815828979015, + 0.012106511741876602, + 0.008681681007146835, + 0.006985520012676716, + 0.007756984792649746, + 0.012386810034513474, + 0.006972943432629108, + 0.005581929814070463, + 0.015211493708193302, + 0.010746688582003117, + 0.006851430516690016, + 0.005968667101114988, + 0.006127479486167431, + 0.004217613022774458, + 0.0059972601011395454, + 0.004221497569233179, + 0.004944643471390009, + 0.007703685201704502, + 0.0046657840721309185, + 0.006176547147333622, + 0.0032023920211941004, + 0.0023955574724823236, + 0.004455599468201399, + 0.0024435122031718493, + 0.0024308860301971436, + 0.003964523784816265, + 0.002422307152301073, + 0.003972797654569149, + 0.0040144119411706924, + 0.005392381455749273, + 0.005256061442196369, + 0.006508997175842524, + 0.007409921381622553, + 0.005696646403521299, + 0.005272527225315571, + 0.005854444578289986, + 0.01133850309997797, + 0.018991921097040176, + 0.0158978421241045, + 0.02583472616970539, + 0.03485286980867386, + 0.01830451749265194, + 0.017488932237029076, + 0.015315139666199684, + 0.009826933965086937, + 0.00828383769840002, + 0.036881327629089355, + 0.006627438589930534, + 0.011400071904063225, + 0.0497233010828495, + 0.006531403865665197, + 0.009413252584636211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.3497401177883148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.002851707860827446 + ], + "opposite_logprob_tokens": [ + -5.877851486206055 + ], + "logprob_reward": -0.002851707860827446, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002851707860827446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31364700198173523, + 0.0741843655705452, + 0.03658897429704666, + 0.07777854800224304, + 0.029091009870171547, + 0.08087048679590225, + 0.02275952510535717, + 0.07187342643737793, + 0.1095036044716835, + 0.18370305001735687 + ], + "credit_map": [ + 0.027353540062904358, + 0.01923828013241291, + 0.06634068489074707, + 0.02315044403076172, + 0.029849987477064133, + 0.07849524170160294, + 0.03562738373875618, + 0.033591434359550476, + 0.005558714270591736, + 0.00702213030308485, + 0.014928719960153103, + 0.005962559953331947, + 0.006955521646887064, + 0.01802203245460987, + 0.0074254912324249744, + 0.008309205062687397, + 0.0037494737189263105, + 0.003774855984374881, + 0.005239096004515886, + 0.005001912824809551, + 0.003414405044168234, + 0.005275125149637461, + 0.004917457699775696, + 0.005216649267822504, + 0.00856354832649231, + 0.011317290365695953, + 0.007412812206894159, + 0.009604844264686108, + 0.012622224166989326, + 0.009626112878322601, + 0.010532357729971409, + 0.008099352940917015, + 0.0037253566551953554, + 0.0023073528427630663, + 0.004889638628810644, + 0.002494808752089739, + 0.002686866791918874, + 0.007099739275872707, + 0.002757234964519739, + 0.0031300135888159275, + 0.008832081221044064, + 0.009116963483393192, + 0.010484262369573116, + 0.008696685545146465, + 0.01102328673005104, + 0.011488230898976326, + 0.010201593860983849, + 0.011027388274669647, + 0.002554247621446848, + 0.002534231636673212, + 0.0035328140947967768, + 0.0021467877086251974, + 0.0035037773195654154, + 0.004484771750867367, + 0.0019096225732937455, + 0.00209327251650393, + 0.0057723065838217735, + 0.00545092299580574, + 0.010915081016719341, + 0.01122860237956047, + 0.008218720555305481, + 0.013738218694925308, + 0.007855191826820374, + 0.008694383315742016, + 0.006827516946941614, + 0.01154902670532465, + 0.015003765933215618, + 0.016400614753365517, + 0.021531065925955772, + 0.018132586032152176, + 0.010826480574905872, + 0.009232548996806145, + 0.012445629574358463, + 0.013055949471890926, + 0.047397106885910034, + 0.011895528994500637, + 0.01804153434932232, + 0.06062827259302139, + 0.009705811738967896, + 0.010533221997320652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.002851707860827446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.001947531825862825 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.001947531825862825, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001947531825862825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31931981444358826, + 0.057431310415267944, + 0.04592588543891907, + 0.05413522571325302, + 0.06026626005768776, + 0.03434189409017563, + 0.028423495590686798, + 0.04959449544548988, + 0.10868000239133835, + 0.2418815940618515 + ], + "credit_map": [ + 0.030751211568713188, + 0.021199075505137444, + 0.06568026542663574, + 0.03191378340125084, + 0.028380833566188812, + 0.06516164541244507, + 0.042284660041332245, + 0.03394836559891701, + 0.004798589739948511, + 0.005272052716463804, + 0.011041252873837948, + 0.004390007350593805, + 0.006102224346250296, + 0.011675559915602207, + 0.00637842807918787, + 0.00777319585904479, + 0.004739969037473202, + 0.006217848043888807, + 0.005785444751381874, + 0.006134538911283016, + 0.005618682596832514, + 0.006827771198004484, + 0.005249439273029566, + 0.005352195352315903, + 0.005854958202689886, + 0.008486313745379448, + 0.00580415828153491, + 0.003965958021581173, + 0.010479484684765339, + 0.009136316366493702, + 0.005282559897750616, + 0.005125473253428936, + 0.005537119694054127, + 0.012748154811561108, + 0.006017260253429413, + 0.004516624379903078, + 0.014986297115683556, + 0.006815486587584019, + 0.005472705699503422, + 0.0041726091876626015, + 0.0032705587800592184, + 0.003929821774363518, + 0.004687869921326637, + 0.0034282102715224028, + 0.0052671972662210464, + 0.005707618314772844, + 0.0041615888476371765, + 0.0038890300784260035, + 0.0031858491711318493, + 0.0020523376297205687, + 0.005899061448872089, + 0.0019546167459338903, + 0.0020540845580399036, + 0.007616005372256041, + 0.002141741570085287, + 0.0035197974648326635, + 0.00435223150998354, + 0.004132658243179321, + 0.006944482214748859, + 0.007357553578913212, + 0.006177997682243586, + 0.007391077000647783, + 0.005968860350549221, + 0.00726963160559535, + 0.008264396339654922, + 0.00702969403937459, + 0.025241350755095482, + 0.007697203662246466, + 0.011514606885612011, + 0.0330841951072216, + 0.006408071611076593, + 0.009440484456717968, + 0.015832122415304184, + 0.025911523029208183, + 0.05185092240571976, + 0.012336722575128078, + 0.02896825410425663, + 0.0707428902387619, + 0.01600726507604122, + 0.020231910049915314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.001947531825862825 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.054967496544122696 + ], + "opposite_logprob_tokens": [ + -2.9299674034118652 + ], + "logprob_reward": -0.054967496544122696, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054967496544122696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2546040415763855, + 0.06543562561273575, + 0.048847462981939316, + 0.045402564108371735, + 0.058266934007406235, + 0.10447204858064651, + 0.03406832739710808, + 0.05845395103096962, + 0.1345362663269043, + 0.19591274857521057 + ], + "credit_map": [ + 0.030088700354099274, + 0.020507387816905975, + 0.03136755898594856, + 0.03414120525121689, + 0.044000230729579926, + 0.025895792990922928, + 0.04456797614693642, + 0.024035193026065826, + 0.006727165076881647, + 0.006418094504624605, + 0.013088556937873363, + 0.005244195461273193, + 0.007824991829693317, + 0.011416538618505001, + 0.006341170519590378, + 0.008374911732971668, + 0.00587271386757493, + 0.006721945013850927, + 0.005927601363509893, + 0.006172426510602236, + 0.005871320143342018, + 0.006973565556108952, + 0.006091726943850517, + 0.005216165911406279, + 0.004972207825630903, + 0.003343458054587245, + 0.006067960988730192, + 0.007153988815844059, + 0.0040851617231965065, + 0.006776103284209967, + 0.007099103648215532, + 0.005904576275497675, + 0.004922635853290558, + 0.009114387445151806, + 0.005232387222349644, + 0.007929939776659012, + 0.01314732525497675, + 0.006398250348865986, + 0.007140642032027245, + 0.0043813628144562244, + 0.008561081252992153, + 0.014233216643333435, + 0.01319858618080616, + 0.010174441151320934, + 0.020094089210033417, + 0.014611048623919487, + 0.012675309553742409, + 0.010924268513917923, + 0.0038751461543142796, + 0.002798312110826373, + 0.004166689235717058, + 0.004780855029821396, + 0.003741257358342409, + 0.004828655160963535, + 0.004175884649157524, + 0.005701528396457434, + 0.005375846289098263, + 0.0037756217643618584, + 0.012310116551816463, + 0.006219359580427408, + 0.004261220805346966, + 0.014630506746470928, + 0.005466844886541367, + 0.006414433941245079, + 0.010314490646123886, + 0.007211549673229456, + 0.0291905514895916, + 0.01292647048830986, + 0.011957032606005669, + 0.03862079232931137, + 0.010850039310753345, + 0.01346534676849842, + 0.01602284424006939, + 0.02815309911966324, + 0.03281016647815704, + 0.019232170656323433, + 0.027904769405722618, + 0.03223574161529541, + 0.019449396058917046, + 0.020104557275772095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.054967496544122696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370891332626343 + ], + "opposite_logprob_tokens": [ + -0.3870891034603119 + ], + "logprob_reward": -1.1370891332626343, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370891332626343, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2132808417081833, + 0.06049465388059616, + 0.10536416620016098, + 0.05826328322291374, + 0.08697662502527237, + 0.032999202609062195, + 0.055699512362480164, + 0.10085377842187881, + 0.04546822980046272, + 0.24059976637363434 + ], + "credit_map": [ + 0.026444831863045692, + 0.019256992265582085, + 0.025904769077897072, + 0.02243562787771225, + 0.027537990361452103, + 0.030064253136515617, + 0.02722785994410515, + 0.03440850228071213, + 0.005288132932037115, + 0.007099483627825975, + 0.007230744231492281, + 0.007970979437232018, + 0.009228921495378017, + 0.0077928476966917515, + 0.007646738085895777, + 0.008236794732511044, + 0.011144168674945831, + 0.010501446202397346, + 0.015829047188162804, + 0.0127332154661417, + 0.009673594497144222, + 0.018810436129570007, + 0.012019376270473003, + 0.01465286873281002, + 0.0052531505934894085, + 0.005160746164619923, + 0.008156337775290012, + 0.007848148234188557, + 0.006839573848992586, + 0.008655899204313755, + 0.007492604199796915, + 0.008856813423335552, + 0.007317009847611189, + 0.015696417540311813, + 0.007922559045255184, + 0.009180490858852863, + 0.020970817655324936, + 0.011016505770385265, + 0.009243466891348362, + 0.005629344843327999, + 0.003620428964495659, + 0.0037881662137806416, + 0.004604967311024666, + 0.0032184929586946964, + 0.005004508886486292, + 0.006328572519123554, + 0.002761244075372815, + 0.003672819584608078, + 0.005599519703537226, + 0.0033817999064922333, + 0.009828086011111736, + 0.004213675856590271, + 0.0054616727866232395, + 0.014423215761780739, + 0.005136138293892145, + 0.007655398454517126, + 0.007530210539698601, + 0.006823749281466007, + 0.02086733467876911, + 0.014592341147363186, + 0.008343088440597057, + 0.021840903908014297, + 0.010052476078271866, + 0.010803665034472942, + 0.0036273121368139982, + 0.004600109998136759, + 0.006731160916388035, + 0.006651382427662611, + 0.006551704835146666, + 0.005900139454752207, + 0.004964265041053295, + 0.006442150101065636, + 0.022395795211195946, + 0.016679957509040833, + 0.055619724094867706, + 0.014859708957374096, + 0.02418721467256546, + 0.06657429039478302, + 0.018584856763482094, + 0.021698208525776863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -1.1370891332626343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, smooth, 360-degree orbiting shot places the audience in an absolutely pitch-black, boundless void. An intense, concentrated, dramatic overhead light, like a stage spotlight, pierces this darkness vertically, creating a distinct, volumetric column of light. Within this column, countless tiny golden particles slowly drift. Rather than simply plummeting, they gracefully swirl and float through the air, as if drawn by an invisible breeze, creating a silent, stardust-like ballet. Each particle resembles a miniature diamond, varying in size and brightness. As they pass through the beam, their faceted structures capture and refract the dazzling light, creating a fleeting, dazzling starburst effect and subtle rainbow dispersions before the camera. As the camera rotates, the relationship between light and shadow shifts, with some particles illuminating momentarily upon entering the column and fading quietly into darkness upon exiting, creating a perpetually dynamic aesthetic. The main colors of the entire picture are deep black and warm and dazzling gold, creating a magical atmosphere that is mysterious, grand, peaceful and dreamy, perfectly capturing this cosmic dance carefully woven by light and shadow.", + "seed": 8901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_6353_seed_8901_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": -0.35142279306892305, + "vlm_logprob": -0.35142279306892305, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.35142279306892305, + "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 background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.4760582447052002 + ], + "opposite_logprob_tokens": [ + -0.9760582447052002 + ], + "logprob_reward": -0.4760582447052002, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4760582447052002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2828005850315094, + 0.06526584178209305, + 0.11199161410331726, + 0.05141625180840492, + 0.13534832000732422, + 0.0881548747420311, + 0.025488337501883507, + 0.038927674293518066, + 0.17029860615730286, + 0.030307911336421967 + ], + "credit_map": [ + 0.02961972914636135, + 0.0214712992310524, + 0.07274819910526276, + 0.02805965021252632, + 0.020619891583919525, + 0.0650053322315216, + 0.025947773829102516, + 0.01932869479060173, + 0.004591465461999178, + 0.010308339260518551, + 0.009275764226913452, + 0.007388955447822809, + 0.008958984166383743, + 0.007727793883532286, + 0.00998708140105009, + 0.007027455139905214, + 0.00870391633361578, + 0.010606412775814533, + 0.020829109475016594, + 0.012277757748961449, + 0.010670164600014687, + 0.02368794195353985, + 0.011130725964903831, + 0.014085580594837666, + 0.004968392662703991, + 0.0053617381490767, + 0.007798460312187672, + 0.011280364356935024, + 0.004414469003677368, + 0.005825165659189224, + 0.006763921584933996, + 0.005003736354410648, + 0.016522910445928574, + 0.020520174875855446, + 0.02004832960665226, + 0.021460188552737236, + 0.017874427139759064, + 0.014922970905900002, + 0.01523116696625948, + 0.008768140338361263, + 0.010688858106732368, + 0.01537505816668272, + 0.01595989428460598, + 0.008903148584067822, + 0.011777867563068867, + 0.011491924524307251, + 0.008244282566010952, + 0.005713843274861574, + 0.002397386124357581, + 0.002295912243425846, + 0.004843481350690126, + 0.002199477283284068, + 0.004518700297921896, + 0.003694540588185191, + 0.0036092365626245737, + 0.0019296039827167988, + 0.0028796824626624584, + 0.0038041609805077314, + 0.005586868617683649, + 0.004795314744114876, + 0.006097420118749142, + 0.004403896164149046, + 0.007280832156538963, + 0.004079500213265419, + 0.015379021875560284, + 0.017937680706381798, + 0.02511981502175331, + 0.033954594284296036, + 0.02105187624692917, + 0.028462903574109077, + 0.015288243070244789, + 0.01310446485877037, + 0.0025069217663258314, + 0.0025823041796684265, + 0.004363652318716049, + 0.006590977776795626, + 0.003430112497881055, + 0.0038433086592704058, + 0.004606945440173149, + 0.0023836863692849874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.4760582447052002 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.007664084900170565 + ], + "opposite_logprob_tokens": [ + -4.882664203643799 + ], + "logprob_reward": -0.007664084900170565, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007664084900170565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28730037808418274, + 0.03426743298768997, + 0.05090108513832092, + 0.05051572248339653, + 0.05685262382030487, + 0.04992799833416939, + 0.022899368777871132, + 0.0551854707300663, + 0.19933955371379852, + 0.19281040132045746 + ], + "credit_map": [ + 0.022544540464878082, + 0.02336954139173031, + 0.07573840767145157, + 0.026554908603429794, + 0.01893802173435688, + 0.05355332791805267, + 0.041781991720199585, + 0.024819614365696907, + 0.00247999862767756, + 0.0046151019632816315, + 0.005468466784805059, + 0.0046842978335917, + 0.004753932822495699, + 0.0041820937767624855, + 0.006173354107886553, + 0.001910185907036066, + 0.0045012058690190315, + 0.005394758656620979, + 0.008851787075400352, + 0.0067389532923698425, + 0.005714871920645237, + 0.008560528978705406, + 0.00539267947897315, + 0.0057463026605546474, + 0.0052741593681275845, + 0.004821802489459515, + 0.008160918019711971, + 0.006245676893740892, + 0.005816316232085228, + 0.00814591534435749, + 0.0063244933262467384, + 0.005726438947021961, + 0.008812098763883114, + 0.013029193505644798, + 0.007850571535527706, + 0.007102016359567642, + 0.0048860302194952965, + 0.006413895636796951, + 0.004776172339916229, + 0.00398264778777957, + 0.00522638950496912, + 0.007522833999246359, + 0.011237801983952522, + 0.0049653141759335995, + 0.00576964532956481, + 0.0078510083258152, + 0.003916293848305941, + 0.0034387116320431232, + 0.002533290069550276, + 0.002130696317180991, + 0.004248779267072678, + 0.002381375525146723, + 0.0037601313088089228, + 0.0029207903426140547, + 0.0031346851028501987, + 0.0017896209610626101, + 0.006305949296802282, + 0.004419481847435236, + 0.009211370721459389, + 0.005365421064198017, + 0.007965799421072006, + 0.006090995389968157, + 0.00851184781640768, + 0.007314604241400957, + 0.012503317557275295, + 0.017550203949213028, + 0.02509250119328499, + 0.03732782229781151, + 0.027339039370417595, + 0.03885070979595184, + 0.0266377292573452, + 0.014038224704563618, + 0.013024838641285896, + 0.017122963443398476, + 0.026472559198737144, + 0.028015485033392906, + 0.029397135600447655, + 0.022999266162514687, + 0.04316408932209015, + 0.012614077888429165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.007664084900170565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.0010484919184818864 + ], + "opposite_logprob_tokens": [ + -6.876048564910889 + ], + "logprob_reward": -0.0010484919184818864, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010484919184818864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23968341946601868, + 0.04401267319917679, + 0.05852917209267616, + 0.048700157552957535, + 0.05720669776201248, + 0.06163130700588226, + 0.0234540868550539, + 0.017594555392861366, + 0.2195420265197754, + 0.2296459674835205 + ], + "credit_map": [ + 0.029902877286076546, + 0.025146260857582092, + 0.034965939819812775, + 0.034120798110961914, + 0.019621921703219414, + 0.028906693682074547, + 0.048382218927145004, + 0.018636686727404594, + 0.0029261340387165546, + 0.007262316532433033, + 0.005939848255366087, + 0.005601238925009966, + 0.006462723016738892, + 0.004547677468508482, + 0.006999034434556961, + 0.00427369587123394, + 0.004896866157650948, + 0.005367945414036512, + 0.010104983113706112, + 0.006563441827893257, + 0.005827412009239197, + 0.010345442220568657, + 0.006746797356754541, + 0.008676281198859215, + 0.004436108283698559, + 0.0051742843352258205, + 0.00730166956782341, + 0.007333057466894388, + 0.006105121690779924, + 0.00749658327549696, + 0.005783518310636282, + 0.005069813225418329, + 0.006354648154228926, + 0.008361555635929108, + 0.00923849456012249, + 0.008381904102861881, + 0.007507589180022478, + 0.005842657294124365, + 0.006730740424245596, + 0.004789105616509914, + 0.006277039647102356, + 0.009175452403724194, + 0.010090506635606289, + 0.008237474597990513, + 0.007801366038620472, + 0.007468774449080229, + 0.007277185562998056, + 0.005303504876792431, + 0.001794106443412602, + 0.001797507400624454, + 0.005165802780538797, + 0.0025291242636740208, + 0.003903327276930213, + 0.0031851991079747677, + 0.0032819476909935474, + 0.001797069446183741, + 0.002031287644058466, + 0.0016940549248829484, + 0.0030960002914071083, + 0.00240605091676116, + 0.002580290660262108, + 0.0018731593154370785, + 0.0024181785993278027, + 0.0014955319929867983, + 0.01722034066915512, + 0.018598925322294235, + 0.03578448295593262, + 0.042820077389478683, + 0.03362541273236275, + 0.029539627954363823, + 0.02434886246919632, + 0.017604272812604904, + 0.016138063743710518, + 0.021376170217990875, + 0.033183738589286804, + 0.03346901386976242, + 0.040390852838754654, + 0.03209556266665459, + 0.036155808717012405, + 0.0168367400765419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0010484919184818864 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.04866431653499603 + ], + "opposite_logprob_tokens": [ + -3.0486643314361572 + ], + "logprob_reward": -0.04866431653499603, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866431653499603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18730449676513672, + 0.05999309942126274, + 0.03807990252971649, + 0.030187608674168587, + 0.05283854529261589, + 0.04193086549639702, + 0.06882744282484055, + 0.08941339701414108, + 0.13716915249824524, + 0.29425549507141113 + ], + "credit_map": [ + 0.017988666892051697, + 0.019133104011416435, + 0.030550163239240646, + 0.024349069222807884, + 0.016102580353617668, + 0.028558366000652313, + 0.03589896485209465, + 0.014723584055900574, + 0.00598698016256094, + 0.008759230375289917, + 0.008801618590950966, + 0.007720927242189646, + 0.008602875284850597, + 0.006964860484004021, + 0.008131919428706169, + 0.005024690646678209, + 0.004584840498864651, + 0.0038874915335327387, + 0.006459672935307026, + 0.0047125546261668205, + 0.0035297651775181293, + 0.005941420793533325, + 0.003880413481965661, + 0.005083743948489428, + 0.003328015562146902, + 0.0037185545079410076, + 0.004735862836241722, + 0.00382354948669672, + 0.0027825108263641596, + 0.00410588039085269, + 0.0033826311118900776, + 0.0043106055818498135, + 0.00668170303106308, + 0.007989246398210526, + 0.007567278109490871, + 0.007221761159598827, + 0.006629898678511381, + 0.005848212167620659, + 0.0063994890078902245, + 0.004500952549278736, + 0.004143455997109413, + 0.005938652437180281, + 0.006739725358784199, + 0.006938563659787178, + 0.004784094635397196, + 0.0052648489363491535, + 0.00507901469245553, + 0.003042511409148574, + 0.0061621288768947124, + 0.005894694477319717, + 0.014430990442633629, + 0.00720922602340579, + 0.010000055655837059, + 0.008952903561294079, + 0.00844215415418148, + 0.007735293358564377, + 0.006946362089365721, + 0.006871412508189678, + 0.014778426848351955, + 0.009766979143023491, + 0.013192497193813324, + 0.010171046480536461, + 0.016887275502085686, + 0.010799406096339226, + 0.010769844986498356, + 0.014170050621032715, + 0.020828325301408768, + 0.027482785284519196, + 0.017591407522559166, + 0.019789420068264008, + 0.01526390016078949, + 0.011273427866399288, + 0.022102762013673782, + 0.02619560994207859, + 0.043351467698812485, + 0.04439811408519745, + 0.04694032669067383, + 0.044636305421590805, + 0.044173818081617355, + 0.022457091137766838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04866431653499603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236788272857666 + ], + "opposite_logprob_tokens": [ + -0.3486787974834442 + ], + "logprob_reward": -1.2236788272857666, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236788272857666, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21803423762321472, + 0.04865477979183197, + 0.0692778155207634, + 0.09725409001111984, + 0.10599922388792038, + 0.05445989966392517, + 0.07037553191184998, + 0.04276030883193016, + 0.1879449337720871, + 0.10523918271064758 + ], + "credit_map": [ + 0.022404959425330162, + 0.02197759971022606, + 0.043907713145017624, + 0.025773968547582626, + 0.023724349215626717, + 0.02994832955300808, + 0.035272467881441116, + 0.015024862252175808, + 0.003956162370741367, + 0.007840510457754135, + 0.007738009560853243, + 0.006658024154603481, + 0.006625811103731394, + 0.005344298202544451, + 0.006381231360137463, + 0.004110727459192276, + 0.006433078553527594, + 0.007104235701262951, + 0.012586039490997791, + 0.007376533932983875, + 0.007221200503408909, + 0.013705713674426079, + 0.00777364382520318, + 0.007077372167259455, + 0.00956906471401453, + 0.009937373921275139, + 0.01474023051559925, + 0.017871426418423653, + 0.011077179573476315, + 0.013169104233384132, + 0.010956375859677792, + 0.009933331981301308, + 0.009442739188671112, + 0.022081363946199417, + 0.01607205718755722, + 0.01273509580641985, + 0.012526744976639748, + 0.01289195567369461, + 0.01238533016294241, + 0.007863936945796013, + 0.006098931189626455, + 0.009857870638370514, + 0.009571434929966927, + 0.005324850790202618, + 0.006323919165879488, + 0.00688915466889739, + 0.005989175755530596, + 0.0044045643880963326, + 0.008099053986370564, + 0.00756070576608181, + 0.01114885788410902, + 0.00915487390011549, + 0.00992756336927414, + 0.009013419970870018, + 0.008347352035343647, + 0.007123702205717564, + 0.004596345592290163, + 0.004335848148912191, + 0.006539410911500454, + 0.005269848275929689, + 0.007328469306230545, + 0.003864137688651681, + 0.00702808378264308, + 0.0037981641944497824, + 0.01679571531713009, + 0.019427547231316566, + 0.023779377341270447, + 0.0321030393242836, + 0.022248439490795135, + 0.04000625014305115, + 0.016990218311548233, + 0.016594350337982178, + 0.008951177820563316, + 0.008208278566598892, + 0.015558385290205479, + 0.020682021975517273, + 0.011145930737257004, + 0.01901850476861, + 0.013164552859961987, + 0.008510332554578781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -1.2236788272857666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, smooth, 360-degree orbiting shot places the audience in an absolutely pitch-black, boundless void. An intense, concentrated, dramatic overhead light, like a stage spotlight, pierces this darkness vertically, creating a distinct, volumetric column of light. Within this column, countless tiny golden particles slowly drift. Rather than simply plummeting, they gracefully swirl and float through the air, as if drawn by an invisible breeze, creating a silent, stardust-like ballet. Each particle resembles a miniature diamond, varying in size and brightness. As they pass through the beam, their faceted structures capture and refract the dazzling light, creating a fleeting, dazzling starburst effect and subtle rainbow dispersions before the camera. As the camera rotates, the relationship between light and shadow shifts, with some particles illuminating momentarily upon entering the column and fading quietly into darkness upon exiting, creating a perpetually dynamic aesthetic. The main colors of the entire picture are deep black and warm and dazzling gold, creating a magical atmosphere that is mysterious, grand, peaceful and dreamy, perfectly capturing this cosmic dance carefully woven by light and shadow.", + "seed": 8902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_11/index_6353_seed_8902_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": -0.601380706904456, + "vlm_logprob": -0.601380706904456, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.601380706904456, + "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 background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.7035400867462158 + ], + "opposite_logprob_tokens": [ + -0.20354004204273224 + ], + "logprob_reward": -1.7035400867462158, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.7035400867462158, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10067188739776611, + 0.052493270486593246, + 0.039866287261247635, + 0.1525145322084427, + 0.22557534277439117, + 0.11088091135025024, + 0.058271829038858414, + 0.09714854508638382, + 0.07084669172763824, + 0.09173072129487991 + ], + "credit_map": [ + 0.00929918885231018, + 0.010660809464752674, + 0.014892748557031155, + 0.011788376607000828, + 0.012962008826434612, + 0.014496412128210068, + 0.01441315095871687, + 0.012159191071987152, + 0.005567846819758415, + 0.0070855203084647655, + 0.007981891743838787, + 0.0077078677713871, + 0.0029150820337235928, + 0.010140523314476013, + 0.0035893316380679607, + 0.007505206856876612, + 0.0031707945745438337, + 0.0033375327475368977, + 0.005987132433801889, + 0.004660008009523153, + 0.0025738643016666174, + 0.010758455842733383, + 0.0026168006006628275, + 0.006761695723980665, + 0.018127046525478363, + 0.01905740424990654, + 0.031983572989702225, + 0.01468854770064354, + 0.013522020541131496, + 0.026794135570526123, + 0.014973047189414501, + 0.01336877141147852, + 0.014034227468073368, + 0.012972833588719368, + 0.04770299047231674, + 0.02402590773999691, + 0.013012160547077656, + 0.06951230764389038, + 0.020566517487168312, + 0.023748403415083885, + 0.008696319535374641, + 0.006425356026738882, + 0.01923060417175293, + 0.017143111675977707, + 0.008520412258803844, + 0.022938648238778114, + 0.014989320188760757, + 0.012937137857079506, + 0.004591336939483881, + 0.00477202981710434, + 0.014194237999618053, + 0.008299936540424824, + 0.00422796281054616, + 0.010575923137366772, + 0.007196309044957161, + 0.0044140927493572235, + 0.005200004670768976, + 0.0048614852130413055, + 0.022912433370947838, + 0.005748300813138485, + 0.007137347478419542, + 0.03336602821946144, + 0.006661556661128998, + 0.011261389590799809, + 0.00653265742585063, + 0.007127326913177967, + 0.0105040417984128, + 0.00894197914749384, + 0.010049746371805668, + 0.007714519277215004, + 0.013398401439189911, + 0.006578016094863415, + 0.007623990532010794, + 0.00556610431522131, + 0.014696333557367325, + 0.011173529550433159, + 0.014074761420488358, + 0.008013064041733742, + 0.02025262825191021, + 0.010330310091376305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -1.7035400867462158 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.006007116287946701 + ], + "opposite_logprob_tokens": [ + -5.131007194519043 + ], + "logprob_reward": -0.006007116287946701, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006007116287946701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13111135363578796, + 0.046918775886297226, + 0.03280545026063919, + 0.1560789942741394, + 0.17325061559677124, + 0.12102026492357254, + 0.04921426251530647, + 0.12097708135843277, + 0.06804265081882477, + 0.10058058053255081 + ], + "credit_map": [ + 0.010385069064795971, + 0.014959302730858326, + 0.014991801232099533, + 0.011837029829621315, + 0.01350291259586811, + 0.024300334975123405, + 0.016531789675354958, + 0.024603107944130898, + 0.0051628644578158855, + 0.0061522796750068665, + 0.006961905397474766, + 0.00678754597902298, + 0.0028198817744851112, + 0.00852975994348526, + 0.004004063084721565, + 0.006500474642962217, + 0.003074555192142725, + 0.0031614634208381176, + 0.005718874745070934, + 0.005337702576071024, + 0.0024204649962484837, + 0.006230858154594898, + 0.0027102928142994642, + 0.004151239525526762, + 0.016347983852028847, + 0.017618296667933464, + 0.03420940786600113, + 0.012622468173503876, + 0.013096267357468605, + 0.03514394164085388, + 0.012257496826350689, + 0.014783128164708614, + 0.011034800671041012, + 0.010478914715349674, + 0.03597944602370262, + 0.019180821254849434, + 0.010650508105754852, + 0.05335284024477005, + 0.013755380176007748, + 0.01881788857281208, + 0.008299877867102623, + 0.005968146491795778, + 0.02740212343633175, + 0.014723778702318668, + 0.00798852276057005, + 0.03183625265955925, + 0.012435752898454666, + 0.012365802191197872, + 0.004270038567483425, + 0.0048378948122262955, + 0.007060665171593428, + 0.007561716716736555, + 0.004304662346839905, + 0.008946893736720085, + 0.0065876394510269165, + 0.00564474705606699, + 0.006657852325588465, + 0.005615752190351486, + 0.03256053477525711, + 0.008043832145631313, + 0.008868148550391197, + 0.04037982225418091, + 0.008097006939351559, + 0.010754127986729145, + 0.006397079676389694, + 0.007055239286273718, + 0.009601502679288387, + 0.009014482609927654, + 0.009184107184410095, + 0.007588261738419533, + 0.012855597771704197, + 0.006346377078443766, + 0.008829965256154537, + 0.006595006212592125, + 0.01686926744878292, + 0.013560284860432148, + 0.012745910324156284, + 0.010221870616078377, + 0.019769836217164993, + 0.011988434009253979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.006007116287946701 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.003628814360126853 + ], + "opposite_logprob_tokens": [ + -5.628628730773926 + ], + "logprob_reward": -0.003628814360126853, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003628814360126853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12168312817811966, + 0.05502120032906532, + 0.03630894050002098, + 0.19107691943645477, + 0.14746715128421783, + 0.0941426083445549, + 0.058180712163448334, + 0.1152833104133606, + 0.08841383457183838, + 0.09242218732833862 + ], + "credit_map": [ + 0.008484043180942535, + 0.012331105768680573, + 0.01464938372373581, + 0.010394754819571972, + 0.010892248712480068, + 0.026409443467855453, + 0.013513058423995972, + 0.025009088218212128, + 0.006397430785000324, + 0.007478263229131699, + 0.008355126716196537, + 0.00770476832985878, + 0.0029262122698128223, + 0.010789565742015839, + 0.0033392696641385555, + 0.008030564524233341, + 0.0027977696154266596, + 0.0035630809143185616, + 0.0057767173275351524, + 0.005086183547973633, + 0.002131133805960417, + 0.008926223032176495, + 0.0028315018862485886, + 0.00519633200019598, + 0.017961451783776283, + 0.019819386303424835, + 0.04348215460777283, + 0.0145201925188303, + 0.01590772531926632, + 0.04836313799023628, + 0.012629333883523941, + 0.01839355006814003, + 0.009909887798130512, + 0.009264419786632061, + 0.02915516123175621, + 0.01830204948782921, + 0.009465121664106846, + 0.042841505259275436, + 0.013065106235444546, + 0.015463894233107567, + 0.00829879567027092, + 0.006313502788543701, + 0.015073129907250404, + 0.013048535212874413, + 0.00962872989475727, + 0.017429856583476067, + 0.011020851321518421, + 0.013329208828508854, + 0.00510355643928051, + 0.0037052419502288103, + 0.015098223462700844, + 0.005550782196223736, + 0.004598365165293217, + 0.01150374673306942, + 0.008373668417334557, + 0.0042471326887607574, + 0.006667137145996094, + 0.004457420203834772, + 0.03411570191383362, + 0.007242181338369846, + 0.007088731043040752, + 0.03789953887462616, + 0.01025846041738987, + 0.007554135750979185, + 0.007103237323462963, + 0.008730003610253334, + 0.012879236601293087, + 0.01180977188050747, + 0.015032785944640636, + 0.00867195799946785, + 0.015204062685370445, + 0.00898277573287487, + 0.008214927278459072, + 0.007792506832629442, + 0.011980297975242138, + 0.011264928616583347, + 0.012297083623707294, + 0.014459011144936085, + 0.01702188514173031, + 0.009391534142196178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.003628814360126853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": [ + -0.07004249095916748 + ], + "opposite_logprob_tokens": [ + -2.695042610168457 + ], + "logprob_reward": -0.07004249095916748, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004249095916748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11546136438846588, + 0.06380967795848846, + 0.0357169546186924, + 0.13982433080673218, + 0.045068852603435516, + 0.06220543757081032, + 0.045393116772174835, + 0.20015481114387512, + 0.13562211394309998, + 0.1567433476448059 + ], + "credit_map": [ + 0.014254159294068813, + 0.017721323296427727, + 0.02168276347219944, + 0.012978477403521538, + 0.008999217301607132, + 0.013576066121459007, + 0.015606395900249481, + 0.010642966255545616, + 0.006572452373802662, + 0.0078115314245224, + 0.009200765751302242, + 0.008733183145523071, + 0.003828603308647871, + 0.01370741706341505, + 0.003534613875672221, + 0.010421110317111015, + 0.002428083447739482, + 0.0027544673066586256, + 0.00473649799823761, + 0.004493394400924444, + 0.0020384571980684996, + 0.008855937048792839, + 0.002691984875127673, + 0.007718131877481937, + 0.013105401769280434, + 0.016184132546186447, + 0.02840074524283409, + 0.01066895667463541, + 0.012872829101979733, + 0.03367050364613533, + 0.010734716430306435, + 0.014187044464051723, + 0.003721828805282712, + 0.0037162648513913155, + 0.006205326411873102, + 0.008343352936208248, + 0.00469721807166934, + 0.007963304407894611, + 0.00618399865925312, + 0.0042375619523227215, + 0.004532276652753353, + 0.0038136520888656378, + 0.018878107890486717, + 0.005927262827754021, + 0.004262261092662811, + 0.016416803002357483, + 0.003616855014115572, + 0.004758219234645367, + 0.003882594173774123, + 0.0028074507135897875, + 0.011378323659300804, + 0.004805414006114006, + 0.004205760080367327, + 0.007213587872684002, + 0.006623874884098768, + 0.004476108122617006, + 0.011217053979635239, + 0.006801224779337645, + 0.062497057020664215, + 0.012770503759384155, + 0.012691738083958626, + 0.06435061991214752, + 0.01759696938097477, + 0.012229641899466515, + 0.010727686807513237, + 0.009017751552164555, + 0.02552102506160736, + 0.01823941431939602, + 0.015953872352838516, + 0.021647578105330467, + 0.020831480622291565, + 0.013683303259313107, + 0.01311566587537527, + 0.012356272898614407, + 0.024003805592656136, + 0.018511245027184486, + 0.021426985040307045, + 0.023827631026506424, + 0.030686164274811745, + 0.012815576046705246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.07004249095916748 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the background a pitch-black, boundless void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an intense overhead light source creating a distinct vertical volumetric column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do countless tiny golden particles drift and swirl within the light column?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the particles exhibit faceted structures that refract light into starburst effects and rainbow dispersions?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, smooth 360-degree orbiting 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": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236850261688232 + ], + "opposite_logprob_tokens": [ + -0.34868499636650085 + ], + "logprob_reward": -1.2236850261688232, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236850261688232, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1057611033320427, + 0.03599442541599274, + 0.044782668352127075, + 0.04702598974108696, + 0.059857890009880066, + 0.37776297330856323, + 0.07808684557676315, + 0.062268491834402084, + 0.07459373027086258, + 0.11386586725711823 + ], + "credit_map": [ + 0.009552974253892899, + 0.009736862033605576, + 0.012647571042180061, + 0.008442150428891182, + 0.0068760886788368225, + 0.02650715783238411, + 0.00667470321059227, + 0.025323593989014626, + 0.0033641518093645573, + 0.004221959505230188, + 0.006500761955976486, + 0.006124085746705532, + 0.002009287243708968, + 0.006753760389983654, + 0.002171456813812256, + 0.004848960787057877, + 0.0028128959238529205, + 0.002928047440946102, + 0.005749412812292576, + 0.0060509201139211655, + 0.002422359073534608, + 0.012396679259836674, + 0.003393758088350296, + 0.009028594009578228, + 0.004601478576660156, + 0.003396443324163556, + 0.008536121807992458, + 0.0049358997493982315, + 0.005921354051679373, + 0.008483946323394775, + 0.00635787146165967, + 0.004792874678969383, + 0.0054235816933214664, + 0.004739308264106512, + 0.01039642933756113, + 0.006997513584792614, + 0.007668728940188885, + 0.009898790158331394, + 0.010667344555258751, + 0.004066192079335451, + 0.021543435752391815, + 0.01014119852334261, + 0.14171841740608215, + 0.021816937252879143, + 0.012290249578654766, + 0.13802999258041382, + 0.016195526346564293, + 0.016027191653847694, + 0.0059422398917376995, + 0.005438460968434811, + 0.01743946224451065, + 0.012783526442945004, + 0.0071589224971830845, + 0.014816268347203732, + 0.009274189360439777, + 0.005233782343566418, + 0.0056211682967841625, + 0.003171428106725216, + 0.015298456884920597, + 0.006203909870237112, + 0.005491703283041716, + 0.012445567175745964, + 0.00858303438872099, + 0.005453222431242466, + 0.005849871318787336, + 0.004671467002481222, + 0.017088301479816437, + 0.008117923513054848, + 0.008570301346480846, + 0.010425327345728874, + 0.00967080146074295, + 0.010199730284512043, + 0.009481549263000488, + 0.008625307120382786, + 0.017091909423470497, + 0.01384496595710516, + 0.016363361850380898, + 0.012393265962600708, + 0.024390311911702156, + 0.011675205081701279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -1.2236850261688232 + } + ], + "vlm_decoded_outputs": [ + "No", + "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/0000158/group_12/index_tensor([9420, 9420])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_tensor([9420, 9420])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa14204ba85ec9f6d66e85a710604e1dd8369fd7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_tensor([9420, 9420])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, low-angle, static shot frames a steel paperclip as it executes a perfect, steady gyroscopic motion, hovering above the smooth surface of a rain-soaked alley puddle. The polished metal in the clip, rendered in stunning high-contrast detail, captures the flickering neon signs of the surrounding dystopian cyberpunk cityscape, its curved bodies twisting vibrant hues of fuchsia, cyan, and deep blue into flowing, dancing streaks. The camera's low perspective emphasizes the impossible physics at play, placing the small, gleaming object against a backdrop of dark, wet asphalt and grimy, towering building facades, gently billowing with steam. A fine mist hangs in the air, suffused with the glaring glow of holographic advertisements, while the puddle below reflects the clip's silent, hypnotic gyrations, creating a surreal and captivating focal point within the gritty, oppressive urban environment, evoking an atmosphere of quiet mystery and technological wonder.", + "seed": 19561, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_9420_seed_19561_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": -0.06570375021547079, + "vlm_logprob": -0.06570375021547079, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.06570375021547079, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2319270223379135 + ], + "opposite_logprob_tokens": [ + -1.6069270372390747 + ], + "logprob_reward": -0.2319270223379135, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2319270223379135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1839275360107422, + 0.03223378583788872, + 0.08159393072128296, + 0.035165417939424515, + 0.012975038960576057, + 0.027763620018959045, + 0.05585000291466713, + 0.0812065601348877, + 0.18717840313911438, + 0.30210578441619873 + ], + "credit_map": [ + 0.011183606460690498, + 0.029092280194163322, + 0.014254939742386341, + 0.014319198206067085, + 0.03493332862854004, + 0.023662758991122246, + 0.029249247163534164, + 0.02723216451704502, + 0.003076378721743822, + 0.002610910916700959, + 0.002440636046230793, + 0.0035105396527796984, + 0.0038207881152629852, + 0.006662444677203894, + 0.0025580348446965218, + 0.007554049603641033, + 0.007999333553016186, + 0.013493753969669342, + 0.007369454950094223, + 0.006911640986800194, + 0.013817324303090572, + 0.013381287455558777, + 0.0083469832316041, + 0.010274138301610947, + 0.004058565013110638, + 0.004911115858703852, + 0.00437044445425272, + 0.004041934851557016, + 0.004164815414696932, + 0.006293905433267355, + 0.0040027364157140255, + 0.003321897704154253, + 0.0020402565132826567, + 0.001599186914972961, + 0.0017634188989177346, + 0.0017344842199236155, + 0.001256836112588644, + 0.0014872042229399085, + 0.001683546113781631, + 0.0014101057313382626, + 0.003676192369312048, + 0.0038333148695528507, + 0.003969329409301281, + 0.004675503820180893, + 0.003208730136975646, + 0.002549446653574705, + 0.0034275345969945192, + 0.002423566998913884, + 0.005788929294794798, + 0.01242409460246563, + 0.007160167209804058, + 0.008000906556844711, + 0.005771235097199678, + 0.003927850164473057, + 0.007062163203954697, + 0.005714653991162777, + 0.006361914798617363, + 0.009612883441150188, + 0.011026364751160145, + 0.019264906644821167, + 0.009565521962940693, + 0.00493860337883234, + 0.012408571317791939, + 0.008027787320315838, + 0.014811481349170208, + 0.044519390910863876, + 0.03868536651134491, + 0.01827191561460495, + 0.025510530918836594, + 0.009427542798221111, + 0.02842698059976101, + 0.007525183260440826, + 0.03076120652258396, + 0.03759074956178665, + 0.045682426542043686, + 0.032219357788562775, + 0.051996853202581406, + 0.031074821949005127, + 0.04127397760748863, + 0.03150637075304985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.2319270223379135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043117064982652664 + ], + "opposite_logprob_tokens": [ + -3.168117046356201 + ], + "logprob_reward": -0.043117064982652664, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043117064982652664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23895679414272308, + 0.044338569045066833, + 0.056590430438518524, + 0.03975178673863411, + 0.02146225981414318, + 0.04375970736145973, + 0.04553299397230148, + 0.08171656727790833, + 0.12234805524349213, + 0.30554282665252686 + ], + "credit_map": [ + 0.019161483272910118, + 0.022596390917897224, + 0.019389385357499123, + 0.01837635599076748, + 0.05686292424798012, + 0.03394465148448944, + 0.038074687123298645, + 0.030550912022590637, + 0.006353008095175028, + 0.0041054170578718185, + 0.004537490662187338, + 0.004403913393616676, + 0.006781832315027714, + 0.007029062602669001, + 0.004226625431329012, + 0.00690121715888381, + 0.006923032458871603, + 0.008037143386900425, + 0.006524872500449419, + 0.0050024669617414474, + 0.0068678646348416805, + 0.009512258693575859, + 0.006261000409722328, + 0.007461788132786751, + 0.004489913582801819, + 0.003936800640076399, + 0.007760568987578154, + 0.0037349823396652937, + 0.004328238312155008, + 0.0046100676991045475, + 0.005096444394439459, + 0.005794770549982786, + 0.003087362041696906, + 0.002874207217246294, + 0.0025602676905691624, + 0.00269909156486392, + 0.002009305637329817, + 0.0028716789092868567, + 0.0028977489564567804, + 0.002462597331032157, + 0.004923746921122074, + 0.004745191428810358, + 0.008372689597308636, + 0.008922946639358997, + 0.004421649966388941, + 0.0031893805135041475, + 0.005773443728685379, + 0.003410658333450556, + 0.006443909835070372, + 0.008129153400659561, + 0.009266766719520092, + 0.005963260307908058, + 0.004621667787432671, + 0.0027304450049996376, + 0.004403776489198208, + 0.003974012099206448, + 0.007413338869810104, + 0.013148126192390919, + 0.010570088401436806, + 0.014988417737185955, + 0.00802859477698803, + 0.006423576269298792, + 0.011811542324721813, + 0.009332886897027493, + 0.010116847231984138, + 0.016292210668325424, + 0.0307953879237175, + 0.014779808931052685, + 0.011832845397293568, + 0.007360605522990227, + 0.023691797628998756, + 0.00747855007648468, + 0.027420399710536003, + 0.022884655743837357, + 0.0893588736653328, + 0.05717071518301964, + 0.029486343264579773, + 0.02316032350063324, + 0.026371728628873825, + 0.029689785093069077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.043117064982652664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09624877572059631, + 0.11125839501619339, + 0.13568611443042755, + 0.19058933854103088, + 0.06957859545946121, + 0.045234665274620056, + 0.04610107094049454, + 0.020280372351408005, + 0.13714510202407837, + 0.147877499461174 + ], + "credit_map": [ + 0.008383399806916714, + 0.01192423328757286, + 0.010313360020518303, + 0.007111392915248871, + 0.014728127978742123, + 0.014191767200827599, + 0.015041077509522438, + 0.014555414207279682, + 0.007617810275405645, + 0.006620725616812706, + 0.010279550217092037, + 0.009456422179937363, + 0.007430320605635643, + 0.0294854324311018, + 0.008549715392291546, + 0.03181841969490051, + 0.016213033348321915, + 0.01593654416501522, + 0.015464283525943756, + 0.012141264975070953, + 0.0180879645049572, + 0.020279129967093468, + 0.017630966380238533, + 0.019932923838496208, + 0.021089429035782814, + 0.01614944450557232, + 0.05852467566728592, + 0.01468680053949356, + 0.00999414175748825, + 0.01584080047905445, + 0.011728575453162193, + 0.04257545992732048, + 0.00752048147842288, + 0.01318070013076067, + 0.008521419018507004, + 0.0113905631005764, + 0.007023153360933065, + 0.007171682082116604, + 0.008748817257583141, + 0.006021774839609861, + 0.008891014382243156, + 0.005934474989771843, + 0.006192236207425594, + 0.007853356190025806, + 0.004238770809024572, + 0.004800830036401749, + 0.0032689308281987906, + 0.004055047873407602, + 0.005224073771387339, + 0.00806425977498293, + 0.006052193697541952, + 0.004794938024133444, + 0.007538253907114267, + 0.0033103141468018293, + 0.006377832498401403, + 0.004739201627671719, + 0.001670246827416122, + 0.0027692553121596575, + 0.0024127860087901354, + 0.0038936296477913857, + 0.002101385733112693, + 0.0024744318798184395, + 0.0027843371499329805, + 0.0021743003744632006, + 0.011253503151237965, + 0.0157158300280571, + 0.03164415434002876, + 0.01682269386947155, + 0.011622274294495583, + 0.01199482660740614, + 0.026879096403717995, + 0.011212728917598724, + 0.011726959608495235, + 0.01498420163989067, + 0.0276987012475729, + 0.023413874208927155, + 0.02059023268520832, + 0.013114790432155132, + 0.020090479403734207, + 0.01625826023519039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.014195770025253296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011809049174189568 + ], + "opposite_logprob_tokens": [ + -6.751181125640869 + ], + "logprob_reward": -0.0011809049174189568, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011809049174189568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13002251088619232, + 0.08087917417287827, + 0.046807195991277695, + 0.12225843966007233, + 0.02303239144384861, + 0.04481028765439987, + 0.05136961117386818, + 0.05226079374551773, + 0.19022700190544128, + 0.25833261013031006 + ], + "credit_map": [ + 0.012137727811932564, + 0.01348140649497509, + 0.015745241194963455, + 0.01259058527648449, + 0.017453407868742943, + 0.013094158843159676, + 0.025892281904816628, + 0.019627705216407776, + 0.009602637961506844, + 0.009417223744094372, + 0.007817734964191914, + 0.01163962110877037, + 0.006632356438785791, + 0.012804023921489716, + 0.009488886222243309, + 0.01347669679671526, + 0.007426189724355936, + 0.0076665026135742664, + 0.0044522713869810104, + 0.004201495088636875, + 0.0060808719135820866, + 0.006380549632012844, + 0.004006073344498873, + 0.006593243684619665, + 0.011635169386863708, + 0.01654953695833683, + 0.017288444563746452, + 0.017455291002988815, + 0.011950792744755745, + 0.01702347956597805, + 0.017831698060035706, + 0.012524028308689594, + 0.002285161754116416, + 0.003880795557051897, + 0.002822178415954113, + 0.0034626475535333157, + 0.002516017062589526, + 0.002593055833131075, + 0.002157921437174082, + 0.003314613364636898, + 0.006254913751035929, + 0.0066126873716712, + 0.006367572583258152, + 0.007202160079032183, + 0.005108772777020931, + 0.005320224445313215, + 0.004560434725135565, + 0.0033835216891020536, + 0.005867590196430683, + 0.008301938883960247, + 0.00796278566122055, + 0.007241441402584314, + 0.005368164274841547, + 0.004604396875947714, + 0.00577143207192421, + 0.006251858081668615, + 0.004845747258514166, + 0.007386246230453253, + 0.007657163310796022, + 0.006642673164606094, + 0.006110910326242447, + 0.0045716105960309505, + 0.008085046894848347, + 0.0069613978266716, + 0.01922728680074215, + 0.025400124490261078, + 0.04352011904120445, + 0.02534588798880577, + 0.014328588731586933, + 0.013126382604241371, + 0.037179842591285706, + 0.012098757550120354, + 0.025893978774547577, + 0.03470733016729355, + 0.03653748333454132, + 0.029675349593162537, + 0.03729228302836418, + 0.03049519471824169, + 0.04047473147511482, + 0.02325626090168953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0011809049174189568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097988814115524 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.038097988814115524, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097988814115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2014700025320053, + 0.11170520633459091, + 0.047951340675354004, + 0.12616394460201263, + 0.020341549068689346, + 0.05146842449903488, + 0.05664545297622681, + 0.12754188477993011, + 0.06921420246362686, + 0.18749792873859406 + ], + "credit_map": [ + 0.014267097227275372, + 0.0175174567848444, + 0.017983101308345795, + 0.03271040320396423, + 0.02467316947877407, + 0.027387117967009544, + 0.03470764681696892, + 0.03222402557730675, + 0.007522303611040115, + 0.008865331299602985, + 0.006080916617065668, + 0.010752362199127674, + 0.009782982990145683, + 0.030429845675826073, + 0.007212615106254816, + 0.03105885721743107, + 0.006210765801370144, + 0.008639178238809109, + 0.004320546984672546, + 0.004796839319169521, + 0.005945471115410328, + 0.006329912692308426, + 0.004847277421504259, + 0.006861352361738682, + 0.014665555208921432, + 0.019189631566405296, + 0.016689732670783997, + 0.01706358790397644, + 0.011117430403828621, + 0.022713258862495422, + 0.01302544865757227, + 0.011699318885803223, + 0.002201527589932084, + 0.0033211896661669016, + 0.0025230697356164455, + 0.003679113695397973, + 0.0023869480937719345, + 0.0019841112662106752, + 0.0020352727733552456, + 0.002210318110883236, + 0.005845237523317337, + 0.006986586377024651, + 0.01009670365601778, + 0.01182025857269764, + 0.004931002855300903, + 0.003241132479161024, + 0.0047385613434016705, + 0.0038089503068476915, + 0.011153267696499825, + 0.007372378371655941, + 0.013883001171052456, + 0.007119466084986925, + 0.005545026622712612, + 0.0037623713724315166, + 0.004337486810982227, + 0.0034724590368568897, + 0.012560163624584675, + 0.016808833926916122, + 0.018106743693351746, + 0.026137428358197212, + 0.01403353177011013, + 0.008136807940900326, + 0.016729149967432022, + 0.015029225498437881, + 0.006897579412907362, + 0.01256128866225481, + 0.015886744484305382, + 0.007561965379863977, + 0.007157389540225267, + 0.004780894611030817, + 0.010982248932123184, + 0.003386096330359578, + 0.01814008317887783, + 0.00903815496712923, + 0.058916885405778885, + 0.03878211975097656, + 0.013504675589501858, + 0.014565293677151203, + 0.016224585473537445, + 0.018326139077544212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.038097988814115524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static shot frames a steel paperclip as it executes a perfect, steady gyroscopic motion, hovering above the smooth surface of a rain-soaked alley puddle. The polished metal in the clip, rendered in stunning high-contrast detail, captures the flickering neon signs of the surrounding dystopian cyberpunk cityscape, its curved bodies twisting vibrant hues of fuchsia, cyan, and deep blue into flowing, dancing streaks. The camera's low perspective emphasizes the impossible physics at play, placing the small, gleaming object against a backdrop of dark, wet asphalt and grimy, towering building facades, gently billowing with steam. A fine mist hangs in the air, suffused with the glaring glow of holographic advertisements, while the puddle below reflects the clip's silent, hypnotic gyrations, creating a surreal and captivating focal point within the gritty, oppressive urban environment, evoking an atmosphere of quiet mystery and technological wonder.", + "seed": 19562, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_9420_seed_19562_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": -0.10389206922845914, + "vlm_logprob": -0.10389206922845914, + "vlm_yes_no_margin": 3.999999988079071, + "vlm_token_credit": -0.10389206922845914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44210463762283325 + ], + "opposite_logprob_tokens": [ + -1.0671045780181885 + ], + "logprob_reward": -0.44210463762283325, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.44210463762283325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2215123176574707, + 0.10069427639245987, + 0.0775260180234909, + 0.12599241733551025, + 0.027321675792336464, + 0.03159618005156517, + 0.05491150915622711, + 0.040117692202329636, + 0.2803346812725067, + 0.03999321162700653 + ], + "credit_map": [ + 0.015626033768057823, + 0.020653944462537766, + 0.016168417409062386, + 0.014734080992639065, + 0.037793226540088654, + 0.04526010528206825, + 0.02512107975780964, + 0.046155430376529694, + 0.011313426308333874, + 0.010550448670983315, + 0.006935541983693838, + 0.008336308412253857, + 0.019001813605427742, + 0.018491949886083603, + 0.00961204245686531, + 0.016452744603157043, + 0.010531652718782425, + 0.010991768911480904, + 0.006474468857049942, + 0.007803075015544891, + 0.016844844445586205, + 0.010129030793905258, + 0.005982278846204281, + 0.008768899366259575, + 0.005919305607676506, + 0.0047888667322695255, + 0.004560463130474091, + 0.004329257179051638, + 0.006975808180868626, + 0.047743458300828934, + 0.007254933472722769, + 0.04442032054066658, + 0.003594915848225355, + 0.0019423244521021843, + 0.0034646496642380953, + 0.0026963637210428715, + 0.002132806461304426, + 0.004721923265606165, + 0.004722255747765303, + 0.00404643639922142, + 0.0038492216262966394, + 0.0026530055329203606, + 0.0035838354378938675, + 0.00248401565477252, + 0.0031523772049695253, + 0.0066322097554802895, + 0.0036003782879561186, + 0.005641135852783918, + 0.0077903554774820805, + 0.0035864640958607197, + 0.006282765418291092, + 0.00561542809009552, + 0.007486043032258749, + 0.009609155356884003, + 0.006093974225223064, + 0.008447326719760895, + 0.006088980473577976, + 0.004155708942562342, + 0.006058428902179003, + 0.004138019401580095, + 0.004532353486865759, + 0.0054259696044027805, + 0.003112431149929762, + 0.006605799775570631, + 0.03439825773239136, + 0.020875845104455948, + 0.030023660510778427, + 0.01897502690553665, + 0.06861840188503265, + 0.03596526011824608, + 0.03051576390862465, + 0.040962472558021545, + 0.005776943638920784, + 0.0047677671536803246, + 0.005362380761653185, + 0.003947088960558176, + 0.003886287799105048, + 0.007486382964998484, + 0.003159522544592619, + 0.005606839433312416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.44210463762283325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871484264731407 + ], + "opposite_logprob_tokens": [ + -3.0487148761749268 + ], + "logprob_reward": -0.04871484264731407, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871484264731407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19216793775558472, + 0.24367505311965942, + 0.08979646861553192, + 0.13223153352737427, + 0.041644807904958725, + 0.048268094658851624, + 0.05114500969648361, + 0.08522343635559082, + 0.0659874677658081, + 0.049860142171382904 + ], + "credit_map": [ + 0.019671346992254257, + 0.029270844534039497, + 0.018085522577166557, + 0.0150228850543499, + 0.0269608311355114, + 0.0331423245370388, + 0.02041730098426342, + 0.02959689497947693, + 0.03469710424542427, + 0.030580000951886177, + 0.021696433424949646, + 0.016170227900147438, + 0.045622240751981735, + 0.040797051042318344, + 0.022265151143074036, + 0.031846851110458374, + 0.008990638889372349, + 0.008995813317596912, + 0.0070831188932061195, + 0.009429388679564, + 0.023141155019402504, + 0.011407456360757351, + 0.010082502849400043, + 0.010666406713426113, + 0.008122405037283897, + 0.008934208191931248, + 0.006078706588596106, + 0.006594025995582342, + 0.010666385293006897, + 0.04257417842745781, + 0.010222798213362694, + 0.03903885558247566, + 0.005924124736338854, + 0.003063357202336192, + 0.004633279517292976, + 0.004030810669064522, + 0.004598375875502825, + 0.007863478735089302, + 0.004553474020212889, + 0.006977907847613096, + 0.004811438266187906, + 0.004375355783849955, + 0.004965557716786861, + 0.004807725083082914, + 0.005920995492488146, + 0.009339572861790657, + 0.006629161536693573, + 0.007418291177600622, + 0.0061385659500956535, + 0.0032522042747586966, + 0.006161999423056841, + 0.006559122819453478, + 0.00544583797454834, + 0.010538076981902122, + 0.004913088865578175, + 0.008136117830872536, + 0.01045860443264246, + 0.006114702206104994, + 0.01434390340000391, + 0.008390706963837147, + 0.007683768402785063, + 0.014108824543654919, + 0.010933431796729565, + 0.013189501129090786, + 0.007312773261219263, + 0.006780190858989954, + 0.006646790076047182, + 0.005273687653243542, + 0.014676008373498917, + 0.008760853670537472, + 0.0074984245002269745, + 0.009038745425641537, + 0.006363503634929657, + 0.004278202541172504, + 0.006301965098828077, + 0.005682065617293119, + 0.007115390617400408, + 0.007281801663339138, + 0.006812568753957748, + 0.006024652160704136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04871484264731407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463933777064085 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.00463933777064085, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463933777064085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26894569396972656, + 0.07107435911893845, + 0.06678317487239838, + 0.040642913430929184, + 0.0770033672451973, + 0.06930127739906311, + 0.17357686161994934, + 0.06344033777713776, + 0.057815879583358765, + 0.11141610145568848 + ], + "credit_map": [ + 0.02383175864815712, + 0.029003934934735298, + 0.03389347344636917, + 0.0297095887362957, + 0.04174758121371269, + 0.036210354417562485, + 0.039312709122896194, + 0.03523631393909454, + 0.008727353066205978, + 0.009122252464294434, + 0.007739429362118244, + 0.005470003932714462, + 0.009283190593123436, + 0.01366163045167923, + 0.005321291275322437, + 0.011749214492738247, + 0.006418860051780939, + 0.006176250521093607, + 0.006430068984627724, + 0.006776566617190838, + 0.014866488054394722, + 0.009227865375578403, + 0.007254069205373526, + 0.009633012115955353, + 0.0054832762107253075, + 0.002726177219301462, + 0.0032631836365908384, + 0.0026570933405309916, + 0.0034189557190984488, + 0.010130184702575207, + 0.0030092839151620865, + 0.009954762645065784, + 0.010401695966720581, + 0.007322653662413359, + 0.00693837134167552, + 0.009206581860780716, + 0.00483713299036026, + 0.012530006468296051, + 0.016359319910407066, + 0.009407608769834042, + 0.011040166020393372, + 0.009224887005984783, + 0.006486237980425358, + 0.005943932570517063, + 0.007312080357223749, + 0.012214447371661663, + 0.007817226462066174, + 0.009262308478355408, + 0.018541598692536354, + 0.013390100561082363, + 0.017032472416758537, + 0.012729772366583347, + 0.016871541738510132, + 0.03918468952178955, + 0.025982186198234558, + 0.02984449453651905, + 0.010785591788589954, + 0.007709619123488665, + 0.006304088514298201, + 0.004842290189117193, + 0.007120517548173666, + 0.01084151491522789, + 0.005765029229223728, + 0.010071692056953907, + 0.008241266943514347, + 0.0038162721320986748, + 0.007168713491410017, + 0.0037814450915902853, + 0.011450804769992828, + 0.008290169760584831, + 0.005815454293042421, + 0.00925175566226244, + 0.01393743883818388, + 0.009633035399019718, + 0.015779919922351837, + 0.010482671670615673, + 0.013311496935784817, + 0.01827690191566944, + 0.015809323638677597, + 0.014185315929353237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.00463933777064085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007188120507635176 + ], + "opposite_logprob_tokens": [ + -7.250718593597412 + ], + "logprob_reward": -0.0007188120507635176, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007188120507635176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15808188915252686, + 0.1240975558757782, + 0.13099788129329681, + 0.10325302928686142, + 0.07705903053283691, + 0.0786065086722374, + 0.08598297834396362, + 0.04364117234945297, + 0.14559638500213623, + 0.05268358439207077 + ], + "credit_map": [ + 0.01611156575381756, + 0.020372316241264343, + 0.012904237024486065, + 0.014848309569060802, + 0.029762273654341698, + 0.02306818775832653, + 0.022977888584136963, + 0.01803709752857685, + 0.01706766150891781, + 0.015724817290902138, + 0.011397499591112137, + 0.00955406203866005, + 0.017180806025862694, + 0.022506890818476677, + 0.012842397205531597, + 0.01782342977821827, + 0.01376771554350853, + 0.014853269793093204, + 0.00931531097739935, + 0.01359133142977953, + 0.03082064352929592, + 0.01903541013598442, + 0.012924245558679104, + 0.016689950600266457, + 0.015743955969810486, + 0.01381809264421463, + 0.007462666369974613, + 0.009749957360327244, + 0.009615647606551647, + 0.021042119711637497, + 0.0066428096033632755, + 0.01917778141796589, + 0.007897850126028061, + 0.008229980245232582, + 0.008375276811420918, + 0.012563621625304222, + 0.006460906006395817, + 0.009057385846972466, + 0.01582161895930767, + 0.00865238904953003, + 0.011316141113638878, + 0.012089764699339867, + 0.01102125272154808, + 0.007112424820661545, + 0.008441407233476639, + 0.010562676005065441, + 0.008255094289779663, + 0.009807748720049858, + 0.014043775387108326, + 0.008577860891819, + 0.00756740290671587, + 0.004659534431993961, + 0.008985442109405994, + 0.0165669247508049, + 0.011084847152233124, + 0.014497194439172745, + 0.006027058232575655, + 0.003489078488200903, + 0.005685870070010424, + 0.004000301472842693, + 0.003449272131547332, + 0.006364357657730579, + 0.008228710852563381, + 0.006396522745490074, + 0.01902407594025135, + 0.009653315879404545, + 0.01757691614329815, + 0.010126831941306591, + 0.03378219157457352, + 0.018358225002884865, + 0.015375496819615364, + 0.021699339151382446, + 0.007868901826441288, + 0.005377668887376785, + 0.007763427682220936, + 0.00545327365398407, + 0.004941822960972786, + 0.008187764324247837, + 0.005311955232173204, + 0.007778769358992577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0007188120507635176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282716050744057 + ], + "opposite_logprob_tokens": [ + -3.773282766342163 + ], + "logprob_reward": -0.023282716050744057, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282716050744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18372298777103424, + 0.08517351001501083, + 0.08647587895393372, + 0.14013010263442993, + 0.036541398614645004, + 0.07834842801094055, + 0.0743948295712471, + 0.06500564515590668, + 0.20035311579704285, + 0.049854062497615814 + ], + "credit_map": [ + 0.01605958119034767, + 0.018389230594038963, + 0.02053791470825672, + 0.014770273119211197, + 0.03141448646783829, + 0.024863243103027344, + 0.028187962248921394, + 0.029500307515263557, + 0.007658252026885748, + 0.010543751530349255, + 0.00734811881557107, + 0.00925326906144619, + 0.018001902848482132, + 0.013335911557078362, + 0.010452670976519585, + 0.008579636923968792, + 0.013876255601644516, + 0.007583692669868469, + 0.008103538304567337, + 0.006682942155748606, + 0.014302686788141727, + 0.014399991370737553, + 0.00903682317584753, + 0.012489952147006989, + 0.01416689995676279, + 0.010934021323919296, + 0.007952242158353329, + 0.006881659384816885, + 0.013493414036929607, + 0.03610365837812424, + 0.016430648043751717, + 0.034167561680078506, + 0.004936030600219965, + 0.00526073994114995, + 0.0043348390609025955, + 0.005116283893585205, + 0.0028740346897393465, + 0.005604968406260014, + 0.0037866835482418537, + 0.004627820570021868, + 0.00824892707169056, + 0.007433407474309206, + 0.012976329773664474, + 0.008760503493249416, + 0.005967657081782818, + 0.012671898119151592, + 0.010162284597754478, + 0.012127422727644444, + 0.01653105393052101, + 0.00643569091334939, + 0.007969210855662823, + 0.006818235851824284, + 0.006130029913038015, + 0.010544832795858383, + 0.008004697971045971, + 0.011961087584495544, + 0.010847892612218857, + 0.0071343122981488705, + 0.007694349158555269, + 0.005664754658937454, + 0.007196909748017788, + 0.00813477672636509, + 0.009125585667788982, + 0.009207069873809814, + 0.023838886991143227, + 0.01506172027438879, + 0.026241829618811607, + 0.015228926204144955, + 0.0461181215941906, + 0.023952791467308998, + 0.02157377265393734, + 0.028337085619568825, + 0.007321331650018692, + 0.00546545535326004, + 0.006632206495851278, + 0.004087173845618963, + 0.0059160394594073296, + 0.008246528916060925, + 0.00567836407572031, + 0.006506965495646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.023282716050744057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static shot frames a steel paperclip as it executes a perfect, steady gyroscopic motion, hovering above the smooth surface of a rain-soaked alley puddle. The polished metal in the clip, rendered in stunning high-contrast detail, captures the flickering neon signs of the surrounding dystopian cyberpunk cityscape, its curved bodies twisting vibrant hues of fuchsia, cyan, and deep blue into flowing, dancing streaks. The camera's low perspective emphasizes the impossible physics at play, placing the small, gleaming object against a backdrop of dark, wet asphalt and grimy, towering building facades, gently billowing with steam. A fine mist hangs in the air, suffused with the glaring glow of holographic advertisements, while the puddle below reflects the clip's silent, hypnotic gyrations, creating a surreal and captivating focal point within the gritty, oppressive urban environment, evoking an atmosphere of quiet mystery and technological wonder.", + "seed": 19563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_9420_seed_19563_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": -0.09214195194654166, + "vlm_logprob": -0.09214195194654166, + "vlm_yes_no_margin": 3.799999988079071, + "vlm_token_credit": -0.09214195194654166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35848504304885864 + ], + "opposite_logprob_tokens": [ + -1.2334849834442139 + ], + "logprob_reward": -0.35848504304885864, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.35848504304885864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1539342850446701, + 0.06949475407600403, + 0.08565831929445267, + 0.2280939519405365, + 0.03327269107103348, + 0.08723223954439163, + 0.13687868416309357, + 0.0788724422454834, + 0.10230297595262527, + 0.024259688332676888 + ], + "credit_map": [ + 0.014472467824816704, + 0.01189353596419096, + 0.018468068912625313, + 0.016978751868009567, + 0.017884351313114166, + 0.02262413688004017, + 0.030973859131336212, + 0.020639119669795036, + 0.005011135712265968, + 0.007577917072921991, + 0.008221013471484184, + 0.006526934448629618, + 0.007103030104190111, + 0.011576374992728233, + 0.011930039152503014, + 0.011548308655619621, + 0.007929010316729546, + 0.014909581281244755, + 0.009506914764642715, + 0.008283184841275215, + 0.008954263292253017, + 0.01204672921448946, + 0.010709581896662712, + 0.013319051824510098, + 0.020837148651480675, + 0.0312193650752306, + 0.018515683710575104, + 0.019174300134181976, + 0.04446425288915634, + 0.024309124797582626, + 0.041507720947265625, + 0.028066348284482956, + 0.0038265802431851625, + 0.004069917369633913, + 0.003182177199050784, + 0.0026030319277197123, + 0.0044038365595042706, + 0.005673412699252367, + 0.004726171027868986, + 0.004787564743310213, + 0.007183815352618694, + 0.005511322990059853, + 0.00755487522110343, + 0.007020339369773865, + 0.011907300911843777, + 0.010746631771326065, + 0.022958481684327126, + 0.014349477365612984, + 0.011043227277696133, + 0.008437719196081161, + 0.012261178344488144, + 0.018702425062656403, + 0.016203701496124268, + 0.016478978097438812, + 0.030577659606933594, + 0.023173782974481583, + 0.005540840793401003, + 0.005137724801898003, + 0.007449739146977663, + 0.009067193605005741, + 0.010029580444097519, + 0.008699691854417324, + 0.019385095685720444, + 0.013562572188675404, + 0.008169429376721382, + 0.006331117358058691, + 0.008937896229326725, + 0.009096725843846798, + 0.015436561778187752, + 0.018380288034677505, + 0.019083835184574127, + 0.016867121681571007, + 0.00280360272154212, + 0.002148222643882036, + 0.003262045094743371, + 0.0019690559711307287, + 0.002087557455524802, + 0.003710772143676877, + 0.004922644700855017, + 0.0033557892311364412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.35848504304885864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209971010684967 + ], + "opposite_logprob_tokens": [ + -2.8120996952056885 + ], + "logprob_reward": -0.06209971010684967, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209971010684967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20344635844230652, + 0.10857974737882614, + 0.10277898609638214, + 0.2346567064523697, + 0.04493321105837822, + 0.0567668192088604, + 0.1572858989238739, + 0.033378541469573975, + 0.020662205293774605, + 0.03751149773597717 + ], + "credit_map": [ + 0.015289305709302425, + 0.012415723875164986, + 0.030418381094932556, + 0.02273339033126831, + 0.02402924746274948, + 0.02843143790960312, + 0.04451625421643257, + 0.025612639263272285, + 0.010274987667798996, + 0.01121860183775425, + 0.01101069524884224, + 0.008280050940811634, + 0.011709347367286682, + 0.019792545586824417, + 0.01855812780559063, + 0.017735393717885017, + 0.009956926107406616, + 0.0174256544560194, + 0.01182154007256031, + 0.009507372975349426, + 0.012603743001818657, + 0.01461054291576147, + 0.010949477553367615, + 0.015903737396001816, + 0.020409900695085526, + 0.030332183465361595, + 0.017582803964614868, + 0.019165096804499626, + 0.04081863537430763, + 0.03318190202116966, + 0.04104584455490112, + 0.03212035819888115, + 0.006551113910973072, + 0.0042583998292684555, + 0.003969626035541296, + 0.0034569280687719584, + 0.005349054932594299, + 0.007574252784252167, + 0.007436106447130442, + 0.006337732542306185, + 0.004422780591994524, + 0.004686696920543909, + 0.004505148623138666, + 0.004090204834938049, + 0.00852508470416069, + 0.0066931708715856075, + 0.014683512039482594, + 0.009160226210951805, + 0.01036305446177721, + 0.008982140570878983, + 0.011058520525693893, + 0.016682414337992668, + 0.02104194462299347, + 0.018761835992336273, + 0.042746879160404205, + 0.02764911577105522, + 0.003506898880004883, + 0.002558482810854912, + 0.0034733109641820192, + 0.003886251477524638, + 0.003920030780136585, + 0.004171579144895077, + 0.006855672225356102, + 0.005006314720958471, + 0.0016135453479364514, + 0.0016087109688669443, + 0.002179164905101061, + 0.0025080982595682144, + 0.0019326108740642667, + 0.0028538431506603956, + 0.004978844430297613, + 0.0029873885214328766, + 0.0021265815012156963, + 0.0032851197756826878, + 0.0032884718384593725, + 0.003469109535217285, + 0.005485157947987318, + 0.005334919318556786, + 0.009377976879477501, + 0.005144163966178894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.06209971010684967 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251897498965263 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251897498965263, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251897498965263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21272481977939606, + 0.0957818329334259, + 0.06556802242994308, + 0.13200674951076508, + 0.047597337514162064, + 0.12400563806295395, + 0.20413148403167725, + 0.06197386980056763, + 0.0355168916285038, + 0.020693406462669373 + ], + "credit_map": [ + 0.018391286954283714, + 0.0133203761652112, + 0.029716456308960915, + 0.0281910952180624, + 0.017786329612135887, + 0.040570031851530075, + 0.031386133283376694, + 0.03336308151483536, + 0.006981867831200361, + 0.010548330843448639, + 0.009885773994028568, + 0.007918172515928745, + 0.011790662072598934, + 0.015055857598781586, + 0.020971624180674553, + 0.012629539705812931, + 0.00531531497836113, + 0.00958186388015747, + 0.006625196896493435, + 0.006507429759949446, + 0.007336278446018696, + 0.010191135108470917, + 0.009723703376948833, + 0.010287094861268997, + 0.010480569675564766, + 0.018672209233045578, + 0.010715387761592865, + 0.010068369098007679, + 0.024127725511789322, + 0.01791395992040634, + 0.022464973852038383, + 0.01756354607641697, + 0.005538989324122667, + 0.003968521952629089, + 0.006068023853003979, + 0.004035792779177427, + 0.005605571903288364, + 0.0069630444049835205, + 0.008599776774644852, + 0.006817616522312164, + 0.012015672400593758, + 0.010105925612151623, + 0.012832139618694782, + 0.010275254026055336, + 0.01585373282432556, + 0.016902495175600052, + 0.028005246073007584, + 0.018015164881944656, + 0.01774810627102852, + 0.012388383969664574, + 0.016548100858926773, + 0.024070965126156807, + 0.023459043353796005, + 0.022522471845149994, + 0.05224812030792236, + 0.03514627739787102, + 0.004540805239230394, + 0.0040023610927164555, + 0.006078529171645641, + 0.008039315231144428, + 0.007300234399735928, + 0.006798771675676107, + 0.014438237994909286, + 0.010775616392493248, + 0.0027438951656222343, + 0.0025685608852654696, + 0.003420370165258646, + 0.003986133262515068, + 0.005461592227220535, + 0.004965892527252436, + 0.006930920761078596, + 0.005439524073153734, + 0.0024271581787616014, + 0.0024423524737358093, + 0.002331510419026017, + 0.002086882246658206, + 0.0022106249816715717, + 0.002602681051939726, + 0.004126282874494791, + 0.0024659123737365007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.005251897498965263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011803095694631338 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011803095694631338, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011803095694631338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1662837564945221, + 0.12958623468875885, + 0.09611739218235016, + 0.19919127225875854, + 0.03655844181776047, + 0.06957263499498367, + 0.07373347133398056, + 0.08889704942703247, + 0.055278584361076355, + 0.08478114008903503 + ], + "credit_map": [ + 0.017971182242035866, + 0.013225533999502659, + 0.023606684058904648, + 0.01699334941804409, + 0.021894237026572227, + 0.02279798686504364, + 0.029116477817296982, + 0.020678309723734856, + 0.01316350232809782, + 0.02085723541676998, + 0.012254451401531696, + 0.011522244662046432, + 0.014218450523912907, + 0.0194773580878973, + 0.019269771873950958, + 0.01882322132587433, + 0.00875688437372446, + 0.006990686058998108, + 0.012652267701923847, + 0.010240193456411362, + 0.008060639724135399, + 0.01680176891386509, + 0.01498833391815424, + 0.017626618966460228, + 0.018234627321362495, + 0.029545875266194344, + 0.016701703891158104, + 0.01667044498026371, + 0.034188125282526016, + 0.025561679154634476, + 0.03310016170144081, + 0.02518865093588829, + 0.002981225261464715, + 0.005203971639275551, + 0.0032120314426720142, + 0.0026753763668239117, + 0.005428404547274113, + 0.005147223826497793, + 0.006404748652130365, + 0.005505458451807499, + 0.009406928904354572, + 0.006793494801968336, + 0.007159266155213118, + 0.006032759323716164, + 0.008069577626883984, + 0.011077399365603924, + 0.011390691623091698, + 0.009642518125474453, + 0.008921676315367222, + 0.0044892025180161, + 0.006537077948451042, + 0.008280039764940739, + 0.00860573723912239, + 0.007994472980499268, + 0.017363321036100388, + 0.011541948653757572, + 0.006084919907152653, + 0.0074329436756670475, + 0.008195278234779835, + 0.013158167712390423, + 0.010330747812986374, + 0.009097700007259846, + 0.020152555778622627, + 0.014444735832512379, + 0.005091656930744648, + 0.004899138119071722, + 0.006004865746945143, + 0.004520439077168703, + 0.008378490805625916, + 0.008897126652300358, + 0.0092976288869977, + 0.008189239539206028, + 0.009893168695271015, + 0.010668964125216007, + 0.01038068812340498, + 0.008847708813846111, + 0.010148031637072563, + 0.010517511516809464, + 0.014190531335771084, + 0.010134536772966385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0011803095694631338 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033692799508571625 + ], + "opposite_logprob_tokens": [ + -3.4086928367614746 + ], + "logprob_reward": -0.033692799508571625, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033692799508571625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09557297080755234, + 0.1476876586675644, + 0.11852139979600906, + 0.2629375755786896, + 0.06785613298416138, + 0.052466366440057755, + 0.05750415474176407, + 0.08906514942646027, + 0.03780100494623184, + 0.07058759033679962 + ], + "credit_map": [ + 0.009427434764802456, + 0.008837733417749405, + 0.010421985760331154, + 0.01187389250844717, + 0.015261758118867874, + 0.015731366351246834, + 0.011707750149071217, + 0.012311052531003952, + 0.012499891221523285, + 0.015501145273447037, + 0.013927152380347252, + 0.018376469612121582, + 0.014334464445710182, + 0.019932419061660767, + 0.03245817869901657, + 0.020657936111092567, + 0.010611720383167267, + 0.01993374712765217, + 0.014848651364445686, + 0.01223608385771513, + 0.015489443205296993, + 0.017284773290157318, + 0.012022238224744797, + 0.016094738617539406, + 0.021990004926919937, + 0.03614417836070061, + 0.024815984070301056, + 0.021686432883143425, + 0.045403722673654556, + 0.034088097512722015, + 0.04307058826088905, + 0.03573857247829437, + 0.007790316361933947, + 0.005799437407404184, + 0.008688988164067268, + 0.006283447612076998, + 0.009140921756625175, + 0.010507293045520782, + 0.01061137393116951, + 0.00903435330837965, + 0.008550023660063744, + 0.005634843837469816, + 0.004486204590648413, + 0.004638698883354664, + 0.006184005178511143, + 0.005029346328228712, + 0.009808513335883617, + 0.00813472829759121, + 0.007416082080453634, + 0.005694667808711529, + 0.004882400389760733, + 0.004780143033713102, + 0.010216576978564262, + 0.007314356975257397, + 0.009031571447849274, + 0.00816835556179285, + 0.006524702068418264, + 0.005926903802901506, + 0.0081837959587574, + 0.012515312060713768, + 0.01052155252546072, + 0.010581135749816895, + 0.020666731521487236, + 0.0141450185328722, + 0.0053329686634242535, + 0.003112061647698283, + 0.005066444631665945, + 0.0035357940942049026, + 0.004752052016556263, + 0.005594891030341387, + 0.005192139185965061, + 0.00521465577185154, + 0.0056913793087005615, + 0.005577585194259882, + 0.011492224410176277, + 0.005764828063547611, + 0.007055021356791258, + 0.008595773950219154, + 0.015812136232852936, + 0.010598641820251942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.033692799508571625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, low-angle, static shot frames a steel paperclip as it executes a perfect, steady gyroscopic motion, hovering above the smooth surface of a rain-soaked alley puddle. The polished metal in the clip, rendered in stunning high-contrast detail, captures the flickering neon signs of the surrounding dystopian cyberpunk cityscape, its curved bodies twisting vibrant hues of fuchsia, cyan, and deep blue into flowing, dancing streaks. The camera's low perspective emphasizes the impossible physics at play, placing the small, gleaming object against a backdrop of dark, wet asphalt and grimy, towering building facades, gently billowing with steam. A fine mist hangs in the air, suffused with the glaring glow of holographic advertisements, while the puddle below reflects the clip's silent, hypnotic gyrations, creating a surreal and captivating focal point within the gritty, oppressive urban environment, evoking an atmosphere of quiet mystery and technological wonder.", + "seed": 19564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_12/index_9420_seed_19564_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": -0.16166720144683494, + "vlm_logprob": -0.16166720144683494, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.16166720144683494, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7023047208786011 + ], + "opposite_logprob_tokens": [ + -0.7023047208786011 + ], + "logprob_reward": -0.7023047208786011, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7023047208786011, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2540469467639923, + 0.04777342081069946, + 0.043284401297569275, + 0.03774084150791168, + 0.06122921034693718, + 0.07400733977556229, + 0.288289874792099, + 0.1000993400812149, + 0.03571303188800812, + 0.05781562253832817 + ], + "credit_map": [ + 0.028871305286884308, + 0.046303048729896545, + 0.03536776825785637, + 0.032568398863077164, + 0.02967066504061222, + 0.03368854150176048, + 0.022933827713131905, + 0.024643393233418465, + 0.005541343241930008, + 0.006013193633407354, + 0.006387715227901936, + 0.005801195744425058, + 0.0038661842700093985, + 0.005007003899663687, + 0.004772321321070194, + 0.010384462773799896, + 0.006348425056785345, + 0.006419084966182709, + 0.008493487723171711, + 0.005888848565518856, + 0.004106077365577221, + 0.004448742140084505, + 0.0038059393409639597, + 0.003773797070607543, + 0.004345029126852751, + 0.003825035411864519, + 0.006151392124593258, + 0.005712665617465973, + 0.002819692948833108, + 0.003130799625068903, + 0.0071681649424135685, + 0.004588059615343809, + 0.005904064513742924, + 0.006416131276637316, + 0.010081330314278603, + 0.012859598733484745, + 0.003906827885657549, + 0.004527285695075989, + 0.0058066751807928085, + 0.01172730140388012, + 0.006867722142487764, + 0.0055753677152097225, + 0.01334547158330679, + 0.013910998590290546, + 0.0034263983834534883, + 0.006387441419064999, + 0.007610378321260214, + 0.016883565112948418, + 0.025533979758620262, + 0.04605108126997948, + 0.031840983778238297, + 0.04844387248158455, + 0.031659021973609924, + 0.0268343947827816, + 0.05373386666178703, + 0.024192677810788155, + 0.010226881131529808, + 0.012234731577336788, + 0.01494914386421442, + 0.017658358439803123, + 0.009129314683377743, + 0.008716648444533348, + 0.00785118993371725, + 0.019333073869347572, + 0.0028178775683045387, + 0.004281768575310707, + 0.004903609398752451, + 0.007557624485343695, + 0.0031385270413011312, + 0.004337677266448736, + 0.003984690178185701, + 0.004691257141530514, + 0.0072669037617743015, + 0.006740186829119921, + 0.007779786363244057, + 0.010646478272974491, + 0.00583159364759922, + 0.007156983949244022, + 0.0066408258862793446, + 0.005752866622060537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.7023047208786011 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006405293941498 + ], + "opposite_logprob_tokens": [ + -2.695064067840576 + ], + "logprob_reward": -0.07006405293941498, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006405293941498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2556714713573456, + 0.04502582177519798, + 0.15453267097473145, + 0.0774473175406456, + 0.04578957334160805, + 0.05698271840810776, + 0.17877505719661713, + 0.09077965468168259, + 0.060929957777261734, + 0.03406578674912453 + ], + "credit_map": [ + 0.028340192511677742, + 0.03334244340658188, + 0.031478364020586014, + 0.03131992369890213, + 0.04676007479429245, + 0.03140854090452194, + 0.026890059933066368, + 0.026131868362426758, + 0.004567874129861593, + 0.005983404815196991, + 0.00610702857375145, + 0.006562949623912573, + 0.0042185294441878796, + 0.0044696638360619545, + 0.005567713174968958, + 0.0075486572459340096, + 0.017631594091653824, + 0.01837906427681446, + 0.029574789106845856, + 0.03602168709039688, + 0.013649898581206799, + 0.009659508243203163, + 0.015908174216747284, + 0.013707943260669708, + 0.009035010822117329, + 0.008545851334929466, + 0.011459959670901299, + 0.010493576526641846, + 0.00825541466474533, + 0.010723589919507504, + 0.011256438679993153, + 0.007677479647099972, + 0.005241259001195431, + 0.005071796011179686, + 0.006552219390869141, + 0.007315800990909338, + 0.005100545939058065, + 0.0039664385840296745, + 0.005858608987182379, + 0.006682908162474632, + 0.005593296606093645, + 0.004499081056565046, + 0.008110511116683483, + 0.015678957104682922, + 0.003164974506944418, + 0.005367457866668701, + 0.005095792468637228, + 0.009472648613154888, + 0.015272886492311954, + 0.024186424911022186, + 0.022628171369433403, + 0.033361200243234634, + 0.01759314350783825, + 0.013130223378539085, + 0.03802426904439926, + 0.014578722417354584, + 0.00711316242814064, + 0.007608556188642979, + 0.012728020548820496, + 0.015478424727916718, + 0.0066849286668002605, + 0.010295973159372807, + 0.011400998570024967, + 0.019469594582915306, + 0.004111859947443008, + 0.005540646146982908, + 0.008989892899990082, + 0.014286809600889683, + 0.004369652830064297, + 0.004135825205594301, + 0.006046983413398266, + 0.01344828587025404, + 0.003979349508881569, + 0.003721121931448579, + 0.0070380703546106815, + 0.006180969998240471, + 0.0024792728945612907, + 0.0031622524838894606, + 0.0037029904779046774, + 0.0038017581682652235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.07006405293941498 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251778755337 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251778755337, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251778755337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1799301952123642, + 0.0623709037899971, + 0.09855947643518448, + 0.06318245083093643, + 0.06877650320529938, + 0.06901279091835022, + 0.23341169953346252, + 0.13315270841121674, + 0.04293128848075867, + 0.04867199435830116 + ], + "credit_map": [ + 0.017223309725522995, + 0.02757512591779232, + 0.024631621316075325, + 0.025208408012986183, + 0.02011692523956299, + 0.028019996359944344, + 0.015562789514660835, + 0.021592015400528908, + 0.008123517967760563, + 0.008630313910543919, + 0.009556249715387821, + 0.007364779245108366, + 0.005537583492696285, + 0.006135432980954647, + 0.007071203552186489, + 0.009951823391020298, + 0.015478170476853848, + 0.011887744069099426, + 0.022906258702278137, + 0.01053945068269968, + 0.00934789888560772, + 0.01332717016339302, + 0.007040674332529306, + 0.00803210586309433, + 0.006758084520697594, + 0.006465161219239235, + 0.0093925129622221, + 0.013150840066373348, + 0.004605783149600029, + 0.006256340071558952, + 0.011648179963231087, + 0.004905551206320524, + 0.006868026684969664, + 0.007234781514853239, + 0.011236202903091908, + 0.019495749846100807, + 0.004350963048636913, + 0.005429090932011604, + 0.00800766795873642, + 0.006154020316898823, + 0.007259332109242678, + 0.004397771321237087, + 0.014542781747877598, + 0.014708779752254486, + 0.0033157491125166416, + 0.00531252846121788, + 0.0071711004711687565, + 0.012304741889238358, + 0.025656752288341522, + 0.04901351407170296, + 0.03022862784564495, + 0.03296538442373276, + 0.020054254680871964, + 0.020465431734919548, + 0.032908909022808075, + 0.022118834778666496, + 0.01247425563633442, + 0.01589573733508587, + 0.017777759581804276, + 0.02121274173259735, + 0.017084360122680664, + 0.019679535180330276, + 0.014672922901809216, + 0.014355395920574665, + 0.003346006153151393, + 0.0037149707786738873, + 0.006404727231711149, + 0.010692285373806953, + 0.00253479927778244, + 0.0035057207569479942, + 0.004965511616319418, + 0.007767264265567064, + 0.004886717069894075, + 0.004786074161529541, + 0.01319494005292654, + 0.007312868721783161, + 0.002784998156130314, + 0.003182819811627269, + 0.003012404777109623, + 0.009511171840131283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.005251778755337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009211826254613698 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009211826254613698, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009211826254613698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2619282603263855, + 0.05498186871409416, + 0.05302578583359718, + 0.04822792857885361, + 0.21589180827140808, + 0.04489261284470558, + 0.13204434514045715, + 0.08110279589891434, + 0.04601290449500084, + 0.06189162656664848 + ], + "credit_map": [ + 0.027655156329274178, + 0.034384891390800476, + 0.03368610888719559, + 0.03467869013547897, + 0.042810726910829544, + 0.030778763815760612, + 0.03125372156500816, + 0.02668019011616707, + 0.008331631310284138, + 0.005115365609526634, + 0.008575878106057644, + 0.005648819729685783, + 0.0042321402579545975, + 0.004287660121917725, + 0.007250438444316387, + 0.011539937928318977, + 0.00596263725310564, + 0.006791961379349232, + 0.00829379353672266, + 0.006396864540874958, + 0.008746166713535786, + 0.006640092935413122, + 0.004620735067874193, + 0.005573537666350603, + 0.005948565900325775, + 0.0058262827806174755, + 0.00686283316463232, + 0.007385780569165945, + 0.004844113253057003, + 0.006875849328935146, + 0.0062901731580495834, + 0.004194331355392933, + 0.020478054881095886, + 0.020335836336016655, + 0.03335242345929146, + 0.05259144678711891, + 0.019931448623538017, + 0.01830826699733734, + 0.0267574954777956, + 0.024136841297149658, + 0.004737334791570902, + 0.005379472393542528, + 0.006347158923745155, + 0.00675389776006341, + 0.0033378698863089085, + 0.0037847463972866535, + 0.004411593079566956, + 0.010140541009604931, + 0.01474783942103386, + 0.025519976392388344, + 0.014278563670814037, + 0.022228654474020004, + 0.012069301679730415, + 0.01114675309509039, + 0.0216675978153944, + 0.010385650210082531, + 0.008064541034400463, + 0.006520396564155817, + 0.012839870527386665, + 0.010533194057643414, + 0.0045643155463039875, + 0.008437247946858406, + 0.006735216360539198, + 0.023408006876707077, + 0.0052589536644518375, + 0.004965787287801504, + 0.007396444212645292, + 0.008107955567538738, + 0.002847940195351839, + 0.003848163178190589, + 0.0038885846734046936, + 0.009699076414108276, + 0.00457370188087225, + 0.0030354750342667103, + 0.01846170425415039, + 0.00903997104614973, + 0.0027166015934199095, + 0.002871783683076501, + 0.0051003508269786835, + 0.01609204337000847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0009211826254613698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a steel paperclip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paperclip executing a steady gyroscopic motion while hovering above a puddle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a rain-soaked alley with dark wet asphalt and grimy towering building facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature a dystopian cyberpunk cityscape with flickering neon signs in fuchsia, cyan, and deep blue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed from a low angle with a static camera position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794272035360336 + ], + "opposite_logprob_tokens": [ + -3.529794216156006 + ], + "logprob_reward": -0.029794272035360336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794272035360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15337464213371277, + 0.0326479971408844, + 0.05847498029470444, + 0.07758835703134537, + 0.06689140945672989, + 0.0665166825056076, + 0.10250349342823029, + 0.17592380940914154, + 0.2227085530757904, + 0.04337000101804733 + ], + "credit_map": [ + 0.016242995858192444, + 0.019157851114869118, + 0.018561067059636116, + 0.02029660902917385, + 0.03080858662724495, + 0.018911734223365784, + 0.01335669681429863, + 0.016039106994867325, + 0.004291244316846132, + 0.003827004926279187, + 0.004668125882744789, + 0.004700100980699062, + 0.002632739255204797, + 0.0029277545399963856, + 0.004404174629598856, + 0.005196855403482914, + 0.010323821566998959, + 0.0070746177807450294, + 0.014310872182250023, + 0.00632296921685338, + 0.004202938172966242, + 0.0075078243389725685, + 0.004502781666815281, + 0.004229165147989988, + 0.008882997557520866, + 0.009944611229002476, + 0.01231429073959589, + 0.011181352660059929, + 0.008569816127419472, + 0.007952147163450718, + 0.010348008945584297, + 0.008395139127969742, + 0.006937897764146328, + 0.007078200578689575, + 0.011274047195911407, + 0.016413941979408264, + 0.004596381913870573, + 0.005083530675619841, + 0.0077305566519498825, + 0.007776857819408178, + 0.004928504582494497, + 0.004732691682875156, + 0.0116743678227067, + 0.01652705855667591, + 0.003549838438630104, + 0.0049690259620547295, + 0.00620732456445694, + 0.013927873224020004, + 0.00882019754499197, + 0.01495105680078268, + 0.012606672011315823, + 0.015870163217186928, + 0.009515673853456974, + 0.009231225587427616, + 0.019290298223495483, + 0.012218213640153408, + 0.018509091809391975, + 0.01210485678166151, + 0.025869322940707207, + 0.026083383709192276, + 0.01090157963335514, + 0.023366227746009827, + 0.017019221559166908, + 0.04207014665007591, + 0.017486346885561943, + 0.011811256408691406, + 0.03739577904343605, + 0.06357020139694214, + 0.009699594229459763, + 0.012614766135811806, + 0.02098977193236351, + 0.04914084076881409, + 0.004522399045526981, + 0.0038666357286274433, + 0.011502739042043686, + 0.007496529724448919, + 0.0022351553197950125, + 0.002856654580682516, + 0.002746925223618746, + 0.008142967708408833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.029794272035360336 + } + ], + "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/0000158/group_13/index_tensor([5135, 5135])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_tensor([5135, 5135])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0896ce008e8110ab02532f537e3a7c3ebef9954 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_tensor([5135, 5135])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a low-angle shot, the camera smoothly and deliberately sliding from left to right, revealing a stunning concept sports car on display at a prestigious auto show. Perched on a minimalist, polished black circular platform, the car boasts an incredibly sleek silhouette, painted a deep Candy Apple Red with a flawless, mirror-like metallic finish. Bright, focused spotlights from overhead rigs traverse its aerodynamic curves, creating sharp, moving highlights and emphasizing its radical design, including a visible carbon fiber diffuser and large, intricate alloy wheels. The surroundings of the vast showroom, deliberately darker and cooler in tone, the blurred figures of attendees and the ambient glow of the exhibits in the distance create a sense of scale and excitement. The entire scene is imbued with an aura of avant-garde luxury and anticipation, the camera's reverent movements and the dramatic play of light and shadow working together to present the car not just as a machine but as a work of art.", + "seed": 8189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8189_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": -0.03058434762060642, + "vlm_logprob": -0.03058434762060642, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.03058434762060642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058297645300626755 + ], + "opposite_logprob_tokens": [ + -4.683297634124756 + ], + "logprob_reward": -0.058297645300626755, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.058297645300626755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08165369927883148, + 0.08509492874145508, + 0.05754416063427925, + 0.09365981817245483, + 0.06626496464014053, + 0.34989166259765625, + 0.08770428597927094, + 0.04319796338677406, + 0.033857062458992004, + 0.10113149136304855 + ], + "credit_map": [ + 0.008193868212401867, + 0.009052150882780552, + 0.008409260772168636, + 0.011579753831028938, + 0.007818878628313541, + 0.009328796528279781, + 0.015008765272796154, + 0.012262219563126564, + 0.007360700983554125, + 0.006072946824133396, + 0.012389753013849258, + 0.01143390778452158, + 0.009340180084109306, + 0.018594538792967796, + 0.009762142784893513, + 0.01014076266437769, + 0.003955252468585968, + 0.004143045283854008, + 0.004317678511142731, + 0.006790028885006905, + 0.01317877322435379, + 0.012998787686228752, + 0.005740404594689608, + 0.006420188117772341, + 0.006318270228803158, + 0.0063662840984761715, + 0.015760058537125587, + 0.016026120632886887, + 0.010562688112258911, + 0.008492358960211277, + 0.012847958132624626, + 0.017286069691181183, + 0.006174784153699875, + 0.008894834667444229, + 0.01064135879278183, + 0.012202420271933079, + 0.004834309685975313, + 0.005466577131301165, + 0.009979826398193836, + 0.008070849813520908, + 0.04371946305036545, + 0.01896258071064949, + 0.030696462839841843, + 0.02209417335689068, + 0.026143670082092285, + 0.022080549970269203, + 0.09218449145555496, + 0.09401025623083115, + 0.011118839494884014, + 0.007474976591765881, + 0.011909944005310535, + 0.011156616732478142, + 0.005828304681926966, + 0.012677645310759544, + 0.009988385252654552, + 0.01754956692457199, + 0.003817559452727437, + 0.0028337950352579355, + 0.006343034096062183, + 0.005078611895442009, + 0.0038868922274559736, + 0.009986207820475101, + 0.005689321085810661, + 0.005562539212405682, + 0.002685523359104991, + 0.0023998902179300785, + 0.004760392010211945, + 0.0036455586086958647, + 0.004432029090821743, + 0.00786641426384449, + 0.004116091877222061, + 0.003951163496822119, + 0.014444944448769093, + 0.027565494179725647, + 0.01045366283506155, + 0.007278380915522575, + 0.0054944781586527824, + 0.008839180693030357, + 0.009726407006382942, + 0.01732894591987133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 3 + ], + "selected_reward": -0.058297645300626755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004271447658539 + ], + "opposite_logprob_tokens": [ + -2.695042610168457 + ], + "logprob_reward": -0.07004271447658539, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004271447658539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07270102202892303, + 0.06376732140779495, + 0.05326063930988312, + 0.0791679099202156, + 0.02497590333223343, + 0.27502208948135376, + 0.02151748724281788, + 0.028676563873887062, + 0.07409942895174026, + 0.3068116009235382 + ], + "credit_map": [ + 0.008313024416565895, + 0.008964784443378448, + 0.009433849714696407, + 0.010596142150461674, + 0.007170944940298796, + 0.00829275231808424, + 0.01112999115139246, + 0.008799527771770954, + 0.006095753517001867, + 0.004459913820028305, + 0.009577877819538116, + 0.006632171105593443, + 0.008040782064199448, + 0.01175775844603777, + 0.006451823748648167, + 0.010751240886747837, + 0.002706970786675811, + 0.0027248591650277376, + 0.0037820269353687763, + 0.005850803107023239, + 0.010762831196188927, + 0.015433419495821, + 0.004715348593890667, + 0.007284378632903099, + 0.009484091773629189, + 0.004756410606205463, + 0.017932191491127014, + 0.01450972817838192, + 0.006612792611122131, + 0.007178504951298237, + 0.009088653139770031, + 0.009605537168681622, + 0.002802591072395444, + 0.004484641831368208, + 0.0034218281507492065, + 0.0036941245198249817, + 0.002407332183793187, + 0.0018522454192861915, + 0.003636052366346121, + 0.002677089534699917, + 0.03837892785668373, + 0.021559910848736763, + 0.027800455689430237, + 0.022668104618787766, + 0.021779464557766914, + 0.01748417690396309, + 0.05872375890612602, + 0.06662729382514954, + 0.002162921242415905, + 0.00224506389349699, + 0.0025272201746702194, + 0.003113153390586376, + 0.0018752939067780972, + 0.003310718573629856, + 0.003471140284091234, + 0.0028119762428104877, + 0.002628396498039365, + 0.002025813329964876, + 0.004195644054561853, + 0.004150977823883295, + 0.0025684067513793707, + 0.004845270421355963, + 0.0039006699807941914, + 0.004361387807875872, + 0.006108274683356285, + 0.005354755558073521, + 0.015830013900995255, + 0.012953869067132473, + 0.006318586878478527, + 0.010212486609816551, + 0.008679014630615711, + 0.008642424829304218, + 0.06645458191633224, + 0.04872974753379822, + 0.02630491741001606, + 0.014823658391833305, + 0.01652505248785019, + 0.03269730508327484, + 0.03643852099776268, + 0.06483784317970276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 3 + ], + "selected_reward": -0.07004271447658539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015235256869345903 + ], + "opposite_logprob_tokens": [ + -6.501523494720459 + ], + "logprob_reward": -0.0015235256869345903, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015235256869345903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055745840072631836, + 0.10967998951673508, + 0.031099170446395874, + 0.06644917279481888, + 0.04333663359284401, + 0.4848695397377014, + 0.07274229824542999, + 0.0770203024148941, + 0.0393526516854763, + 0.019704457372426987 + ], + "credit_map": [ + 0.006238980684429407, + 0.006106233224272728, + 0.007733477745205164, + 0.008588832803070545, + 0.005620971322059631, + 0.005720234476029873, + 0.007928759790956974, + 0.007808346301317215, + 0.009260297752916813, + 0.0068643661215901375, + 0.016008948907256126, + 0.014703151769936085, + 0.010820017196238041, + 0.02797757461667061, + 0.011990495957434177, + 0.012055136263370514, + 0.002587755210697651, + 0.0031058690510690212, + 0.0028332157526165247, + 0.003711891360580921, + 0.006441880017518997, + 0.00529860146343708, + 0.0037192616146057844, + 0.003400694578886032, + 0.006988400127738714, + 0.004308378789573908, + 0.013298554345965385, + 0.009313147515058517, + 0.007072392851114273, + 0.007689585909247398, + 0.009117729030549526, + 0.008660984225571156, + 0.0044015683233737946, + 0.0063468473963439465, + 0.007968897931277752, + 0.005906013771891594, + 0.004037943203002214, + 0.004029395990073681, + 0.005659875459969044, + 0.004986090119928122, + 0.06410171836614609, + 0.03214936703443527, + 0.04722175747156143, + 0.03685597702860832, + 0.03732748702168465, + 0.029759718105196953, + 0.11574698984622955, + 0.12170647084712982, + 0.006095779128372669, + 0.005368920974433422, + 0.009997210465371609, + 0.009381876327097416, + 0.007111311890184879, + 0.011811267584562302, + 0.011434783227741718, + 0.01154114305973053, + 0.006738901603966951, + 0.0045088669285178185, + 0.011236142367124557, + 0.011437266133725643, + 0.0059298379346728325, + 0.016661731526255608, + 0.009101870469748974, + 0.011405686847865582, + 0.003360797418281436, + 0.0033998703584074974, + 0.0073601496405899525, + 0.005327825900167227, + 0.0038703090976923704, + 0.006233866326510906, + 0.00458048889413476, + 0.0052193449810147285, + 0.004363310057669878, + 0.00618447782471776, + 0.0017117720562964678, + 0.0013848997186869383, + 0.0013382347533479333, + 0.0019935143645852804, + 0.0015618550824001431, + 0.0011663928162306547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.0015235256869345903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09455177187919617, + 0.04444868490099907, + 0.03480951488018036, + 0.04783939570188522, + 0.07074135541915894, + 0.3202473521232605, + 0.17945602536201477, + 0.08227518200874329, + 0.07763951271772385, + 0.047991231083869934 + ], + "credit_map": [ + 0.011188501492142677, + 0.008763832971453667, + 0.010161158628761768, + 0.013547524809837341, + 0.010206210426986217, + 0.01068961899727583, + 0.012076286599040031, + 0.017918642610311508, + 0.00419030012562871, + 0.003376094391569495, + 0.005775463301688433, + 0.004966712091118097, + 0.005953631363809109, + 0.0077282446436584, + 0.00510121276602149, + 0.0073570264503359795, + 0.0031924324575811625, + 0.00231308420188725, + 0.003458772785961628, + 0.003835817566141486, + 0.006412927992641926, + 0.005253463517874479, + 0.004360077902674675, + 0.005982936359941959, + 0.005640756338834763, + 0.0031990676652640104, + 0.009348439052700996, + 0.0068276203237473965, + 0.005915816407650709, + 0.004043295048177242, + 0.0056963455863296986, + 0.00716805225238204, + 0.009683582000434399, + 0.0075963810086250305, + 0.010884156450629234, + 0.008412329480051994, + 0.009141224436461926, + 0.008005192503333092, + 0.00979333184659481, + 0.007225155830383301, + 0.04383726045489311, + 0.021839866414666176, + 0.03139730542898178, + 0.0226855780929327, + 0.02623800002038479, + 0.02057942934334278, + 0.07442597299814224, + 0.07924392074346542, + 0.014800154604017735, + 0.015070243738591671, + 0.02801608294248581, + 0.024350421503186226, + 0.012829686515033245, + 0.02917644754052162, + 0.025010501965880394, + 0.03020247258245945, + 0.006175173446536064, + 0.005940575152635574, + 0.012591883540153503, + 0.01189484167844057, + 0.009176951833069324, + 0.018716953694820404, + 0.008849240839481354, + 0.008929559960961342, + 0.005686743184924126, + 0.005463097710162401, + 0.008717386052012444, + 0.007182408589869738, + 0.011782224290072918, + 0.02035280503332615, + 0.008195938542485237, + 0.010258905589580536, + 0.006286286283284426, + 0.012811318039894104, + 0.004678262863308191, + 0.004123759921640158, + 0.0035270436201244593, + 0.003967765253037214, + 0.00494124973192811, + 0.0076555428095161915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.002489208010956645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020568644627928734 + ], + "opposite_logprob_tokens": [ + -3.895568609237671 + ], + "logprob_reward": -0.020568644627928734, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020568644627928734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06291113048791885, + 0.035451315343379974, + 0.020123811438679695, + 0.03170182928442955, + 0.041996218264102936, + 0.1290023922920227, + 0.19997376203536987, + 0.2842780351638794, + 0.11763028055429459, + 0.07693125307559967 + ], + "credit_map": [ + 0.007262576837092638, + 0.006785753648728132, + 0.008791957050561905, + 0.0053367712534964085, + 0.006352270487695932, + 0.010394505225121975, + 0.00626877648755908, + 0.01171851810067892, + 0.0031580650247633457, + 0.0022635934874415398, + 0.005431776866316795, + 0.005225593689829111, + 0.0025392158422619104, + 0.00731851439923048, + 0.005980109330266714, + 0.0035344490315765142, + 0.001913868822157383, + 0.0019243410788476467, + 0.0024609884712845087, + 0.0032967356964945793, + 0.002711343113332987, + 0.0027970452792942524, + 0.0023338368628174067, + 0.0026856507174670696, + 0.00244522956199944, + 0.0021663708612322807, + 0.0052206371910870075, + 0.005932682193815708, + 0.003630884690210223, + 0.0025654055643826723, + 0.004500325303524733, + 0.005240289960056543, + 0.006046507507562637, + 0.004653122741729021, + 0.006761278957128525, + 0.004742288030683994, + 0.004309849347919226, + 0.005394366569817066, + 0.005088613368570805, + 0.005000193137675524, + 0.017004134133458138, + 0.008881048299372196, + 0.011042192578315735, + 0.011119826696813107, + 0.009230929426848888, + 0.00720216752961278, + 0.03180501237511635, + 0.03271709010004997, + 0.01483132503926754, + 0.01461417879909277, + 0.031234972178936005, + 0.028490101918578148, + 0.01393171213567257, + 0.031852640211582184, + 0.031130392104387283, + 0.03388843685388565, + 0.027106191962957382, + 0.018988927826285362, + 0.040407586842775345, + 0.03268305957317352, + 0.026983113959431648, + 0.05260675400495529, + 0.03218444809317589, + 0.05331794172525406, + 0.008410405367612839, + 0.009128937497735023, + 0.02155839093029499, + 0.017890064045786858, + 0.014003118500113487, + 0.017320532351732254, + 0.014518167823553085, + 0.014800670556724072, + 0.014766376465559006, + 0.006863032002002001, + 0.006532197818160057, + 0.002790948608890176, + 0.0032069324515759945, + 0.011071120388805866, + 0.009669732302427292, + 0.022030910477042198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.020568644627928734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera smoothly and deliberately sliding from left to right, revealing a stunning concept sports car on display at a prestigious auto show. Perched on a minimalist, polished black circular platform, the car boasts an incredibly sleek silhouette, painted a deep Candy Apple Red with a flawless, mirror-like metallic finish. Bright, focused spotlights from overhead rigs traverse its aerodynamic curves, creating sharp, moving highlights and emphasizing its radical design, including a visible carbon fiber diffuser and large, intricate alloy wheels. The surroundings of the vast showroom, deliberately darker and cooler in tone, the blurred figures of attendees and the ambient glow of the exhibits in the distance create a sense of scale and excitement. The entire scene is imbued with an aura of avant-garde luxury and anticipation, the camera's reverent movements and the dramatic play of light and shadow working together to present the car not just as a machine but as a work of art.", + "seed": 8190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8190_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": -0.02018997472478077, + "vlm_logprob": -0.02018997472478077, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.02018997472478077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044670283794403076 + ], + "opposite_logprob_tokens": [ + -4.544670104980469 + ], + "logprob_reward": -0.044670283794403076, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.044670283794403076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2852160632610321, + 0.10726699978113174, + 0.10799915343523026, + 0.054396726191043854, + 0.08171015977859497, + 0.07680165767669678, + 0.06376910209655762, + 0.032605595886707306, + 0.15321984887123108, + 0.037014663219451904 + ], + "credit_map": [ + 0.02810988761484623, + 0.02326267771422863, + 0.023945217952132225, + 0.043666042387485504, + 0.04687598720192909, + 0.030692147091031075, + 0.04648502916097641, + 0.04217909649014473, + 0.009156301617622375, + 0.03284299001097679, + 0.007531868759542704, + 0.01271392684429884, + 0.010061145760118961, + 0.007534088101238012, + 0.015191606245934963, + 0.012235073372721672, + 0.01252495776861906, + 0.02424931898713112, + 0.01586061529815197, + 0.01008435059338808, + 0.0073435064405202866, + 0.010056528262794018, + 0.010552477091550827, + 0.017327401787042618, + 0.005806408356875181, + 0.0036670442204922438, + 0.0111404312774539, + 0.0068976739421486855, + 0.005143295042216778, + 0.006707357242703438, + 0.0058966148644685745, + 0.009137902408838272, + 0.007412842009216547, + 0.007986788637936115, + 0.017986169084906578, + 0.01320839673280716, + 0.007588864304125309, + 0.008552421815693378, + 0.008928951807320118, + 0.010045729577541351, + 0.00931386835873127, + 0.007289474830031395, + 0.016755415126681328, + 0.011213712394237518, + 0.005552150774747133, + 0.005005563143640757, + 0.01255143154412508, + 0.009120042435824871, + 0.004438434727489948, + 0.010152523405849934, + 0.012176532298326492, + 0.010151523165404797, + 0.006169719155877829, + 0.0058553097769618034, + 0.006070269737392664, + 0.008754788897931576, + 0.002719544107094407, + 0.0025963482912629843, + 0.0061991713009774685, + 0.008344952017068863, + 0.0023654755204916, + 0.0023294391576200724, + 0.003992035053670406, + 0.004058633930981159, + 0.013405033387243748, + 0.009775573387742043, + 0.020083541050553322, + 0.046291254460811615, + 0.013153989799320698, + 0.00797327607870102, + 0.024995991960167885, + 0.01754118502140045, + 0.005288542713969946, + 0.0036521037109196186, + 0.005386771634221077, + 0.006827894598245621, + 0.004064078442752361, + 0.0034675253555178642, + 0.0036869607865810394, + 0.004640788305550814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.044670283794403076 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043084751814603806 + ], + "opposite_logprob_tokens": [ + -3.1680848598480225 + ], + "logprob_reward": -0.043084751814603806, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043084751814603806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07308443635702133, + 0.04689974710345268, + 0.10092268139123917, + 0.07253658026456833, + 0.0353979617357254, + 0.052978139370679855, + 0.23731480538845062, + 0.14123652875423431, + 0.14876289665699005, + 0.09086626023054123 + ], + "credit_map": [ + 0.008541526272892952, + 0.0064549935050308704, + 0.007584624458104372, + 0.01579643413424492, + 0.0063372026197612286, + 0.005833918694406748, + 0.011242878623306751, + 0.011292858049273491, + 0.004502167459577322, + 0.003248275490477681, + 0.00935772992670536, + 0.006509849335998297, + 0.00442820368334651, + 0.006343279033899307, + 0.00699646444991231, + 0.005513778887689114, + 0.006382763851433992, + 0.024267390370368958, + 0.012261980213224888, + 0.009604431688785553, + 0.008815649896860123, + 0.011565975844860077, + 0.013613578863441944, + 0.014410911127924919, + 0.006813742686063051, + 0.004934041295200586, + 0.014605510048568249, + 0.010545323602855206, + 0.0059191398322582245, + 0.007087355479598045, + 0.010851491242647171, + 0.01177997700870037, + 0.0029168035835027695, + 0.0029621762223541737, + 0.008872601203620434, + 0.005364615935832262, + 0.0036692991852760315, + 0.003590876469388604, + 0.004187216516584158, + 0.003834372851997614, + 0.0058334022760391235, + 0.005408380646258593, + 0.01173421274870634, + 0.007513977587223053, + 0.003635312896221876, + 0.003297693096101284, + 0.008434657007455826, + 0.007120506837964058, + 0.01769089326262474, + 0.042813144624233246, + 0.03695126995444298, + 0.03376154229044914, + 0.022763462737202644, + 0.020768655464053154, + 0.022726058959960938, + 0.03983978182077408, + 0.011273430660367012, + 0.013972800225019455, + 0.02089536376297474, + 0.01860763691365719, + 0.019996579736471176, + 0.014309617690742016, + 0.021892361342906952, + 0.020288744941353798, + 0.015163027681410313, + 0.017851216718554497, + 0.021774806082248688, + 0.02612244337797165, + 0.01783052273094654, + 0.015536153689026833, + 0.019178112968802452, + 0.015306604094803333, + 0.013955814763903618, + 0.01019199751317501, + 0.011083437129855156, + 0.013566012494266033, + 0.011845231987535954, + 0.009732550010085106, + 0.009268775582313538, + 0.011222440749406815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.043084751814603806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009310677996836603 + ], + "opposite_logprob_tokens": [ + -7.000931262969971 + ], + "logprob_reward": -0.0009310677996836603, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009310677996836603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2896521985530853, + 0.05756315588951111, + 0.08111002296209335, + 0.05649915710091591, + 0.09701735526323318, + 0.06472170352935791, + 0.08336244523525238, + 0.14471018314361572, + 0.1045708954334259, + 0.02079286426305771 + ], + "credit_map": [ + 0.024223001673817635, + 0.02368483692407608, + 0.025906171649694443, + 0.04283613711595535, + 0.04670393839478493, + 0.02945602685213089, + 0.04904903471469879, + 0.04779304191470146, + 0.005154496990144253, + 0.0049003916792571545, + 0.007042095065116882, + 0.0078092943876981735, + 0.007134477142244577, + 0.008081907406449318, + 0.009964686818420887, + 0.007475807331502438, + 0.011963200755417347, + 0.017225300893187523, + 0.00895724818110466, + 0.007025128696113825, + 0.007676198147237301, + 0.007881761528551579, + 0.008968205191195011, + 0.011412984691560268, + 0.005323212128132582, + 0.0037738222163170576, + 0.011811837553977966, + 0.007234453689306974, + 0.0034451393876224756, + 0.009181623347103596, + 0.006701794918626547, + 0.009027274325489998, + 0.012101230211555958, + 0.015712639316916466, + 0.016930507495999336, + 0.014994481578469276, + 0.007971652783453465, + 0.008149024099111557, + 0.00900298822671175, + 0.0121548380702734, + 0.007174205034971237, + 0.0062744030728936195, + 0.010797960683703423, + 0.009415804408490658, + 0.0061725955456495285, + 0.006031505763530731, + 0.009501738473773003, + 0.009353486821055412, + 0.006392165087163448, + 0.006986475549638271, + 0.013595069758594036, + 0.01341196522116661, + 0.007604214828461409, + 0.0074988193809986115, + 0.017907287925481796, + 0.009966435842216015, + 0.011228545568883419, + 0.013045033439993858, + 0.021149449050426483, + 0.023195073008537292, + 0.018162140622735023, + 0.01331062987446785, + 0.02122851461172104, + 0.023390797898173332, + 0.010286299511790276, + 0.009608639404177666, + 0.015860166400671005, + 0.02120555005967617, + 0.010935996659100056, + 0.010358510538935661, + 0.014015361666679382, + 0.012300373986363411, + 0.0026985087897628546, + 0.002304480643942952, + 0.002479524351656437, + 0.0026878979988396168, + 0.0030404527205973864, + 0.0020554345101118088, + 0.0023708001244813204, + 0.0031557658221572638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0009310677996836603 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024894457310438156 + ], + "opposite_logprob_tokens": [ + -6.0024895668029785 + ], + "logprob_reward": -0.0024894457310438156, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024894457310438156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15271982550621033, + 0.08559735864400864, + 0.058424897491931915, + 0.05601800978183746, + 0.05296556279063225, + 0.04997947812080383, + 0.14876101911067963, + 0.15450964868068695, + 0.19681087136268616, + 0.04421333968639374 + ], + "credit_map": [ + 0.014360387809574604, + 0.018193233758211136, + 0.016623428091406822, + 0.01621495932340622, + 0.025409923866391182, + 0.01798662170767784, + 0.023384375497698784, + 0.020546890795230865, + 0.008433207869529724, + 0.008306554518640041, + 0.01068166270852089, + 0.00914544053375721, + 0.00976607482880354, + 0.01333257183432579, + 0.012381122447550297, + 0.013550717383623123, + 0.004970101173967123, + 0.012388385832309723, + 0.009606845676898956, + 0.004931468050926924, + 0.0044768149964511395, + 0.006626171059906483, + 0.0064607965759932995, + 0.008964316919445992, + 0.005651678889989853, + 0.00304907513782382, + 0.008383450098335743, + 0.007745799142867327, + 0.003793861949816346, + 0.011015218682587147, + 0.007782006170600653, + 0.008596920408308506, + 0.0041483500972390175, + 0.0036395054776221514, + 0.015772229060530663, + 0.007415055297315121, + 0.0037065159995108843, + 0.0047425501979887486, + 0.006237534340471029, + 0.007303822319954634, + 0.005956105887889862, + 0.006093068514019251, + 0.00736188655719161, + 0.00753252487629652, + 0.003980494569987059, + 0.0042342497035861015, + 0.005266718566417694, + 0.009554433636367321, + 0.013110728934407234, + 0.014170401729643345, + 0.024110760539770126, + 0.023437215015292168, + 0.01325727067887783, + 0.011676644906401634, + 0.029347313567996025, + 0.01965068280696869, + 0.010418318212032318, + 0.014520230703055859, + 0.02357187308371067, + 0.02586839161813259, + 0.02164154313504696, + 0.011591295711696148, + 0.02004051022231579, + 0.026857487857341766, + 0.019150113686919212, + 0.017597759142518044, + 0.026181330904364586, + 0.05334289371967316, + 0.021535061299800873, + 0.012734535150229931, + 0.026721306145191193, + 0.01954786479473114, + 0.007883289828896523, + 0.005977017339318991, + 0.006048090755939484, + 0.004717160016298294, + 0.004401653539389372, + 0.0034870512317866087, + 0.0045720902271568775, + 0.0071269855834543705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0024894457310438156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774324484169483 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774324484169483, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774324484169483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19365718960762024, + 0.20334018766880035, + 0.08371682465076447, + 0.12712973356246948, + 0.08910681307315826, + 0.03649304434657097, + 0.08999407291412354, + 0.07641300559043884, + 0.07046777009963989, + 0.029681293293833733 + ], + "credit_map": [ + 0.018471788614988327, + 0.018440278246998787, + 0.01610015332698822, + 0.023856982588768005, + 0.03240180388092995, + 0.019058844074606895, + 0.03218429908156395, + 0.033143047243356705, + 0.019848152995109558, + 0.04755735397338867, + 0.02039993181824684, + 0.01823551207780838, + 0.020893044769763947, + 0.022159233689308167, + 0.027342423796653748, + 0.026904534548521042, + 0.007146219257265329, + 0.014774765819311142, + 0.013656155206263065, + 0.010806417092680931, + 0.006975430529564619, + 0.01033713947981596, + 0.00752375740557909, + 0.012496939860284328, + 0.015369758941233158, + 0.010589586570858955, + 0.02111152932047844, + 0.014525665901601315, + 0.008214322850108147, + 0.020224817097187042, + 0.017235372215509415, + 0.019858671352267265, + 0.006236181128770113, + 0.011673659086227417, + 0.01967824622988701, + 0.01520074624568224, + 0.00687369192019105, + 0.009880665689706802, + 0.008881314657628536, + 0.010682309046387672, + 0.003920138813555241, + 0.002641500672325492, + 0.006895055994391441, + 0.005208178889006376, + 0.003566671395674348, + 0.003264129161834717, + 0.0037564411759376526, + 0.007240929640829563, + 0.006455426104366779, + 0.01048208586871624, + 0.01419867668300867, + 0.014799878001213074, + 0.00815758015960455, + 0.006013144738972187, + 0.018407555297017097, + 0.011479726992547512, + 0.005026756785809994, + 0.007159522268921137, + 0.012629382312297821, + 0.01189811248332262, + 0.010353120975196362, + 0.006194594781845808, + 0.010340706445276737, + 0.012810811400413513, + 0.006058905739337206, + 0.007287986576557159, + 0.010476991534233093, + 0.01583046466112137, + 0.00760017242282629, + 0.006314561702311039, + 0.00867555383592844, + 0.00822312943637371, + 0.003520198864862323, + 0.00369004113599658, + 0.004043753258883953, + 0.0058919452130794525, + 0.0031942802015691996, + 0.0028908734675496817, + 0.003158765146508813, + 0.003291435306891799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.009774324484169483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera smoothly and deliberately sliding from left to right, revealing a stunning concept sports car on display at a prestigious auto show. Perched on a minimalist, polished black circular platform, the car boasts an incredibly sleek silhouette, painted a deep Candy Apple Red with a flawless, mirror-like metallic finish. Bright, focused spotlights from overhead rigs traverse its aerodynamic curves, creating sharp, moving highlights and emphasizing its radical design, including a visible carbon fiber diffuser and large, intricate alloy wheels. The surroundings of the vast showroom, deliberately darker and cooler in tone, the blurred figures of attendees and the ambient glow of the exhibits in the distance create a sense of scale and excitement. The entire scene is imbued with an aura of avant-garde luxury and anticipation, the camera's reverent movements and the dramatic play of light and shadow working together to present the car not just as a machine but as a work of art.", + "seed": 8191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8191_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": -0.032888023648411036, + "vlm_logprob": -0.032888023648411036, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.032888023648411036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07142961770296097 + ], + "opposite_logprob_tokens": [ + -3.44642972946167 + ], + "logprob_reward": -0.07142961770296097, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.07142961770296097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1222374215722084, + 0.046516429632902145, + 0.4881887137889862, + 0.03721923008561134, + 0.04496636614203453, + 0.035658907145261765, + 0.15803781151771545, + 0.021138830110430717, + 0.02333754114806652, + 0.022698720917105675 + ], + "credit_map": [ + 0.012696359306573868, + 0.017285967245697975, + 0.02583751268684864, + 0.00866871327161789, + 0.016000382602214813, + 0.0093058617785573, + 0.012503575533628464, + 0.019939051941037178, + 0.006224622949957848, + 0.0046096681617200375, + 0.0061251199804246426, + 0.007793552707880735, + 0.006100940052419901, + 0.0034831510856747627, + 0.006382945459336042, + 0.005796428769826889, + 0.0276139285415411, + 0.03368208929896355, + 0.07046981155872345, + 0.02421095035970211, + 0.09128763526678085, + 0.15458530187606812, + 0.03949476778507233, + 0.04684422165155411, + 0.008740783669054508, + 0.0035319256130605936, + 0.005434991326183081, + 0.00539619754999876, + 0.0036967103369534016, + 0.0025806929916143417, + 0.005376302171498537, + 0.002461628057062626, + 0.004685284104198217, + 0.004210660699754953, + 0.007363919634371996, + 0.005007678177207708, + 0.006791639141738415, + 0.005020332522690296, + 0.004134607966989279, + 0.007752240635454655, + 0.0041654896922409534, + 0.004288914147764444, + 0.00586386863142252, + 0.005880322773009539, + 0.003034640336409211, + 0.0022433525882661343, + 0.0031029663514345884, + 0.007079352159053087, + 0.011905197985470295, + 0.027558906003832817, + 0.02599572390317917, + 0.017872564494609833, + 0.03272870555520058, + 0.010174356400966644, + 0.012815245427191257, + 0.018987109884619713, + 0.0026856206823140383, + 0.002129373839125037, + 0.002185055986046791, + 0.0021419674158096313, + 0.002599460771307349, + 0.0026044496335089207, + 0.003200907027348876, + 0.0035919947549700737, + 0.0023523778654634953, + 0.002286482835188508, + 0.002348508220165968, + 0.002670580055564642, + 0.002311872085556388, + 0.005008261650800705, + 0.0028325410094112158, + 0.0035269169602543116, + 0.0025917673483490944, + 0.0016688620671629906, + 0.0035611006896942854, + 0.0022024947684258223, + 0.002101182471960783, + 0.004337403457611799, + 0.0031368860509246588, + 0.003099022665992379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.07142961770296097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790104791522026 + ], + "opposite_logprob_tokens": [ + -2.579010486602783 + ], + "logprob_reward": -0.0790104791522026, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790104791522026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0797947347164154, + 0.03320014849305153, + 0.3899588882923126, + 0.08613124489784241, + 0.07292662560939789, + 0.11916229128837585, + 0.06880287081003189, + 0.07304064184427261, + 0.04010866954922676, + 0.03687387332320213 + ], + "credit_map": [ + 0.00987400021404028, + 0.009777447208762169, + 0.020481735467910767, + 0.008322085253894329, + 0.006828879006206989, + 0.005009860266000032, + 0.008353855460882187, + 0.01114687044173479, + 0.004245180170983076, + 0.0038735107518732548, + 0.004358985926955938, + 0.00679004518315196, + 0.004319156985729933, + 0.003550697350874543, + 0.0028866007924079895, + 0.003175972728058696, + 0.03224039450287819, + 0.027952931821346283, + 0.05648865923285484, + 0.02098115161061287, + 0.06994234770536423, + 0.119461789727211, + 0.028484219685196877, + 0.034407395869493484, + 0.014723483473062515, + 0.0073974765837192535, + 0.011949889361858368, + 0.012259201146662235, + 0.012966478243470192, + 0.007189922966063023, + 0.01403773669153452, + 0.0056070550344884396, + 0.006870454642921686, + 0.005623226519674063, + 0.012479022145271301, + 0.011052612215280533, + 0.013338628225028515, + 0.004545884672552347, + 0.006500614807009697, + 0.012516179122030735, + 0.02192317321896553, + 0.008426457643508911, + 0.023781953379511833, + 0.022267330437898636, + 0.009591172449290752, + 0.008006935939192772, + 0.007923007011413574, + 0.01724226027727127, + 0.0070652225986123085, + 0.009239035658538342, + 0.012475861236453056, + 0.009456271305680275, + 0.011403534561395645, + 0.004620378836989403, + 0.006023846101015806, + 0.00851871445775032, + 0.005931000225245953, + 0.009530738927423954, + 0.006085259839892387, + 0.009197650477290154, + 0.011061797849833965, + 0.008549467660486698, + 0.01044183038175106, + 0.012242898344993591, + 0.002533744089305401, + 0.007701242808252573, + 0.00301350187510252, + 0.004939050413668156, + 0.005190120544284582, + 0.006201957818120718, + 0.004740165546536446, + 0.0057888878509402275, + 0.0032741758041083813, + 0.003751616459339857, + 0.002904807683080435, + 0.004734432324767113, + 0.004514607600867748, + 0.00809838529676199, + 0.004748370964080095, + 0.004847479052841663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.0790104791522026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007316772826015949 + ], + "opposite_logprob_tokens": [ + -7.250731468200684 + ], + "logprob_reward": -0.0007316772826015949, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007316772826015949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13791799545288086, + 0.0641595646739006, + 0.14183977246284485, + 0.03891688957810402, + 0.061059121042490005, + 0.0729614794254303, + 0.22649319469928741, + 0.1795641928911209, + 0.0395602285861969, + 0.037527527660131454 + ], + "credit_map": [ + 0.012631450779736042, + 0.018125014379620552, + 0.02330813929438591, + 0.011559594422578812, + 0.019452570006251335, + 0.010616738349199295, + 0.022183211520314217, + 0.02004127949476242, + 0.0077612814493477345, + 0.005454059224575758, + 0.00903042871505022, + 0.012029493227601051, + 0.006282947026193142, + 0.005258454009890556, + 0.01020076684653759, + 0.008142130449414253, + 0.01922827772796154, + 0.012758294120430946, + 0.021601125597953796, + 0.01381791289895773, + 0.016331419348716736, + 0.032903000712394714, + 0.010558273643255234, + 0.014641462825238705, + 0.005242621526122093, + 0.004091457463800907, + 0.005410743411630392, + 0.005902827251702547, + 0.004221444949507713, + 0.0033894008956849575, + 0.0067809103056788445, + 0.0038774849381297827, + 0.006200803909450769, + 0.004705462604761124, + 0.011252805590629578, + 0.006829440593719482, + 0.012395956553518772, + 0.004131817724555731, + 0.005793799180537462, + 0.0097490344196558, + 0.006066503934562206, + 0.007849667221307755, + 0.018413549289107323, + 0.009936539456248283, + 0.009533202275633812, + 0.0043971058912575245, + 0.0064207687973976135, + 0.010344144888222218, + 0.015318234451115131, + 0.027424143627285957, + 0.05545174330472946, + 0.04544614255428314, + 0.031154179945588112, + 0.0110948970541358, + 0.015767203643918037, + 0.024836644530296326, + 0.019631678238511086, + 0.021734243258833885, + 0.02589188702404499, + 0.018148884177207947, + 0.023008543998003006, + 0.015601061284542084, + 0.026246754452586174, + 0.029301131144165993, + 0.006243723910301924, + 0.003370215417817235, + 0.0038775275461375713, + 0.0041984086856245995, + 0.0035976727958768606, + 0.007878405041992664, + 0.003788215108215809, + 0.006606061477214098, + 0.00333877420052886, + 0.00437556067481637, + 0.004771423060446978, + 0.005341566167771816, + 0.004724704194813967, + 0.0049384902231395245, + 0.005196007434278727, + 0.004841000773012638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.0007316772826015949 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19242335855960846, + 0.08520587533712387, + 0.22801196575164795, + 0.05916998162865639, + 0.0467342734336853, + 0.052963536232709885, + 0.17377334833145142, + 0.04573875293135643, + 0.03699648752808571, + 0.07898249477148056 + ], + "credit_map": [ + 0.01862679235637188, + 0.024656185880303383, + 0.028502659872174263, + 0.016514847055077553, + 0.03212069720029831, + 0.016867276281118393, + 0.030275782570242882, + 0.024859098717570305, + 0.013972069136798382, + 0.00853954441845417, + 0.011374327354133129, + 0.01354788988828659, + 0.008331657387316227, + 0.009058643132448196, + 0.011175942607223988, + 0.009205793961882591, + 0.014306516386568546, + 0.016609106212854385, + 0.03195418417453766, + 0.014202897436916828, + 0.04103163257241249, + 0.06998206675052643, + 0.019029099494218826, + 0.0208964254707098, + 0.005602182820439339, + 0.006141172256320715, + 0.008402909152209759, + 0.008072205819189548, + 0.01277276873588562, + 0.005920024588704109, + 0.008652708493173122, + 0.003606004873290658, + 0.004269119817763567, + 0.004071468021720648, + 0.0068421754986047745, + 0.005398204084485769, + 0.006523295305669308, + 0.004419773817062378, + 0.004999111872166395, + 0.01021112222224474, + 0.008043897338211536, + 0.004761281423270702, + 0.007969536818563938, + 0.00920833833515644, + 0.004975297953933477, + 0.004544578958302736, + 0.003834360744804144, + 0.009626241400837898, + 0.013402457349002361, + 0.018812600523233414, + 0.04730396717786789, + 0.03453037515282631, + 0.019337067380547523, + 0.0088119488209486, + 0.012082156725227833, + 0.019492750987410545, + 0.008285408839583397, + 0.004675391595810652, + 0.004335310310125351, + 0.0059142508544027805, + 0.005119136068969965, + 0.004632221069186926, + 0.006155910436064005, + 0.006621120031923056, + 0.003476447658613324, + 0.005411095451563597, + 0.004527197685092688, + 0.0041710929945111275, + 0.0033814210910350084, + 0.006308624520897865, + 0.004811910446733236, + 0.004908693954348564, + 0.006010599434375763, + 0.009831706993281841, + 0.009564838372170925, + 0.008954275399446487, + 0.010313857346773148, + 0.013216020539402962, + 0.010485059581696987, + 0.010606137104332447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068493127822876 + ], + "opposite_logprob_tokens": [ + -4.511068344116211 + ], + "logprob_reward": -0.011068493127822876, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068493127822876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09328728914260864, + 0.10019839555025101, + 0.49054643511772156, + 0.05383969843387604, + 0.028250694274902344, + 0.02875557355582714, + 0.02773290127515793, + 0.03528342396020889, + 0.09397988766431808, + 0.048125721514225006 + ], + "credit_map": [ + 0.010496330447494984, + 0.008494088426232338, + 0.016248824074864388, + 0.007099113892763853, + 0.012006185948848724, + 0.007932848297059536, + 0.015189092606306076, + 0.015820810571312904, + 0.014743401668965816, + 0.008064836263656616, + 0.01299581490457058, + 0.014379192143678665, + 0.014941160567104816, + 0.014032082632184029, + 0.008536896668374538, + 0.012504995800554752, + 0.02998301200568676, + 0.03251192718744278, + 0.07035662978887558, + 0.023933911696076393, + 0.09119503200054169, + 0.15672555565834045, + 0.03947623446583748, + 0.04636415094137192, + 0.0059249140322208405, + 0.005846090614795685, + 0.008388594724237919, + 0.007967784069478512, + 0.010202988982200623, + 0.004785040393471718, + 0.007221735082566738, + 0.003502551233395934, + 0.0031280084513127804, + 0.0027236687019467354, + 0.004162992350757122, + 0.0036219211760908365, + 0.0037653245963156223, + 0.002609549555927515, + 0.0033871897030621767, + 0.004852037876844406, + 0.0020945577416568995, + 0.0032132018823176622, + 0.004915547091513872, + 0.005715118255466223, + 0.0024702076334506273, + 0.0022224443964660168, + 0.002421922981739044, + 0.005702572874724865, + 0.0027404206339269876, + 0.0031728644389659166, + 0.004468531813472509, + 0.004713735543191433, + 0.0030985285993665457, + 0.0026042258832603693, + 0.002927495399489999, + 0.00400709779933095, + 0.004053700715303421, + 0.0034188851714134216, + 0.004006284289062023, + 0.0045006489381194115, + 0.004204896744340658, + 0.004496817942708731, + 0.00517684081569314, + 0.005425348412245512, + 0.012139173224568367, + 0.009340137243270874, + 0.011941506527364254, + 0.010313735343515873, + 0.008090715855360031, + 0.016622187569737434, + 0.012764803133904934, + 0.012767629697918892, + 0.003543123137205839, + 0.006169498432427645, + 0.005871315021067858, + 0.0059403483755886555, + 0.006153675727546215, + 0.007800633553415537, + 0.00650758296251297, + 0.006139547098428011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.011068493127822876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle shot, the camera smoothly and deliberately sliding from left to right, revealing a stunning concept sports car on display at a prestigious auto show. Perched on a minimalist, polished black circular platform, the car boasts an incredibly sleek silhouette, painted a deep Candy Apple Red with a flawless, mirror-like metallic finish. Bright, focused spotlights from overhead rigs traverse its aerodynamic curves, creating sharp, moving highlights and emphasizing its radical design, including a visible carbon fiber diffuser and large, intricate alloy wheels. The surroundings of the vast showroom, deliberately darker and cooler in tone, the blurred figures of attendees and the ambient glow of the exhibits in the distance create a sense of scale and excitement. The entire scene is imbued with an aura of avant-garde luxury and anticipation, the camera's reverent movements and the dramatic play of light and shadow working together to present the car not just as a machine but as a work of art.", + "seed": 8192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_13/index_5135_seed_8192_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": -0.02359181799693033, + "vlm_logprob": -0.02359181799693033, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.02359181799693033, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0394691526889801 + ], + "opposite_logprob_tokens": [ + -4.664469242095947 + ], + "logprob_reward": -0.0394691526889801, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0394691526889801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04728703945875168, + 0.047991760075092316, + 0.2365502268075943, + 0.02179788611829281, + 0.034255463629961014, + 0.036661844700574875, + 0.12391506135463715, + 0.09673675149679184, + 0.23101241886615753, + 0.12379160523414612 + ], + "credit_map": [ + 0.003537635086104274, + 0.004686554428189993, + 0.00467822328209877, + 0.007856451906263828, + 0.005264653358608484, + 0.008654892444610596, + 0.005258759018033743, + 0.007349870167672634, + 0.004394781310111284, + 0.002633624244481325, + 0.0070892926305532455, + 0.010977211408317089, + 0.0028705394361168146, + 0.0049729072488844395, + 0.0074387420900166035, + 0.007614658679813147, + 0.006976070813834667, + 0.015584629029035568, + 0.05721025541424751, + 0.03075784258544445, + 0.007186326198279858, + 0.013806919567286968, + 0.08639170229434967, + 0.01863647997379303, + 0.00243574776686728, + 0.0019858803134411573, + 0.0037334763910621405, + 0.004384299740195274, + 0.0017605893081054091, + 0.0023352124262601137, + 0.0025625070556998253, + 0.0026001723017543554, + 0.0021314830519258976, + 0.002539348090067506, + 0.0033612630795687437, + 0.005933389067649841, + 0.0022923743817955256, + 0.008052993565797806, + 0.005217794794589281, + 0.004726815037429333, + 0.0026723307091742754, + 0.002434512134641409, + 0.00590807618573308, + 0.006261209025979042, + 0.0020052623003721237, + 0.006858060136437416, + 0.005520756356418133, + 0.005001631565392017, + 0.015289674513041973, + 0.009398378431797028, + 0.018580537289381027, + 0.01874222420156002, + 0.007498258724808693, + 0.012921319343149662, + 0.02242419682443142, + 0.019060466438531876, + 0.011347074061632156, + 0.020599637180566788, + 0.01284126378595829, + 0.009862647391855717, + 0.012119964696466923, + 0.009278271347284317, + 0.012916162610054016, + 0.007771726232022047, + 0.007098720408976078, + 0.010702090337872505, + 0.07811737805604935, + 0.08049467951059341, + 0.00928129255771637, + 0.007978838868439198, + 0.01843874529004097, + 0.0189006719738245, + 0.00854303315281868, + 0.015194782055914402, + 0.022087618708610535, + 0.015493891201913357, + 0.012404650449752808, + 0.011461201123893261, + 0.019766652956604958, + 0.018839770928025246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.0394691526889801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002904266119003 + ], + "opposite_logprob_tokens": [ + -2.6950290203094482 + ], + "logprob_reward": -0.07002904266119003, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002904266119003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08188004046678543, + 0.037527039647102356, + 0.30043014883995056, + 0.029799070209264755, + 0.04723421111702919, + 0.019920384511351585, + 0.11714199185371399, + 0.13960418105125427, + 0.05703277140855789, + 0.16943015158176422 + ], + "credit_map": [ + 0.006114920601248741, + 0.009870476089417934, + 0.014224771410226822, + 0.012052301317453384, + 0.005504642613232136, + 0.008284580893814564, + 0.012965376488864422, + 0.012862974777817726, + 0.0025371406227350235, + 0.00283769890666008, + 0.006291661411523819, + 0.007528468035161495, + 0.002230273559689522, + 0.004197100643068552, + 0.006073593627661467, + 0.005831102374941111, + 0.0090022012591362, + 0.020730558782815933, + 0.07433496415615082, + 0.04163606837391853, + 0.008442738093435764, + 0.0179749745875597, + 0.10766635835170746, + 0.020642312243580818, + 0.004088366404175758, + 0.002637424971908331, + 0.002511379774659872, + 0.0054269470274448395, + 0.002310815965756774, + 0.0029351194389164448, + 0.00508661987259984, + 0.004802399780601263, + 0.0030531734228134155, + 0.0033129665534943342, + 0.0044565945863723755, + 0.009583249688148499, + 0.0033893336076289415, + 0.009249767288565636, + 0.007393538951873779, + 0.006795587483793497, + 0.0018869240302592516, + 0.0017866163980215788, + 0.0038446420803666115, + 0.0028903456404805183, + 0.0013010271359235048, + 0.0024982253089547157, + 0.0029491791501641273, + 0.002763427561149001, + 0.011268886737525463, + 0.006376245990395546, + 0.017447184771299362, + 0.024628646671772003, + 0.007244095206260681, + 0.00966909434646368, + 0.023775793612003326, + 0.016732051968574524, + 0.009541555307805538, + 0.020912576466798782, + 0.026320025324821472, + 0.030698580667376518, + 0.012799940072000027, + 0.007390002720057964, + 0.018059730529785156, + 0.013881771825253963, + 0.00200831051915884, + 0.0029094989877194166, + 0.013858303427696228, + 0.014719590544700623, + 0.0021484363824129105, + 0.005251521710306406, + 0.004537246190011501, + 0.01159986574202776, + 0.00835439469665289, + 0.020241139456629753, + 0.04065367951989174, + 0.02140701748430729, + 0.024692019447684288, + 0.00947529450058937, + 0.03240203112363815, + 0.012204579077661037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 7 + ], + "selected_reward": -0.07002904266119003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000726316764485091 + ], + "opposite_logprob_tokens": [ + -7.250726222991943 + ], + "logprob_reward": -0.000726316764485091, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.000726316764485091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0732579454779625, + 0.08160386234521866, + 0.03298661485314369, + 0.0474403016269207, + 0.03832034021615982, + 0.057818494737148285, + 0.07255809754133224, + 0.40793418884277344, + 0.10234791785478592, + 0.08573225140571594 + ], + "credit_map": [ + 0.006398022640496492, + 0.007055220659822226, + 0.007621955592185259, + 0.01549017708748579, + 0.007539249956607819, + 0.0095441285520792, + 0.009538853541016579, + 0.010070333257317543, + 0.007684993091970682, + 0.004757350776344538, + 0.014596118591725826, + 0.016811106353998184, + 0.005163538735359907, + 0.0077805668115615845, + 0.01287551037967205, + 0.01193467527627945, + 0.002461414085701108, + 0.0021497937850654125, + 0.003998839762061834, + 0.005487323738634586, + 0.0020134486258029938, + 0.002691882196813822, + 0.00760648725554347, + 0.00657742191106081, + 0.004656185396015644, + 0.0036983871832489967, + 0.007327114697545767, + 0.007469819393008947, + 0.004503351170569658, + 0.0076116896234452724, + 0.004857432097196579, + 0.007316317409276962, + 0.00256316177546978, + 0.0030555494595319033, + 0.0036927429027855396, + 0.007658689748495817, + 0.0023335060104727745, + 0.007829605601727962, + 0.007198581472039223, + 0.003988500684499741, + 0.004879482556134462, + 0.0031857232097536325, + 0.011440268717706203, + 0.010382276959717274, + 0.0027075852267444134, + 0.008556941524147987, + 0.008922317996621132, + 0.0077438959851861, + 0.006761195138096809, + 0.00803322158753872, + 0.011550256051123142, + 0.009058577008545399, + 0.006288296543061733, + 0.006418791599571705, + 0.012996516190469265, + 0.011451242491602898, + 0.039358269423246384, + 0.08525121212005615, + 0.06352046132087708, + 0.05916212871670723, + 0.0475999191403389, + 0.03264941647648811, + 0.04720018431544304, + 0.03319256752729416, + 0.0034940445329993963, + 0.0073284562677145, + 0.03027273528277874, + 0.03067934699356556, + 0.005643593147397041, + 0.004461836535483599, + 0.010108476504683495, + 0.010359417647123337, + 0.003546766471117735, + 0.009049947373569012, + 0.0195537768304348, + 0.009704601019620895, + 0.009862764738500118, + 0.008576463907957077, + 0.017233557999134064, + 0.008204363286495209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.000726316764485091 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12789668142795563, + 0.06606366485357285, + 0.15325483679771423, + 0.032626498490571976, + 0.036725107580423355, + 0.0427585169672966, + 0.08194868266582489, + 0.32508915662765503, + 0.09093759953975677, + 0.04269929975271225 + ], + "credit_map": [ + 0.008669854141771793, + 0.010042745620012283, + 0.023867696523666382, + 0.019968777894973755, + 0.010974434204399586, + 0.01799987629055977, + 0.01764349266886711, + 0.01872979663312435, + 0.006510792765766382, + 0.0047472319565713406, + 0.011187375523149967, + 0.009443620219826698, + 0.0053746807388961315, + 0.009537244215607643, + 0.009121199138462543, + 0.010141510516405106, + 0.00487098190933466, + 0.010296721942722797, + 0.03780567646026611, + 0.020367993041872978, + 0.004995228257030249, + 0.009493779391050339, + 0.05510786920785904, + 0.010316555388271809, + 0.0026964228600263596, + 0.003103466471657157, + 0.0033188406378030777, + 0.006775236688554287, + 0.0030043974984437227, + 0.004888111725449562, + 0.003529512556269765, + 0.005310509819537401, + 0.0018660180503502488, + 0.002457269700244069, + 0.0033020111732184887, + 0.009842011146247387, + 0.0024646176025271416, + 0.004612309392541647, + 0.008785367012023926, + 0.0033955019898712635, + 0.0033983425237238407, + 0.003913276828825474, + 0.006629360374063253, + 0.0071990531869232655, + 0.003210890106856823, + 0.007356066256761551, + 0.006043280009180307, + 0.005008243955671787, + 0.011156187392771244, + 0.005985643714666367, + 0.011257145553827286, + 0.013184295035898685, + 0.004020951688289642, + 0.009519275277853012, + 0.013494465500116348, + 0.013330708257853985, + 0.03258771076798439, + 0.07069239020347595, + 0.04819026589393616, + 0.043233513832092285, + 0.04010093957185745, + 0.02660086750984192, + 0.03834430128335953, + 0.025339137762784958, + 0.0028215183410793543, + 0.0043089138343930244, + 0.031144915148615837, + 0.032882846891880035, + 0.0031839143484830856, + 0.004443674348294735, + 0.005852179601788521, + 0.006299624685198069, + 0.0028751788195222616, + 0.004145440645515919, + 0.00725462194532156, + 0.006099688354879618, + 0.0034156464971601963, + 0.006036507897078991, + 0.006531049031764269, + 0.006341161206364632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slide smoothly and deliberately from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a concept sports car painted deep Candy Apple Red visible on a minimalist, polished black circular platform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bright, focused overhead spotlights creating sharp, moving highlights on the car's curves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast, darker showroom with blurred figures of attendees and ambient glow from exhibits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247272551059723 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247272551059723, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247272551059723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.138124018907547, + 0.18641747534275055, + 0.09760107845067978, + 0.06835044920444489, + 0.045865945518016815, + 0.024297352880239487, + 0.05289216339588165, + 0.2343538999557495, + 0.06563200056552887, + 0.08646558970212936 + ], + "credit_map": [ + 0.009035362862050533, + 0.014383502304553986, + 0.02045881189405918, + 0.025598814710974693, + 0.012280366383492947, + 0.020719967782497406, + 0.014313788153231144, + 0.021333394572138786, + 0.013105252757668495, + 0.015944605693221092, + 0.024024933576583862, + 0.031495578587055206, + 0.019099131226539612, + 0.02381521835923195, + 0.031104963272809982, + 0.02782779186964035, + 0.0032123939599841833, + 0.006621176842600107, + 0.023762548342347145, + 0.01294876728206873, + 0.00298222410492599, + 0.00583741394802928, + 0.0347377173602581, + 0.007498841267079115, + 0.004189162980765104, + 0.0035481576342135668, + 0.0060503133572638035, + 0.02373269945383072, + 0.004313552752137184, + 0.006326309870928526, + 0.011600421741604805, + 0.008589823730289936, + 0.0020834957249462605, + 0.0025521598290652037, + 0.0038898868951946497, + 0.009720905683934689, + 0.0021653403528034687, + 0.008503262884914875, + 0.009182997047901154, + 0.007767894305288792, + 0.001734185148961842, + 0.0013591357273980975, + 0.0037782059516757727, + 0.0034824637696146965, + 0.0014750007539987564, + 0.003937867935746908, + 0.00439665000885725, + 0.004133842885494232, + 0.006847279146313667, + 0.00610254006460309, + 0.007846145890653133, + 0.006201489362865686, + 0.00437628198415041, + 0.005619908683001995, + 0.009756139479577541, + 0.006142381578683853, + 0.021814601495862007, + 0.04870384559035301, + 0.0367913544178009, + 0.03516385331749916, + 0.02770686149597168, + 0.017631668597459793, + 0.02706564962863922, + 0.01947607472538948, + 0.003395281033590436, + 0.005863720551133156, + 0.015266933478415012, + 0.017572874203324318, + 0.0058692703023552895, + 0.00495772622525692, + 0.007453133352100849, + 0.005253061652183533, + 0.006672472693026066, + 0.018030041828751564, + 0.011144071817398071, + 0.010375823825597763, + 0.01719149574637413, + 0.00761212594807148, + 0.009592879563570023, + 0.005846673157066107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.005247272551059723 + } + ], + "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/0000158/group_14/index_tensor([6564, 6564])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_tensor([6564, 6564])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6b4ee286d80bbb542d60dca3d01316a62bfd6107 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_tensor([6564, 6564])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic low-angle wide shot captures a massive bronze bull statue, its muscular body gleaming in the blazing midday sun. Set across a vast city square, the camera suddenly shakes violently, mimicking the force of an earth-shaking tremor that sends ripples through the stone pavement. Following the earthquake, the massive statue begins to teeter on its massive granite base, its surface a mixture of polished bronze and a weathered green patina. It tilts forward, its immense weight undergoing a sickening, harrowing shift, its powerful form tilting perilously before swaying backward with terrifying momentum. Dust and small stone fragments erupt from the base, where the statue's hooves scrape and grind against it, leaving tiny crack lines and cobwebs on the surface, creating a sudden, heart-stopping tension and the threat of imminent, catastrophic collapse.", + "seed": 10023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_6564_seed_10023_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": -0.1775285540148616, + "vlm_logprob": -0.1775285540148616, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.1775285540148616, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01112178061157465 + ], + "opposite_logprob_tokens": [ + -4.63612174987793 + ], + "logprob_reward": -0.01112178061157465, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01112178061157465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15912100672721863, + 0.1007087230682373, + 0.16071198880672455, + 0.0622706301510334, + 0.18787319958209991, + 0.04668036475777626, + 0.10165151953697205, + 0.025662966072559357, + 0.12267674505710602, + 0.032642897218465805 + ], + "credit_map": [ + 0.036394890397787094, + 0.01815536804497242, + 0.027549780905246735, + 0.014789876528084278, + 0.011960598640143871, + 0.013845358975231647, + 0.014775311574339867, + 0.0216497965157032, + 0.009087683632969856, + 0.02006811648607254, + 0.010547060519456863, + 0.017833547666668892, + 0.011606866493821144, + 0.009074621833860874, + 0.011138450354337692, + 0.0113523593172431, + 0.00851517915725708, + 0.012499657459557056, + 0.022813253104686737, + 0.018147168681025505, + 0.018137797713279724, + 0.01661018095910549, + 0.023745380342006683, + 0.04024336114525795, + 0.004288131836801767, + 0.010009000077843666, + 0.00957953929901123, + 0.014065640978515148, + 0.006208969745784998, + 0.0037239312659949064, + 0.004456975497305393, + 0.00993843749165535, + 0.022070273756980896, + 0.023353872820734978, + 0.015851790085434914, + 0.030659953132271767, + 0.019027188420295715, + 0.03446430340409279, + 0.02320135198533535, + 0.019244443625211716, + 0.004140205215662718, + 0.0052339741960167885, + 0.003965712618082762, + 0.00544044328853488, + 0.0032241272274404764, + 0.013741964474320412, + 0.003296776907518506, + 0.007637156173586845, + 0.010093972086906433, + 0.011720797047019005, + 0.008503380231559277, + 0.009795820340514183, + 0.005810593720525503, + 0.01607694663107395, + 0.010979392565786839, + 0.028670599684119225, + 0.0021966618951410055, + 0.0023301446344703436, + 0.003998142667114735, + 0.0037272474728524685, + 0.0025381194427609444, + 0.002942565595731139, + 0.004163066390901804, + 0.003767014481127262, + 0.0064183250069618225, + 0.0060567790642380714, + 0.03798971325159073, + 0.012869936414062977, + 0.005799274425953627, + 0.008063950575888157, + 0.007494096178561449, + 0.03798466548323631, + 0.0020985319279134274, + 0.003058692440390587, + 0.0032181316055357456, + 0.005446383263915777, + 0.0035277358256280422, + 0.004195766523480415, + 0.006139060948044062, + 0.004958590492606163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.01112178061157465 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079863533377647 + ], + "opposite_logprob_tokens": [ + -4.141079902648926 + ], + "logprob_reward": -0.016079863533377647, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079863533377647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10916571319103241, + 0.09719233959913254, + 0.10458345711231232, + 0.09192102402448654, + 0.18670213222503662, + 0.030521459877490997, + 0.08455434441566467, + 0.06222476437687874, + 0.14326386153697968, + 0.08987095206975937 + ], + "credit_map": [ + 0.018763450905680656, + 0.015129189938306808, + 0.015808286145329475, + 0.013607032597064972, + 0.00938207283616066, + 0.01078072004020214, + 0.012557492591440678, + 0.013137470930814743, + 0.009129165671765804, + 0.010433681309223175, + 0.008412940427660942, + 0.010628909803926945, + 0.014150680974125862, + 0.023214740678668022, + 0.007829425856471062, + 0.013392791152000427, + 0.006873107049614191, + 0.00819874182343483, + 0.01592765375971794, + 0.01376278605312109, + 0.011997343972325325, + 0.00968344509601593, + 0.014778641052544117, + 0.02336173504590988, + 0.005918869283050299, + 0.00871189497411251, + 0.011169436387717724, + 0.026061812415719032, + 0.005878433585166931, + 0.005801431369036436, + 0.007226162124425173, + 0.021152988076210022, + 0.0290501918643713, + 0.026684490963816643, + 0.01332946214824915, + 0.020777305588126183, + 0.016256144270300865, + 0.03582950308918953, + 0.023855013772845268, + 0.02092001773416996, + 0.0021414475049823523, + 0.0037090745754539967, + 0.0024208759423345327, + 0.00386358005926013, + 0.001731598167680204, + 0.008338802494108677, + 0.003874911228194833, + 0.0044411709532141685, + 0.007667167577892542, + 0.009363342076539993, + 0.006671506445854902, + 0.0070980922318995, + 0.005219666287302971, + 0.01424610335379839, + 0.008126521483063698, + 0.02616194263100624, + 0.0053834468126297, + 0.005026421044021845, + 0.00611831434071064, + 0.008902911096811295, + 0.00529544148594141, + 0.00925110187381506, + 0.014683877117931843, + 0.007563246879726648, + 0.006639217026531696, + 0.004904258530586958, + 0.03628105670213699, + 0.014846224337816238, + 0.015582295134663582, + 0.016069995239377022, + 0.011325845494866371, + 0.037614963948726654, + 0.012169098481535912, + 0.007473732344806194, + 0.01831677556037903, + 0.010678974911570549, + 0.009870165027678013, + 0.010749961249530315, + 0.008654778823256493, + 0.011957471258938313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.016079863533377647 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.8260756731033325 + ], + "opposite_logprob_tokens": [ + -0.5760756731033325 + ], + "logprob_reward": -0.8260756731033325, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8260756731033325, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1391695886850357, + 0.05058075115084648, + 0.06626410782337189, + 0.0984753668308258, + 0.27941763401031494, + 0.07864222675561905, + 0.1123800277709961, + 0.06377946585416794, + 0.08204992115497589, + 0.02924090065062046 + ], + "credit_map": [ + 0.025636734440922737, + 0.0208670012652874, + 0.020877689123153687, + 0.015843462198972702, + 0.014658971689641476, + 0.012544090859591961, + 0.012429196387529373, + 0.016312450170516968, + 0.005823733285069466, + 0.008524136617779732, + 0.006373098585754633, + 0.006316198501735926, + 0.0056449840776622295, + 0.008064019493758678, + 0.005539047531783581, + 0.004295534919947386, + 0.002626511501148343, + 0.005692620296031237, + 0.00531732477247715, + 0.008752424269914627, + 0.006893127225339413, + 0.005537726450711489, + 0.01198562327772379, + 0.019458750262856483, + 0.01009038370102644, + 0.009288049302995205, + 0.011172188445925713, + 0.020558344200253487, + 0.01182345487177372, + 0.009904759004712105, + 0.01090843416750431, + 0.014729754999279976, + 0.026492414996027946, + 0.03226032108068466, + 0.024678915739059448, + 0.05031118914484978, + 0.022188792005181313, + 0.05609779432415962, + 0.02998771145939827, + 0.0374004952609539, + 0.009357165545225143, + 0.008019156754016876, + 0.008458202704787254, + 0.011388501152396202, + 0.009565328247845173, + 0.00855258572846651, + 0.012994552031159401, + 0.010306737385690212, + 0.008655985817313194, + 0.010938756167888641, + 0.0107906898483634, + 0.010995935648679733, + 0.005964178591966629, + 0.02104533277451992, + 0.012902546674013138, + 0.031086599454283714, + 0.004835421219468117, + 0.004365436267107725, + 0.005750211887061596, + 0.008224242366850376, + 0.0058247740380465984, + 0.01134980097413063, + 0.013462909497320652, + 0.009966665878891945, + 0.004323472268879414, + 0.00399078568443656, + 0.025648361071944237, + 0.008880685083568096, + 0.003882330609485507, + 0.005801705177873373, + 0.005920312833040953, + 0.02360226772725582, + 0.0018928664503619075, + 0.0019179392838850617, + 0.0032865279354155064, + 0.005136209074407816, + 0.0029693280812352896, + 0.004492800682783127, + 0.004284837283194065, + 0.005260389298200607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.8260756731033325 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1732141524553299, + 0.057631589472293854, + 0.040023960173130035, + 0.0598805770277977, + 0.26754629611968994, + 0.0713278129696846, + 0.08020073920488358, + 0.043399859219789505, + 0.1727086901664734, + 0.03406631574034691 + ], + "credit_map": [ + 0.0417264923453331, + 0.019542457535862923, + 0.0303784366697073, + 0.013054023496806622, + 0.012486125342547894, + 0.015294262208044529, + 0.01589127629995346, + 0.024841077625751495, + 0.005633033346384764, + 0.009256120771169662, + 0.006222122348845005, + 0.009907953441143036, + 0.006654506549239159, + 0.007742208894342184, + 0.007429502438753843, + 0.0047861430794000626, + 0.00300794979557395, + 0.0034231001045554876, + 0.007725555449724197, + 0.00422027800232172, + 0.005514843389391899, + 0.004277863539755344, + 0.0063226912170648575, + 0.005531681701540947, + 0.004012357909232378, + 0.010625875554978848, + 0.0057796575129032135, + 0.012171560898423195, + 0.008568212389945984, + 0.004466368351131678, + 0.00560003099963069, + 0.008656520396471024, + 0.031001297757029533, + 0.03187102451920509, + 0.02262290194630623, + 0.05290854722261429, + 0.022248361259698868, + 0.031670793890953064, + 0.03756174445152283, + 0.03766164183616638, + 0.006366415414959192, + 0.007943768985569477, + 0.0073998658917844296, + 0.010450221598148346, + 0.005111093632876873, + 0.01733902283012867, + 0.008304859511554241, + 0.008412563242018223, + 0.006717191077768803, + 0.007382944691926241, + 0.006614350248128176, + 0.007770737633109093, + 0.006970624905079603, + 0.013998874463140965, + 0.012576625682413578, + 0.018169397488236427, + 0.004824028350412846, + 0.004226911813020706, + 0.006288470700383186, + 0.006617323495447636, + 0.004841376096010208, + 0.00546000711619854, + 0.00553274666890502, + 0.005608994513750076, + 0.007654184475541115, + 0.007024122402071953, + 0.05557871609926224, + 0.01755600981414318, + 0.006004725582897663, + 0.012158862315118313, + 0.010814731009304523, + 0.055917348712682724, + 0.0031872775871306658, + 0.0028913593851029873, + 0.003996836021542549, + 0.00514519028365612, + 0.004825875163078308, + 0.004577446263283491, + 0.005440301727503538, + 0.0040020328015089035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11376304924488068, + 0.072461798787117, + 0.1334121823310852, + 0.1206781268119812, + 0.22946327924728394, + 0.05603145807981491, + 0.09241745620965958, + 0.05145722255110741, + 0.07729414105415344, + 0.053021278232336044 + ], + "credit_map": [ + 0.028302811086177826, + 0.011916574090719223, + 0.019545646384358406, + 0.009278160519897938, + 0.00918117817491293, + 0.008884926326572895, + 0.011491863988339901, + 0.015161894261837006, + 0.007437847554683685, + 0.011715407483279705, + 0.0063330321572721004, + 0.011352348141372204, + 0.009813293814659119, + 0.009468470700085163, + 0.007762867026031017, + 0.008578528650105, + 0.008598582819104195, + 0.01534181647002697, + 0.014892643317580223, + 0.029566565528512, + 0.014327673241496086, + 0.010375192388892174, + 0.02072370983660221, + 0.019585995003581047, + 0.008976053446531296, + 0.018120573833584785, + 0.01904810406267643, + 0.019549036398530006, + 0.012946710921823978, + 0.010135370306670666, + 0.01191296149045229, + 0.019989313557744026, + 0.028809892013669014, + 0.025559213012456894, + 0.020033469423651695, + 0.02664823643863201, + 0.024263478815555573, + 0.05431267246603966, + 0.02608622796833515, + 0.02375006675720215, + 0.0056620389223098755, + 0.0065435259602963924, + 0.005572105757892132, + 0.009539572522044182, + 0.004933131858706474, + 0.00795311201363802, + 0.007891478016972542, + 0.007936493493616581, + 0.007425480522215366, + 0.008332193829119205, + 0.008340620435774326, + 0.008534736931324005, + 0.005504578817635775, + 0.015179582871496677, + 0.011733454652130604, + 0.02736680768430233, + 0.0036891379859298468, + 0.0032502971589565277, + 0.006131687201559544, + 0.00667907390743494, + 0.0032425695098936558, + 0.014758630655705929, + 0.006976584438234568, + 0.006729239597916603, + 0.005224345251917839, + 0.004140126518905163, + 0.017438463866710663, + 0.010329748503863811, + 0.009744699113070965, + 0.007443207316100597, + 0.007537278812378645, + 0.015436273068189621, + 0.0033612672705203295, + 0.0043965899385511875, + 0.005142820533365011, + 0.007368443068116903, + 0.006140522658824921, + 0.009046887047588825, + 0.009746248833835125, + 0.007818498648703098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.023284463211894035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot captures a massive bronze bull statue, its muscular body gleaming in the blazing midday sun. Set across a vast city square, the camera suddenly shakes violently, mimicking the force of an earth-shaking tremor that sends ripples through the stone pavement. Following the earthquake, the massive statue begins to teeter on its massive granite base, its surface a mixture of polished bronze and a weathered green patina. It tilts forward, its immense weight undergoing a sickening, harrowing shift, its powerful form tilting perilously before swaying backward with terrifying momentum. Dust and small stone fragments erupt from the base, where the statue's hooves scrape and grind against it, leaving tiny crack lines and cobwebs on the surface, creating a sudden, heart-stopping tension and the threat of imminent, catastrophic collapse.", + "seed": 10024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_6564_seed_10024_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": -0.08864420354366302, + "vlm_logprob": -0.08864420354366302, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08864420354366302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0099415872246027 + ], + "opposite_logprob_tokens": [ + -5.009941577911377 + ], + "logprob_reward": -0.0099415872246027, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0099415872246027, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11612503230571747, + 0.031040353700518608, + 0.07258471846580505, + 0.11375965178012848, + 0.1479116678237915, + 0.18247215449810028, + 0.1896168291568756, + 0.0475723072886467, + 0.03430728614330292, + 0.06461001187562943 + ], + "credit_map": [ + 0.013311339542269707, + 0.01250731386244297, + 0.01465991698205471, + 0.01677955687046051, + 0.018132664263248444, + 0.016473328694701195, + 0.013410752639174461, + 0.010850160382688046, + 0.005496092140674591, + 0.0025044248905032873, + 0.004844645969569683, + 0.0028270746115595102, + 0.0035742376931011677, + 0.004736477974802256, + 0.002830268582329154, + 0.004227133467793465, + 0.009944991208612919, + 0.01008206233382225, + 0.0066480147652328014, + 0.009035694412887096, + 0.010861754417419434, + 0.008104261942207813, + 0.010372243821620941, + 0.0075356923043727875, + 0.011411030776798725, + 0.01787523552775383, + 0.01310758013278246, + 0.013530646450817585, + 0.01626279018819332, + 0.009055663831532001, + 0.013196108862757683, + 0.019320594146847725, + 0.016152527183294296, + 0.017488999292254448, + 0.014900525100529194, + 0.018327323719859123, + 0.020549893379211426, + 0.025322426110506058, + 0.01346334908157587, + 0.021706633269786835, + 0.01151782926172018, + 0.030867034569382668, + 0.01801224984228611, + 0.01897832378745079, + 0.03466663509607315, + 0.014522481709718704, + 0.029436267912387848, + 0.024471338838338852, + 0.01858598366379738, + 0.03221158683300018, + 0.019567033275961876, + 0.02498452365398407, + 0.01906348019838333, + 0.02649633213877678, + 0.023134762421250343, + 0.025573113933205605, + 0.0034438730217516422, + 0.00439060851931572, + 0.014496928080916405, + 0.012358469888567924, + 0.003058375557884574, + 0.004023745656013489, + 0.002788712503388524, + 0.0030115973204374313, + 0.002859658794477582, + 0.004630940034985542, + 0.003950534854084253, + 0.004735748283565044, + 0.0035455836914479733, + 0.004096310585737228, + 0.0048385970294475555, + 0.005649914965033531, + 0.004567546769976616, + 0.00665827002376318, + 0.006423471495509148, + 0.0050166212022304535, + 0.004495435394346714, + 0.0173177532851696, + 0.0039474801160395145, + 0.016183428466320038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.0099415872246027 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023309852927923203 + ], + "opposite_logprob_tokens": [ + -3.7733099460601807 + ], + "logprob_reward": -0.023309852927923203, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023309852927923203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07350652664899826, + 0.01859270967543125, + 0.060009267181158066, + 0.08115002512931824, + 0.09910842031240463, + 0.04689920321106911, + 0.14284829795360565, + 0.1480124145746231, + 0.0500456728041172, + 0.27982744574546814 + ], + "credit_map": [ + 0.005719202570617199, + 0.007533828262239695, + 0.009373227134346962, + 0.011822780594229698, + 0.01213616319000721, + 0.008620387874543667, + 0.011554998345673084, + 0.006745940539985895, + 0.0029338879976421595, + 0.001602535485289991, + 0.00280372123233974, + 0.0019834088161587715, + 0.0017602319130674005, + 0.00187193404417485, + 0.0030130355153232813, + 0.0026239538565278053, + 0.006998657248914242, + 0.006479238625615835, + 0.0055473740212619305, + 0.009006069041788578, + 0.010611825622618198, + 0.004658510908484459, + 0.010661173611879349, + 0.006046417634934187, + 0.007216852158308029, + 0.008194650523364544, + 0.007778897415846586, + 0.017102347686886787, + 0.010107731446623802, + 0.00443792250007391, + 0.0129181407392025, + 0.013393481262028217, + 0.010816235095262527, + 0.009114027954638004, + 0.015658793970942497, + 0.0073602464981377125, + 0.009347742423415184, + 0.02278187684714794, + 0.007483303546905518, + 0.01654619351029396, + 0.004996458068490028, + 0.004196780268102884, + 0.011009499430656433, + 0.0042339288629591465, + 0.004588955547660589, + 0.004364411812275648, + 0.004994886927306652, + 0.0085142832249403, + 0.013735397718846798, + 0.02530973218381405, + 0.01447116769850254, + 0.01934325136244297, + 0.015666408464312553, + 0.019199704751372337, + 0.01878770999610424, + 0.016334937885403633, + 0.011374806985259056, + 0.0208117812871933, + 0.0176699161529541, + 0.017511464655399323, + 0.016285676509141922, + 0.02550327405333519, + 0.012576915323734283, + 0.02627858705818653, + 0.003623207099735737, + 0.008256640285253525, + 0.0056226300075650215, + 0.007939721457660198, + 0.0060013397596776485, + 0.0065046800300478935, + 0.005538338329643011, + 0.006559114903211594, + 0.023198243230581284, + 0.028729742392897606, + 0.014400383457541466, + 0.012388271279633045, + 0.014987858943641186, + 0.0919298306107521, + 0.015952875837683678, + 0.07824023813009262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.023309852927923203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34857380390167236 + ], + "opposite_logprob_tokens": [ + -1.2235738039016724 + ], + "logprob_reward": -0.34857380390167236, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34857380390167236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050354935228824615, + 0.028017712756991386, + 0.04260514676570892, + 0.351509690284729, + 0.09401539713144302, + 0.07736419141292572, + 0.10932943969964981, + 0.1716858595609665, + 0.024667585268616676, + 0.05045008659362793 + ], + "credit_map": [ + 0.006714476738125086, + 0.005573399364948273, + 0.006237150169909, + 0.00948937889188528, + 0.0050154561176896095, + 0.004216336645185947, + 0.009216447360813618, + 0.0038922845851629972, + 0.005036122631281614, + 0.002819571876898408, + 0.0034414813853800297, + 0.002908889437094331, + 0.0031988113187253475, + 0.00257514207623899, + 0.00416752602905035, + 0.003870165441185236, + 0.006214980501681566, + 0.004116525407880545, + 0.004565056879073381, + 0.004143234807997942, + 0.005466908682137728, + 0.005698023363947868, + 0.004463339690119028, + 0.007937074638903141, + 0.030809123069047928, + 0.052330370992422104, + 0.03873489424586296, + 0.058831993490457535, + 0.04337979480624199, + 0.027390040457248688, + 0.04886874556541443, + 0.05116472393274307, + 0.008658001199364662, + 0.01334224920719862, + 0.007729590404778719, + 0.018203413113951683, + 0.015500056557357311, + 0.01073300838470459, + 0.009369982406497002, + 0.010479087010025978, + 0.011783644556999207, + 0.004059022292494774, + 0.004183542914688587, + 0.00893339328467846, + 0.005802638363093138, + 0.00510900653898716, + 0.015436235815286636, + 0.02205670438706875, + 0.005748824682086706, + 0.011593679897487164, + 0.006152122747153044, + 0.010099060833454132, + 0.009687197394669056, + 0.01730262115597725, + 0.023447200655937195, + 0.025298723950982094, + 0.012151697650551796, + 0.012025462463498116, + 0.05611199513077736, + 0.04701678827404976, + 0.009313163347542286, + 0.012717667035758495, + 0.007676719222217798, + 0.014672355726361275, + 0.0015159521717578173, + 0.0033309455029666424, + 0.002098482334986329, + 0.0033851347398012877, + 0.0037480255123227835, + 0.002658040728420019, + 0.004116403870284557, + 0.0038145985454320908, + 0.004073573742061853, + 0.004370832350105047, + 0.003528278786689043, + 0.0017788701225072145, + 0.0026845443062484264, + 0.01457142736762762, + 0.0037364137824624777, + 0.015706144273281097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.34857380390167236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300768807530403 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023300768807530403, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300768807530403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039766378700733185, + 0.02412877045571804, + 0.022932643070816994, + 0.3129560053348541, + 0.4666617512702942, + 0.013400980271399021, + 0.018498027697205544, + 0.048576150089502335, + 0.02871842123568058, + 0.024360917508602142 + ], + "credit_map": [ + 0.004282331559807062, + 0.0037219561636447906, + 0.00600806949660182, + 0.0060171764343976974, + 0.004189088474959135, + 0.003612629836425185, + 0.00757288234308362, + 0.004362244624644518, + 0.003979440778493881, + 0.0028501669876277447, + 0.0032944739796221256, + 0.0020397996995598078, + 0.003001370932906866, + 0.00241846963763237, + 0.002814278472214937, + 0.003730770433321595, + 0.0034570619463920593, + 0.003143011126667261, + 0.0022870965767651796, + 0.002552187070250511, + 0.003529794979840517, + 0.0021139374002814293, + 0.0031208298169076443, + 0.0027287232223898172, + 0.027627265080809593, + 0.05336691811680794, + 0.03938706964254379, + 0.04403255879878998, + 0.04545454680919647, + 0.021898146718740463, + 0.04402025043964386, + 0.03716925531625748, + 0.03739652782678604, + 0.046009887009859085, + 0.08625391125679016, + 0.055865753442049026, + 0.05654105916619301, + 0.10981158912181854, + 0.034143295139074326, + 0.040639728307724, + 0.0008962209103628993, + 0.0010891716228798032, + 0.002030800562351942, + 0.0018291122978553176, + 0.001631803228519857, + 0.0009770009201020002, + 0.003504005027934909, + 0.0014428659342229366, + 0.0018260573269799352, + 0.0020938310772180557, + 0.0010607101721689105, + 0.00219342322088778, + 0.0016468806425109506, + 0.002261251676827669, + 0.003882763674482703, + 0.003533109789714217, + 0.002136507537215948, + 0.0051699052564799786, + 0.00927152018994093, + 0.017775217071175575, + 0.0053027840331196785, + 0.002475669141858816, + 0.0036122493911534548, + 0.0028322963044047356, + 0.002144482685253024, + 0.003880580887198448, + 0.002969555091112852, + 0.003909980412572622, + 0.0038461238145828247, + 0.00359330209903419, + 0.004788996651768684, + 0.0035853993613272905, + 0.0018752419855445623, + 0.0027110713999718428, + 0.0026767232920974493, + 0.0044305892661213875, + 0.0025051673874258995, + 0.003797434503212571, + 0.003785534994676709, + 0.0025791535153985023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.023300768807530403 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095004856586456 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038095004856586456, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095004856586456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039953574538230896, + 0.02017299458384514, + 0.014453372918069363, + 0.04481826722621918, + 0.14559757709503174, + 0.16953128576278687, + 0.06432521343231201, + 0.3137739300727844, + 0.04275562986731529, + 0.1446181833744049 + ], + "credit_map": [ + 0.007382963318377733, + 0.0039747185073792934, + 0.0063378154300153255, + 0.005084990058094263, + 0.003623412922024727, + 0.003702927613630891, + 0.005998549051582813, + 0.003848195541650057, + 0.0032739564776420593, + 0.0025341350119560957, + 0.0021761239040642977, + 0.001577467774040997, + 0.002307616639882326, + 0.0035477830097079277, + 0.002091210102662444, + 0.002664702944457531, + 0.001956805121153593, + 0.0016897858586162329, + 0.0012114522978663445, + 0.001681244233623147, + 0.0019005127251148224, + 0.0019392160465940833, + 0.0016106081893667579, + 0.002463748212903738, + 0.0043513658456504345, + 0.005880012176930904, + 0.004706160631030798, + 0.008232975378632545, + 0.005738695152103901, + 0.0030135733541101217, + 0.007083880249410868, + 0.005811604671180248, + 0.012254013679921627, + 0.016825612634420395, + 0.024001095443964005, + 0.017142971977591515, + 0.018958622589707375, + 0.032204821705818176, + 0.011623521335422993, + 0.012586918659508228, + 0.016625050455331802, + 0.013529938645660877, + 0.05156776309013367, + 0.013356250710785389, + 0.01280116755515337, + 0.010245876386761665, + 0.012786678038537502, + 0.03861856088042259, + 0.005088565405458212, + 0.0064358520321547985, + 0.004553896375000477, + 0.006902959197759628, + 0.004171786364167929, + 0.009761125780642033, + 0.012295442633330822, + 0.015115582384169102, + 0.022663215175271034, + 0.019354455173015594, + 0.09202840179204941, + 0.10457563400268555, + 0.020703712478280067, + 0.01806117594242096, + 0.017501339316368103, + 0.018886003643274307, + 0.003090701764449477, + 0.007234625983983278, + 0.004868100397288799, + 0.005459494423121214, + 0.005618819501250982, + 0.005479051731526852, + 0.005860186647623777, + 0.005144649650901556, + 0.011556540615856647, + 0.014696790836751461, + 0.008416689932346344, + 0.005644784774631262, + 0.006192018743604422, + 0.04691389575600624, + 0.008443274535238743, + 0.04275418445467949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.038095004856586456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot captures a massive bronze bull statue, its muscular body gleaming in the blazing midday sun. Set across a vast city square, the camera suddenly shakes violently, mimicking the force of an earth-shaking tremor that sends ripples through the stone pavement. Following the earthquake, the massive statue begins to teeter on its massive granite base, its surface a mixture of polished bronze and a weathered green patina. It tilts forward, its immense weight undergoing a sickening, harrowing shift, its powerful form tilting perilously before swaying backward with terrifying momentum. Dust and small stone fragments erupt from the base, where the statue's hooves scrape and grind against it, leaving tiny crack lines and cobwebs on the surface, creating a sudden, heart-stopping tension and the threat of imminent, catastrophic collapse.", + "seed": 10025, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_6564_seed_10025_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": -0.12751287259161473, + "vlm_logprob": -0.12751287259161473, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.12751287259161473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25673097372055054 + ], + "opposite_logprob_tokens": [ + -1.5067310333251953 + ], + "logprob_reward": -0.25673097372055054, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25673097372055054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5614151358604431, + 0.04980303719639778, + 0.13814368844032288, + 0.020835692062973976, + 0.0396774560213089, + 0.022181827574968338, + 0.04161582142114639, + 0.07614543288946152, + 0.013740385882556438, + 0.03644143044948578 + ], + "credit_map": [ + 0.08357606083154678, + 0.09234808385372162, + 0.06434353440999985, + 0.08434917777776718, + 0.054073404520750046, + 0.05120689049363136, + 0.07027800381183624, + 0.061240073293447495, + 0.004829107318073511, + 0.006766038481146097, + 0.005233343690633774, + 0.005182391963899136, + 0.006426052190363407, + 0.010858394205570221, + 0.0046455031260848045, + 0.005862214136868715, + 0.012493438087403774, + 0.03531818464398384, + 0.019646473228931427, + 0.016125235706567764, + 0.014179999940097332, + 0.011308827437460423, + 0.011189988814294338, + 0.017881549894809723, + 0.002133537782356143, + 0.0035352869890630245, + 0.002352083334699273, + 0.002766701392829418, + 0.002152161207050085, + 0.0021333724725991488, + 0.0032138952519744635, + 0.0025486559607088566, + 0.0026122902054339647, + 0.0033528737258166075, + 0.005755997262895107, + 0.0036345135886222124, + 0.004921309184283018, + 0.004141141194850206, + 0.007221702486276627, + 0.008037635125219822, + 0.0012755942298099399, + 0.0009446791373193264, + 0.0031882787588983774, + 0.004438227508217096, + 0.0021818033419549465, + 0.001262159552425146, + 0.00665763346478343, + 0.0022334526292979717, + 0.0036426205188035965, + 0.003049337537959218, + 0.005023738835006952, + 0.006281089503318071, + 0.006820645648986101, + 0.00686399033293128, + 0.004434195347130299, + 0.005500206258147955, + 0.006480933632701635, + 0.011157485656440258, + 0.0058746254071593285, + 0.020429089665412903, + 0.006075096782296896, + 0.007867666892707348, + 0.010698948055505753, + 0.007561593782156706, + 0.0009755379869602621, + 0.0019063024083152413, + 0.0019456557929515839, + 0.0014312553685158491, + 0.0013323494931682944, + 0.001686836127191782, + 0.0024803353007882833, + 0.001982114277780056, + 0.005306611768901348, + 0.005503281485289335, + 0.0042772451415658, + 0.0035398066975176334, + 0.0043122912757098675, + 0.0026399355847388506, + 0.007109600584954023, + 0.00375265977345407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.25673097372055054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018214579671621323 + ], + "opposite_logprob_tokens": [ + -4.018214702606201 + ], + "logprob_reward": -0.018214579671621323, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018214579671621323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13630160689353943, + 0.09136299043893814, + 0.10207442194223404, + 0.06549538671970367, + 0.0774533748626709, + 0.034732136875391006, + 0.10249532759189606, + 0.07175325602293015, + 0.23426488041877747, + 0.08406669646501541 + ], + "credit_map": [ + 0.015110276639461517, + 0.01882508210837841, + 0.012878035195171833, + 0.01749706268310547, + 0.03285703808069229, + 0.01114482432603836, + 0.015546758659183979, + 0.012442512437701225, + 0.012848157435655594, + 0.011170392856001854, + 0.01150835957378149, + 0.008420631289482117, + 0.009534381330013275, + 0.01737232320010662, + 0.011760852299630642, + 0.008747885003685951, + 0.008018682710826397, + 0.024156896397471428, + 0.014676478691399097, + 0.012074789963662624, + 0.009549888782203197, + 0.00974606815725565, + 0.008894716389477253, + 0.014956891536712646, + 0.005325314588844776, + 0.007709784433245659, + 0.00824348721653223, + 0.012987551279366016, + 0.008875414729118347, + 0.006532962899655104, + 0.008553927764296532, + 0.007266940549015999, + 0.007160321809351444, + 0.005599313415586948, + 0.008332628756761551, + 0.007525784894824028, + 0.009821944870054722, + 0.009098333306610584, + 0.011341491714119911, + 0.018573546782135963, + 0.0031315775122493505, + 0.0038659318815916777, + 0.0037503472995013, + 0.004667337518185377, + 0.0035165916197001934, + 0.004917042329907417, + 0.006000958848744631, + 0.004882349167019129, + 0.007169749587774277, + 0.005964440293610096, + 0.016998032107949257, + 0.03188709542155266, + 0.009116659872233868, + 0.007311650086194277, + 0.013450147584080696, + 0.01059754192829132, + 0.007472677156329155, + 0.008139251731336117, + 0.0064970203675329685, + 0.016147390007972717, + 0.008105850778520107, + 0.006966445129364729, + 0.009437788277864456, + 0.00898682326078415, + 0.01514368038624525, + 0.017110202461481094, + 0.03538699075579643, + 0.047292426228523254, + 0.01632145419716835, + 0.023656414821743965, + 0.04951425641775131, + 0.029839422553777695, + 0.007442656904459, + 0.006478014402091503, + 0.0223491620272398, + 0.005473713390529156, + 0.005045520607382059, + 0.00455642631277442, + 0.026990430429577827, + 0.005730762146413326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.018214579671621323 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28128644824028015 + ], + "opposite_logprob_tokens": [ + -1.4062864780426025 + ], + "logprob_reward": -0.28128644824028015, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28128644824028015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1780664622783661, + 0.05156105011701584, + 0.08417780697345734, + 0.07734591513872147, + 0.05332985147833824, + 0.035967279225587845, + 0.1928960382938385, + 0.03569527715444565, + 0.11935818940401077, + 0.17160208523273468 + ], + "credit_map": [ + 0.038612086325883865, + 0.025774504989385605, + 0.01992662623524666, + 0.017436182126402855, + 0.031195178627967834, + 0.01933218352496624, + 0.014051412232220173, + 0.011738299392163754, + 0.005446607246994972, + 0.004219249356538057, + 0.005163245368748903, + 0.003409502562135458, + 0.010450413450598717, + 0.010551078245043755, + 0.004230613354593515, + 0.008090349845588207, + 0.00764558045193553, + 0.016669118776917458, + 0.011727291159331799, + 0.010119910351932049, + 0.007444214075803757, + 0.006631823722273111, + 0.011418920010328293, + 0.01252095215022564, + 0.007040283642709255, + 0.012564257718622684, + 0.009290889836847782, + 0.012985719367861748, + 0.011023028753697872, + 0.00751122273504734, + 0.007257235236465931, + 0.0096732871606946, + 0.005089126527309418, + 0.007434061728417873, + 0.00559749873355031, + 0.008137728087604046, + 0.005723514128476381, + 0.0048538316041231155, + 0.00963694229722023, + 0.0068571497686207294, + 0.0028622723184525967, + 0.0025329324416816235, + 0.0048169828951358795, + 0.008275927975773811, + 0.003100105095654726, + 0.0024400774855166674, + 0.008278636261820793, + 0.0036603487096726894, + 0.012721649371087551, + 0.011076891794800758, + 0.03913766145706177, + 0.06038522347807884, + 0.01549102459102869, + 0.014281325973570347, + 0.024670250713825226, + 0.015132025815546513, + 0.0041467878036201, + 0.003021219978109002, + 0.00258994079194963, + 0.00480671739205718, + 0.0037963883951306343, + 0.004264738876372576, + 0.006996135227382183, + 0.006073351483792067, + 0.010442550294101238, + 0.014246407896280289, + 0.013165965676307678, + 0.026181159541010857, + 0.011621606536209583, + 0.010109860450029373, + 0.016379371285438538, + 0.017211278900504112, + 0.021757593378424644, + 0.02706325240433216, + 0.025738384574651718, + 0.015068055130541325, + 0.019623763859272003, + 0.01598026044666767, + 0.028285769745707512, + 0.018085023388266563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.28128644824028015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368023827672005 + ], + "opposite_logprob_tokens": [ + -3.6513681411743164 + ], + "logprob_reward": -0.026368023827672005, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368023827672005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12587538361549377, + 0.09351950138807297, + 0.06771805137395859, + 0.13012830913066864, + 0.02675923891365528, + 0.0923207476735115, + 0.20989596843719482, + 0.10854557156562805, + 0.0665200874209404, + 0.07871706038713455 + ], + "credit_map": [ + 0.01598058082163334, + 0.01865966245532036, + 0.020650118589401245, + 0.01669059321284294, + 0.011483893729746342, + 0.010588286444544792, + 0.011872273869812489, + 0.01994999125599861, + 0.013087516650557518, + 0.012301033362746239, + 0.010321595706045628, + 0.007453189231455326, + 0.0088339913636446, + 0.022053593769669533, + 0.008054650388658047, + 0.01141393557190895, + 0.00592269329354167, + 0.013168254867196083, + 0.008803662844002247, + 0.008264344185590744, + 0.006564589217305183, + 0.007654027082026005, + 0.007379747461527586, + 0.00996073242276907, + 0.009748013690114021, + 0.013581202365458012, + 0.016754131764173508, + 0.029650231823325157, + 0.01873839646577835, + 0.010159033350646496, + 0.01893312856554985, + 0.01256418414413929, + 0.002254381775856018, + 0.002083864761516452, + 0.002908567665144801, + 0.0036989080253988504, + 0.0031150744762271643, + 0.003220578655600548, + 0.003544229082763195, + 0.005933636333793402, + 0.007558342069387436, + 0.008210328407585621, + 0.011663036420941353, + 0.017144033685326576, + 0.008211411535739899, + 0.0064579034224152565, + 0.022480305284261703, + 0.01059538684785366, + 0.0121986148878932, + 0.015536286868155003, + 0.042549632489681244, + 0.06846152991056442, + 0.014741925522685051, + 0.01887821964919567, + 0.02162083052098751, + 0.015908941626548767, + 0.010521680116653442, + 0.012568662874400616, + 0.008145289495587349, + 0.027430344372987747, + 0.01005538646131754, + 0.01278772670775652, + 0.014879373833537102, + 0.012157121673226357, + 0.006248851306736469, + 0.007963255047798157, + 0.011019237339496613, + 0.008100842125713825, + 0.006395653355866671, + 0.008803846314549446, + 0.011407747864723206, + 0.006580660119652748, + 0.010776671580970287, + 0.015667540952563286, + 0.012747327797114849, + 0.009351389482617378, + 0.006561901420354843, + 0.007305794395506382, + 0.008319554850459099, + 0.007986889220774174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.026368023827672005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549643374979496 + ], + "opposite_logprob_tokens": [ + -2.929964303970337 + ], + "logprob_reward": -0.0549643374979496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549643374979496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12618789076805115, + 0.10207228362560272, + 0.1221121996641159, + 0.07437962293624878, + 0.04040084406733513, + 0.04471886157989502, + 0.2111712098121643, + 0.05541975796222687, + 0.045626137405633926, + 0.1779112070798874 + ], + "credit_map": [ + 0.011888634413480759, + 0.014728279784321785, + 0.015530661679804325, + 0.015627026557922363, + 0.019331051036715508, + 0.010875318199396133, + 0.012101560831069946, + 0.026105357334017754, + 0.014079166576266289, + 0.012207655236124992, + 0.010957146994769573, + 0.007741654757410288, + 0.0109188137575984, + 0.02365639992058277, + 0.011411266401410103, + 0.011100188829004765, + 0.008959879167377949, + 0.025527706369757652, + 0.016721850261092186, + 0.011202691122889519, + 0.011849967762827873, + 0.011382517404854298, + 0.014874263666570187, + 0.021593322977423668, + 0.006976859178394079, + 0.00981885101646185, + 0.008747056126594543, + 0.013367809355258942, + 0.010082448832690716, + 0.0058191996067762375, + 0.009703649207949638, + 0.009863743558526039, + 0.002888198709115386, + 0.004197155125439167, + 0.003969932906329632, + 0.004029449541121721, + 0.006068467628210783, + 0.004000859335064888, + 0.006624855566769838, + 0.008621926419436932, + 0.0027653963770717382, + 0.002161772456020117, + 0.0065927887335419655, + 0.00799790769815445, + 0.003918364178389311, + 0.003563385456800461, + 0.012113804928958416, + 0.005605438258498907, + 0.014252426102757454, + 0.01663897931575775, + 0.03967434912919998, + 0.06619158387184143, + 0.01730496436357498, + 0.01766354963183403, + 0.020973768085241318, + 0.01847158558666706, + 0.004636677447706461, + 0.007142066024243832, + 0.00387173262424767, + 0.01364689040929079, + 0.004052221309393644, + 0.0054278322495520115, + 0.007250024937093258, + 0.009392310865223408, + 0.004207435995340347, + 0.004921779967844486, + 0.006616187747567892, + 0.008074428886175156, + 0.005176117643713951, + 0.005014521069824696, + 0.0065511600114405155, + 0.005064505618065596, + 0.020555760711431503, + 0.02421773038804531, + 0.027906788513064384, + 0.013664303347468376, + 0.01569674164056778, + 0.01573195867240429, + 0.043065764009952545, + 0.0170721635222435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0549643374979496 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic low-angle wide shot captures a massive bronze bull statue, its muscular body gleaming in the blazing midday sun. Set across a vast city square, the camera suddenly shakes violently, mimicking the force of an earth-shaking tremor that sends ripples through the stone pavement. Following the earthquake, the massive statue begins to teeter on its massive granite base, its surface a mixture of polished bronze and a weathered green patina. It tilts forward, its immense weight undergoing a sickening, harrowing shift, its powerful form tilting perilously before swaying backward with terrifying momentum. Dust and small stone fragments erupt from the base, where the statue's hooves scrape and grind against it, leaving tiny crack lines and cobwebs on the surface, creating a sudden, heart-stopping tension and the threat of imminent, catastrophic collapse.", + "seed": 10026, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_14/index_6564_seed_10026_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": -0.047677562572062016, + "vlm_logprob": -0.047677562572062016, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.047677562572062016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009573973715305328 + ], + "opposite_logprob_tokens": [ + -4.884573936462402 + ], + "logprob_reward": -0.009573973715305328, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.009573973715305328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24966023862361908, + 0.29907622933387756, + 0.06661826372146606, + 0.0714128166437149, + 0.05155550688505173, + 0.03126385062932968, + 0.028941059485077858, + 0.0872395783662796, + 0.043782323598861694, + 0.07045011967420578 + ], + "credit_map": [ + 0.027724703773856163, + 0.034453269094228745, + 0.029696887359023094, + 0.02845434471964836, + 0.03661372885107994, + 0.030818289145827293, + 0.03491627424955368, + 0.0269827451556921, + 0.018347179517149925, + 0.029266370460391045, + 0.009554507210850716, + 0.011707782745361328, + 0.1234247088432312, + 0.07020924985408783, + 0.02187645435333252, + 0.014689967967569828, + 0.006884430069476366, + 0.008212975226342678, + 0.011770624667406082, + 0.006467859260737896, + 0.010387997142970562, + 0.006473386660218239, + 0.011397979222238064, + 0.00502301100641489, + 0.005911362357437611, + 0.008396322838962078, + 0.009338852018117905, + 0.006415545474737883, + 0.014828571118414402, + 0.013239198364317417, + 0.006295942235738039, + 0.006987019442021847, + 0.006643281318247318, + 0.006547699682414532, + 0.007413490209728479, + 0.007639424875378609, + 0.006407762877643108, + 0.004793068859726191, + 0.007247986737638712, + 0.004862791392952204, + 0.0035615439992398024, + 0.0032631526701152325, + 0.004104096908122301, + 0.0034721074625849724, + 0.005083295051008463, + 0.003776485100388527, + 0.005364886950701475, + 0.0026382808573544025, + 0.005535998847335577, + 0.0027670999988913536, + 0.0032816273160278797, + 0.003687104908749461, + 0.0037647108547389507, + 0.003241621656343341, + 0.0033199191093444824, + 0.003342975163832307, + 0.006659303326159716, + 0.009101801551878452, + 0.010829715058207512, + 0.011322101578116417, + 0.009207384660840034, + 0.007283140905201435, + 0.013394193723797798, + 0.019441941753029823, + 0.004592850338667631, + 0.005493999924510717, + 0.0058263326063752174, + 0.01005623210221529, + 0.003103365655988455, + 0.005455269943922758, + 0.005195740144699812, + 0.004058535676449537, + 0.008981375955045223, + 0.009722203947603703, + 0.0115507235750556, + 0.0069315796718001366, + 0.0044477577321231365, + 0.006460890173912048, + 0.013510460965335369, + 0.008845129981637001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.009573973715305328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208228327333927 + ], + "opposite_logprob_tokens": [ + -4.2642083168029785 + ], + "logprob_reward": -0.014208228327333927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208228327333927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.175740584731102, + 0.30485856533050537, + 0.04527511075139046, + 0.03746481239795685, + 0.07793622463941574, + 0.049800559878349304, + 0.03954748436808586, + 0.05883294716477394, + 0.15246877074241638, + 0.058074917644262314 + ], + "credit_map": [ + 0.015174069441854954, + 0.019945502281188965, + 0.018507909029722214, + 0.021354256197810173, + 0.026994191110134125, + 0.022501030936837196, + 0.030329549685120583, + 0.020934073254466057, + 0.04885892570018768, + 0.034129705280065536, + 0.013186526484787464, + 0.01394741702824831, + 0.1285918802022934, + 0.028094083070755005, + 0.02427002042531967, + 0.013779990375041962, + 0.004652072209864855, + 0.006281276699155569, + 0.00839399453252554, + 0.004948692861944437, + 0.00449652224779129, + 0.0028440640307962894, + 0.009861933067440987, + 0.003796559991315007, + 0.004233154933899641, + 0.004946185741573572, + 0.0056784190237522125, + 0.004054037854075432, + 0.006202136632055044, + 0.004628542345017195, + 0.0034873380791395903, + 0.004234998021274805, + 0.013927841559052467, + 0.009656419046223164, + 0.01626170426607132, + 0.008359909988939762, + 0.007123306859284639, + 0.00915644783526659, + 0.007719092071056366, + 0.0057315099984407425, + 0.006502606440335512, + 0.008038172498345375, + 0.007364022545516491, + 0.004062584601342678, + 0.007475539110600948, + 0.006513664964586496, + 0.005335311405360699, + 0.00450865738093853, + 0.004077616147696972, + 0.004737734328955412, + 0.004145129583775997, + 0.0074191465973854065, + 0.005823509767651558, + 0.004651488270610571, + 0.00546399736776948, + 0.003228861838579178, + 0.007809422444552183, + 0.005868050269782543, + 0.006645918358117342, + 0.010626151226460934, + 0.004596581216901541, + 0.009097942151129246, + 0.007757700979709625, + 0.006431180983781815, + 0.014841953292489052, + 0.01805775985121727, + 0.020112168043851852, + 0.04168412834405899, + 0.011584704741835594, + 0.017855312675237656, + 0.016811855137348175, + 0.011520905420184135, + 0.006632482632994652, + 0.010407406836748123, + 0.008590918965637684, + 0.007614024914801121, + 0.005064425524324179, + 0.007222032640129328, + 0.006639352533966303, + 0.005904269404709339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.014208228327333927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16031967103481293 + ], + "opposite_logprob_tokens": [ + -1.9103196859359741 + ], + "logprob_reward": -0.16031967103481293, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16031967103481293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10914582014083862, + 0.41455090045928955, + 0.04248818755149841, + 0.015026147477328777, + 0.2125820368528366, + 0.034962259232997894, + 0.04550851508975029, + 0.06734604388475418, + 0.03522760048508644, + 0.023162441328167915 + ], + "credit_map": [ + 0.00928127858787775, + 0.015137131325900555, + 0.00909476075321436, + 0.014747433364391327, + 0.017626671120524406, + 0.01066615805029869, + 0.022739127278327942, + 0.009853261522948742, + 0.0347401425242424, + 0.04416811466217041, + 0.013259729370474815, + 0.016887051984667778, + 0.1807333379983902, + 0.0752507746219635, + 0.030291825532913208, + 0.019219903275370598, + 0.005580767523497343, + 0.004367103334516287, + 0.007822797633707523, + 0.003435330232605338, + 0.005256848409771919, + 0.004232918377965689, + 0.007479213178157806, + 0.004313213750720024, + 0.0019023598870262504, + 0.0015158412279561162, + 0.0022380435839295387, + 0.001653096522204578, + 0.0019752196967601776, + 0.002948788460344076, + 0.0015308449510484934, + 0.0012619530316442251, + 0.016671789810061455, + 0.04002898186445236, + 0.018023919314146042, + 0.03220813721418381, + 0.040954262018203735, + 0.019798101857304573, + 0.027118850499391556, + 0.017778003588318825, + 0.0034929984249174595, + 0.0036909165792167187, + 0.0034334734082221985, + 0.0031945286318659782, + 0.006315573584288359, + 0.006348256953060627, + 0.004832176491618156, + 0.003654336091130972, + 0.00530812656506896, + 0.0051736366003751755, + 0.0045213038101792336, + 0.008716260083019733, + 0.006114637944847345, + 0.004976070951670408, + 0.0055979229509830475, + 0.005100557580590248, + 0.004340939223766327, + 0.004380327649414539, + 0.009116710163652897, + 0.006010688375681639, + 0.004029238596558571, + 0.006428215652704239, + 0.014507662504911423, + 0.018532266840338707, + 0.0033528052736073732, + 0.003381405957043171, + 0.007260792423039675, + 0.007627975661307573, + 0.0028610187582671642, + 0.003908875398337841, + 0.003319504437968135, + 0.003515219781547785, + 0.002050441922619939, + 0.004076248966157436, + 0.0038194763474166393, + 0.002578041050583124, + 0.0015739869559183717, + 0.0025530748534947634, + 0.003489141119644046, + 0.003022029995918274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.16031967103481293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1670348048210144, + 0.46893590688705444, + 0.03417045250535011, + 0.027875248342752457, + 0.058632176369428635, + 0.04491233080625534, + 0.02886507287621498, + 0.08256997168064117, + 0.044893186539411545, + 0.04211093857884407 + ], + "credit_map": [ + 0.014134017750620842, + 0.029953310266137123, + 0.014079630374908447, + 0.020715346559882164, + 0.031005656346678734, + 0.013855118304491043, + 0.028964489698410034, + 0.014327223412692547, + 0.030580731108784676, + 0.04465203732252121, + 0.01487568486481905, + 0.0189182311296463, + 0.1965990662574768, + 0.10659529268741608, + 0.03370392322540283, + 0.023010896518826485, + 0.0046575418673455715, + 0.002706554252654314, + 0.006576545536518097, + 0.0027930953074246645, + 0.006297636777162552, + 0.003276796778663993, + 0.004965322092175484, + 0.0028969573322683573, + 0.0024061230942606926, + 0.003321350086480379, + 0.0027433184441179037, + 0.004372280091047287, + 0.004210390616208315, + 0.003942849580198526, + 0.003996811807155609, + 0.0028821215964853764, + 0.007800857536494732, + 0.008375711739063263, + 0.011384195648133755, + 0.00683843158185482, + 0.006852013990283012, + 0.0059596518985927105, + 0.007194460369646549, + 0.004226852208375931, + 0.0053552608005702496, + 0.00577785586938262, + 0.00812782347202301, + 0.004638709127902985, + 0.005640245974063873, + 0.006677994504570961, + 0.0045592146925628185, + 0.0041352249681949615, + 0.0037860877346247435, + 0.0027605541981756687, + 0.003266876097768545, + 0.005802736151963472, + 0.003718688152730465, + 0.0022709444165229797, + 0.004253826104104519, + 0.0030053576920181513, + 0.008283650502562523, + 0.0053251441568136215, + 0.011373387649655342, + 0.007960711605846882, + 0.005594944581389427, + 0.007130292244255543, + 0.01690349169075489, + 0.0199983399361372, + 0.0048030661419034, + 0.003387578995898366, + 0.009823125787079334, + 0.009410658851265907, + 0.004009949043393135, + 0.003518804907798767, + 0.0051522389985620975, + 0.004787758458405733, + 0.006172122899442911, + 0.004716941155493259, + 0.006792968139052391, + 0.00608733668923378, + 0.00362905440852046, + 0.004590630531311035, + 0.0051750424318015575, + 0.004946834407746792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.020590949803590775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 bronze bull statue with a muscular body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the statue's surface a mixture of polished bronze and weathered green patina?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle wide shot and then shake violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the statue tilt forward on its granite base, causing dust and stone fragments to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are tiny crack lines and cobwebs visible on the statue's surface and base?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369498997926712 + ], + "opposite_logprob_tokens": [ + -3.4086949825286865 + ], + "logprob_reward": -0.03369498997926712, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369498997926712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2056046426296234, + 0.3805742561817169, + 0.026700522750616074, + 0.03307168558239937, + 0.10904493927955627, + 0.03889833390712738, + 0.06907748430967331, + 0.04288385808467865, + 0.04941638559103012, + 0.04472789913415909 + ], + "credit_map": [ + 0.018426865339279175, + 0.03249388188123703, + 0.017556916922330856, + 0.025912752375006676, + 0.03228697180747986, + 0.01960967667400837, + 0.03239347040653229, + 0.0269241314381361, + 0.02337416633963585, + 0.0381079725921154, + 0.013640085235238075, + 0.015933528542518616, + 0.15640167891979218, + 0.08733194321393967, + 0.02726883813738823, + 0.018516043201088905, + 0.003663875162601471, + 0.0024203695356845856, + 0.0031973044387996197, + 0.0024916883558034897, + 0.004780441522598267, + 0.004171883687376976, + 0.003309551626443863, + 0.002665409119799733, + 0.002572394907474518, + 0.004287981893867254, + 0.003208730136975646, + 0.004460289143025875, + 0.005819360725581646, + 0.004746384918689728, + 0.0050972397439181805, + 0.002879306674003601, + 0.007669130805879831, + 0.0187153909355402, + 0.010242992080748081, + 0.016011875122785568, + 0.022479774430394173, + 0.009267989546060562, + 0.016293466091156006, + 0.008364327251911163, + 0.003986854571849108, + 0.005088112782686949, + 0.004928904585540295, + 0.0038815406151115894, + 0.005205120425671339, + 0.006739309057593346, + 0.005123433656990528, + 0.003945058677345514, + 0.007966568693518639, + 0.00730501115322113, + 0.007510689087212086, + 0.014952331781387329, + 0.007931891828775406, + 0.00750360544770956, + 0.008451110683381557, + 0.007456276565790176, + 0.0048628998920321465, + 0.004200267139822245, + 0.005866586696356535, + 0.007950051687657833, + 0.0037597648333758116, + 0.006058946717530489, + 0.004204315133392811, + 0.00598102854564786, + 0.00525858486071229, + 0.0056205629371106625, + 0.0065095750615000725, + 0.012071140110492706, + 0.0031573749147355556, + 0.005433687008917332, + 0.006489451043307781, + 0.004876014310866594, + 0.006407545879483223, + 0.004377488512545824, + 0.007408793084323406, + 0.0055964309722185135, + 0.004017941188067198, + 0.00786504615098238, + 0.004653081297874451, + 0.004401571117341518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03369498997926712 + } + ], + "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/0000158/group_15/index_tensor([8503, 8503])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_tensor([8503, 8503])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c505b07888daa2e91aa8434c1d90defb4dad2332 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_tensor([8503, 8503])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, overhead tracking shot captures the young hero Nezha plummeting through a swirling, inky black cloud. His signature twin, jet-black buns, bound tightly by a red headband, tremble slightly in the swift currents of air. His face is defined by a furrowed brow and a resolute jawline. His lips are pursed, and his eyes burn with an unyielding fire. He rides upon a blazing Wind Fire Wheel, emitting blazing orange-red flames and visible swirling currents, the only constant source of warm light in the gloomy sky. The famous red Hun Tian Ling, like a living, shimmering silk dragon, violently twirls, coils, and stretches behind him, its dancing path tracing a flowing light trail through the air. The backdrop is a depressing apocalyptic scene, with deep blue, nearly black clouds forming a vast vortex. Dazzling purple lightning periodically rips through the sky, each flash instantly illuminating his red armor, emblazoned with golden lotus patterns, and his resolute face, casting a sharp, shifting shadow across him. Like a flaming meteorite, his entire body streaks steadily and decisively through the sky at astonishing speed, emitting a sharp whistling sound as he hurtles towards an unknown land. The entire scene is sharply contrasted by the cool tones of the sky and the fiery warm tones of Nezha himself, creating an epic sense of doom and a tense atmosphere filled with power and defiance.", + "seed": 31234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_8503_seed_31234_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": -0.07154247565194964, + "vlm_logprob": -0.07154247565194964, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.07154247565194964, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1452726274728775 + ], + "opposite_logprob_tokens": [ + -2.270272731781006 + ], + "logprob_reward": -0.1452726274728775, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1452726274728775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17026358842849731, + 0.23410694301128387, + 0.1034790500998497, + 0.06761931627988815, + 0.10765492171049118, + 0.0382915697991848, + 0.08743160218000412, + 0.06211515888571739, + 0.05596412718296051, + 0.07307372987270355 + ], + "credit_map": [ + 0.028180675581097603, + 0.01823616772890091, + 0.029426082968711853, + 0.023510461673140526, + 0.013455435633659363, + 0.01615181379020214, + 0.014552932232618332, + 0.026750000193715096, + 0.0331043004989624, + 0.02530120685696602, + 0.042299844324588776, + 0.05404694005846977, + 0.021281246095895767, + 0.019221602007746696, + 0.014743159525096416, + 0.024108637124300003, + 0.021938929334282875, + 0.00883933249861002, + 0.015426837839186192, + 0.016785234212875366, + 0.0072886343114078045, + 0.012547681108117104, + 0.008852038532495499, + 0.011800364591181278, + 0.011442883871495724, + 0.013472427614033222, + 0.00815343577414751, + 0.009665438905358315, + 0.008015316911041737, + 0.005299633368849754, + 0.006074221804738045, + 0.00549595570191741, + 0.02124517410993576, + 0.018917206674814224, + 0.027796484529972076, + 0.008691626600921154, + 0.00820859707891941, + 0.007665189448744059, + 0.008605340495705605, + 0.0065253013744950294, + 0.006663306150585413, + 0.002804224845021963, + 0.005148897413164377, + 0.004028943832963705, + 0.0035162896383553743, + 0.005500276573002338, + 0.004130160436034203, + 0.00649946928024292, + 0.012821674346923828, + 0.017679423093795776, + 0.012219890020787716, + 0.01226311270147562, + 0.011037045158445835, + 0.007644748780876398, + 0.007793225347995758, + 0.005972481332719326, + 0.009242172352969646, + 0.00814272090792656, + 0.007081185467541218, + 0.004969725850969553, + 0.006990525405853987, + 0.007997829467058182, + 0.01063499879091978, + 0.007055995985865593, + 0.008658247999846935, + 0.006383734289556742, + 0.006182631477713585, + 0.005050303880125284, + 0.005039955954998732, + 0.008059671148657799, + 0.007548848167061806, + 0.009040730074048042, + 0.020690536126494408, + 0.008743945509195328, + 0.008173135109245777, + 0.008183306083083153, + 0.006808451842516661, + 0.008379408158361912, + 0.006820485461503267, + 0.0052744667045772076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.1452726274728775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870939254760742 + ], + "opposite_logprob_tokens": [ + -3.0487093925476074 + ], + "logprob_reward": -0.04870939254760742, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870939254760742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.229975625872612, + 0.12499742209911346, + 0.09586260467767715, + 0.1022610142827034, + 0.2415522187948227, + 0.03960616886615753, + 0.04768870398402214, + 0.03680272400379181, + 0.04829183220863342, + 0.03296170011162758 + ], + "credit_map": [ + 0.034522268921136856, + 0.02009851112961769, + 0.03867226839065552, + 0.03220977261662483, + 0.0254293791949749, + 0.01819620281457901, + 0.014942111447453499, + 0.04590510576963425, + 0.01620924472808838, + 0.014422990381717682, + 0.025626398622989655, + 0.020871488377451897, + 0.010246079415082932, + 0.012516630813479424, + 0.009753834456205368, + 0.015350757166743279, + 0.01833566278219223, + 0.008141562342643738, + 0.012216860428452492, + 0.019237995147705078, + 0.006735463626682758, + 0.009526116773486137, + 0.009381544776260853, + 0.012287400662899017, + 0.021979577839374542, + 0.015420917421579361, + 0.014380831271409988, + 0.012246882542967796, + 0.00984797440469265, + 0.00866714958101511, + 0.007574118208140135, + 0.012143558822572231, + 0.04395202547311783, + 0.059360306710004807, + 0.050506867468357086, + 0.020245328545570374, + 0.026054037734866142, + 0.01385558769106865, + 0.015528945252299309, + 0.012049123644828796, + 0.005466517526656389, + 0.0039079091511666775, + 0.0048934947699308395, + 0.0038849255070090294, + 0.005588199011981487, + 0.004880235530436039, + 0.005314330104738474, + 0.005670558195561171, + 0.00741916336119175, + 0.007638932205736637, + 0.006529720965772867, + 0.00461886590346694, + 0.005321336444467306, + 0.006216463167220354, + 0.005224181804805994, + 0.004720039665699005, + 0.009326529689133167, + 0.0035182475112378597, + 0.006416473537683487, + 0.004018721636384726, + 0.002510788617655635, + 0.005612424109131098, + 0.002598160644993186, + 0.002801380818709731, + 0.00967989768832922, + 0.00875051598995924, + 0.0054553779773414135, + 0.005901137832552195, + 0.004623884800821543, + 0.006924504414200783, + 0.0036858059465885162, + 0.0032707094214856625, + 0.0051292432472109795, + 0.0029305461794137955, + 0.0036051934584975243, + 0.003633089829236269, + 0.0031201327219605446, + 0.004244180396199226, + 0.0037035916466265917, + 0.006595723330974579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.04870939254760742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013469918631017208 + ], + "opposite_logprob_tokens": [ + -6.626347064971924 + ], + "logprob_reward": -0.0013469918631017208, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013469918631017208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20857717096805573, + 0.1360408216714859, + 0.15689720213413239, + 0.052165739238262177, + 0.1336480677127838, + 0.05423556640744209, + 0.07100243866443634, + 0.08705157041549683, + 0.06523840874433517, + 0.035142965614795685 + ], + "credit_map": [ + 0.0340687595307827, + 0.0197371207177639, + 0.04925090819597244, + 0.03223641961812973, + 0.014882182702422142, + 0.015868404880166054, + 0.013597693294286728, + 0.028935687616467476, + 0.023677939549088478, + 0.014496074058115482, + 0.022215018048882484, + 0.026502829045057297, + 0.010050345212221146, + 0.015850000083446503, + 0.009596855379641056, + 0.013651754707098007, + 0.032511156052351, + 0.013376321643590927, + 0.02325311489403248, + 0.036179497838020325, + 0.007913913577795029, + 0.011414505541324615, + 0.013480930589139462, + 0.01876778155565262, + 0.00945638120174408, + 0.006436069030314684, + 0.008651579730212688, + 0.00683610001578927, + 0.0074457027949392796, + 0.004083261359483004, + 0.005112253129482269, + 0.0041443961672484875, + 0.026101987808942795, + 0.021926606073975563, + 0.028490962460637093, + 0.009782485663890839, + 0.010815967805683613, + 0.013016058132052422, + 0.013147745281457901, + 0.010366260074079037, + 0.010955938138067722, + 0.004440042190253735, + 0.007393215782940388, + 0.004409284796565771, + 0.003454433521255851, + 0.01082647405564785, + 0.00566586758941412, + 0.007090308703482151, + 0.013787389732897282, + 0.008596446365118027, + 0.008745333179831505, + 0.00743363332003355, + 0.0054573058150708675, + 0.0070371320471167564, + 0.00888603925704956, + 0.011059156619012356, + 0.023325176909565926, + 0.007410863880068064, + 0.01629321463406086, + 0.01001370046287775, + 0.004749119281768799, + 0.01046307198703289, + 0.009261700324714184, + 0.005534719675779343, + 0.010700813494622707, + 0.007327266503125429, + 0.006869177334010601, + 0.005750242155045271, + 0.006739550735801458, + 0.010360728949308395, + 0.008445126935839653, + 0.00904550589621067, + 0.005304568912833929, + 0.004233031067997217, + 0.0040760524570941925, + 0.004422858357429504, + 0.003765973960980773, + 0.0036876278463751078, + 0.0046422844752669334, + 0.00501056807115674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0013469918631017208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620557926595211 + ], + "opposite_logprob_tokens": [ + -2.812055826187134 + ], + "logprob_reward": -0.0620557926595211, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620557926595211, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23834730684757233, + 0.12536177039146423, + 0.0959823951125145, + 0.045028917491436005, + 0.16015662252902985, + 0.11601516604423523, + 0.06962653249502182, + 0.09931860864162445, + 0.02767038159072399, + 0.022492287680506706 + ], + "credit_map": [ + 0.037627607583999634, + 0.0378231480717659, + 0.03388393297791481, + 0.03291398286819458, + 0.030249886214733124, + 0.021463191136717796, + 0.017388809472322464, + 0.02699674665927887, + 0.020126016810536385, + 0.014750740490853786, + 0.02237989753484726, + 0.017931818962097168, + 0.010480339638888836, + 0.014702118933200836, + 0.010619520209729671, + 0.014371330849826336, + 0.021758027374744415, + 0.006951771676540375, + 0.011378942057490349, + 0.02243298664689064, + 0.006179160438477993, + 0.007131671998649836, + 0.009082803502678871, + 0.011067024432122707, + 0.009182288311421871, + 0.005064145661890507, + 0.006320893298834562, + 0.006443154998123646, + 0.004732639994472265, + 0.0042984504252672195, + 0.005005784798413515, + 0.003981562796980143, + 0.02824980951845646, + 0.034757599234580994, + 0.03241362422704697, + 0.013456416316330433, + 0.015752699226140976, + 0.01310973521322012, + 0.010999825783073902, + 0.011416909284889698, + 0.020172249525785446, + 0.010317654348909855, + 0.014491554349660873, + 0.00948413461446762, + 0.010795620270073414, + 0.019890328869223595, + 0.012344085611402988, + 0.018519539386034012, + 0.013078059069812298, + 0.011506696231663227, + 0.009417745284736156, + 0.009662463329732418, + 0.006833169609308243, + 0.005818611942231655, + 0.007978830486536026, + 0.005330952815711498, + 0.02427433244884014, + 0.008776375092566013, + 0.017266716808080673, + 0.013032503426074982, + 0.006902651395648718, + 0.014025451615452766, + 0.0074615562334656715, + 0.007579022087156773, + 0.004266120493412018, + 0.0030086019542068243, + 0.003446968039497733, + 0.003200546372681856, + 0.0027024466544389725, + 0.003520272672176361, + 0.004947250708937645, + 0.0025781739968806505, + 0.0030591462273150682, + 0.00231550307944417, + 0.002704842947423458, + 0.002942431950941682, + 0.002015375532209873, + 0.002636135322973132, + 0.0026267694775015116, + 0.0041920836083590984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0620557926595211 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032757371664047 + ], + "opposite_logprob_tokens": [ + -2.350327491760254 + ], + "logprob_reward": -0.10032757371664047, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032757371664047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17112676799297333, + 0.10683701932430267, + 0.17492491006851196, + 0.05822887271642685, + 0.259766161441803, + 0.034582652151584625, + 0.02089942991733551, + 0.11419946700334549, + 0.040384046733379364, + 0.019050708040595055 + ], + "credit_map": [ + 0.02781173773109913, + 0.015944359824061394, + 0.0319613516330719, + 0.023011235520243645, + 0.014684460125863552, + 0.017955061048269272, + 0.011163082905113697, + 0.0285954549908638, + 0.024896983057260513, + 0.009799959138035774, + 0.01726844720542431, + 0.019585542380809784, + 0.006462578661739826, + 0.00949961505830288, + 0.006125577259808779, + 0.013198302127420902, + 0.037337660789489746, + 0.016699301078915596, + 0.025762997567653656, + 0.03270801901817322, + 0.013378271833062172, + 0.014976000413298607, + 0.014248406514525414, + 0.019814252853393555, + 0.00915572326630354, + 0.013952291570603848, + 0.0075433081947267056, + 0.006118236109614372, + 0.007050283718854189, + 0.004790556617081165, + 0.005416303873062134, + 0.004202167969197035, + 0.04351923242211342, + 0.05925178900361061, + 0.05081057921051979, + 0.022747499868273735, + 0.026233941316604614, + 0.020581364631652832, + 0.01889493130147457, + 0.01772681623697281, + 0.008349050767719746, + 0.0034725405275821686, + 0.005605601705610752, + 0.004292062949389219, + 0.0025713799986988306, + 0.0025130314752459526, + 0.004872800782322884, + 0.002906183945015073, + 0.004367194604128599, + 0.0030977819114923477, + 0.002911650575697422, + 0.002151111140847206, + 0.002181828487664461, + 0.0019360866863280535, + 0.002210685983300209, + 0.0020430870354175568, + 0.03289531171321869, + 0.009865347295999527, + 0.0234227292239666, + 0.010842947289347649, + 0.006712694186717272, + 0.015617690049111843, + 0.006529836915433407, + 0.008312899619340897, + 0.006610519252717495, + 0.004605812951922417, + 0.004001228604465723, + 0.003532880451530218, + 0.003449537791311741, + 0.0074824001640081406, + 0.005270699504762888, + 0.005430963821709156, + 0.0033557438291609287, + 0.0018522629979997873, + 0.0022302744910120964, + 0.002743131248280406, + 0.0019631602335721254, + 0.002196525689214468, + 0.0024348730221390724, + 0.002274734666571021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.10032757371664047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot captures the young hero Nezha plummeting through a swirling, inky black cloud. His signature twin, jet-black buns, bound tightly by a red headband, tremble slightly in the swift currents of air. His face is defined by a furrowed brow and a resolute jawline. His lips are pursed, and his eyes burn with an unyielding fire. He rides upon a blazing Wind Fire Wheel, emitting blazing orange-red flames and visible swirling currents, the only constant source of warm light in the gloomy sky. The famous red Hun Tian Ling, like a living, shimmering silk dragon, violently twirls, coils, and stretches behind him, its dancing path tracing a flowing light trail through the air. The backdrop is a depressing apocalyptic scene, with deep blue, nearly black clouds forming a vast vortex. Dazzling purple lightning periodically rips through the sky, each flash instantly illuminating his red armor, emblazoned with golden lotus patterns, and his resolute face, casting a sharp, shifting shadow across him. Like a flaming meteorite, his entire body streaks steadily and decisively through the sky at astonishing speed, emitting a sharp whistling sound as he hurtles towards an unknown land. The entire scene is sharply contrasted by the cool tones of the sky and the fiery warm tones of Nezha himself, creating an epic sense of doom and a tense atmosphere filled with power and defiance.", + "seed": 31235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_8503_seed_31235_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": -0.1155538794817403, + "vlm_logprob": -0.1155538794817403, + "vlm_yes_no_margin": 3.775000011920929, + "vlm_token_credit": -0.1155538794817403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5134457945823669 + ], + "opposite_logprob_tokens": [ + -1.0134458541870117 + ], + "logprob_reward": -0.5134457945823669, + "yes_no_margin_reward": 0.5000000596046448, + "token_credit_reward": -0.5134457945823669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15716885030269623, + 0.06999580562114716, + 0.09218542277812958, + 0.27566856145858765, + 0.0744093656539917, + 0.041086595505476, + 0.042637042701244354, + 0.07878674566745758, + 0.12200465798377991, + 0.046056970953941345 + ], + "credit_map": [ + 0.016213124617934227, + 0.017239481210708618, + 0.02165484055876732, + 0.021405106410384178, + 0.01574224978685379, + 0.029154526069760323, + 0.016561221331357956, + 0.01919829659163952, + 0.00680962111800909, + 0.0048711565323174, + 0.017713163048028946, + 0.00603463826701045, + 0.006190571002662182, + 0.010004620999097824, + 0.009977863170206547, + 0.008394167758524418, + 0.013075005263090134, + 0.005445652641355991, + 0.01821562834084034, + 0.007549053989350796, + 0.006993167567998171, + 0.021805472671985626, + 0.007809163071215153, + 0.011292281560599804, + 0.02747293934226036, + 0.027333049103617668, + 0.03212933987379074, + 0.04141221567988396, + 0.019495191052556038, + 0.05160301923751831, + 0.027221431955695152, + 0.04900136962532997, + 0.0074084121733903885, + 0.009041919372975826, + 0.010475368238985538, + 0.009622082114219666, + 0.009230547584593296, + 0.009810182265937328, + 0.007612459361553192, + 0.01120839174836874, + 0.005863056518137455, + 0.005457918159663677, + 0.005971595644950867, + 0.0034694254864007235, + 0.005008704494684935, + 0.005811881739646196, + 0.0038075351621955633, + 0.005696471780538559, + 0.004750700667500496, + 0.00956538412719965, + 0.004542926326394081, + 0.004902356769889593, + 0.005675090942531824, + 0.0037432387471199036, + 0.004911327734589577, + 0.004546012729406357, + 0.007176295854151249, + 0.009100397117435932, + 0.010666092857718468, + 0.009799941442906857, + 0.014812334440648556, + 0.007958203554153442, + 0.007924136705696583, + 0.01134934090077877, + 0.021872419863939285, + 0.012238739989697933, + 0.008917946368455887, + 0.03300587832927704, + 0.008454733528196812, + 0.01116142701357603, + 0.011318615637719631, + 0.015034887939691544, + 0.007621808908879757, + 0.003936682362109423, + 0.009191280230879784, + 0.004578071646392345, + 0.0044782753102481365, + 0.0034279630053788424, + 0.008879792876541615, + 0.003943095915019512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.5134457945823669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012582560069859028 + ], + "opposite_logprob_tokens": [ + -4.387582778930664 + ], + "logprob_reward": -0.012582560069859028, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012582560069859028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05848565697669983, + 0.030461523681879044, + 0.04445607215166092, + 0.07958541810512543, + 0.03704766929149628, + 0.03399105742573738, + 0.05484858527779579, + 0.04839393496513367, + 0.4990953505039215, + 0.11363473534584045 + ], + "credit_map": [ + 0.006923587992787361, + 0.006471066270023584, + 0.008237926289439201, + 0.007406326476484537, + 0.005941685754805803, + 0.008111043833196163, + 0.005334135610610247, + 0.01005988847464323, + 0.0025764659512788057, + 0.0020563346333801746, + 0.008250855840742588, + 0.0022575485054403543, + 0.00218318123370409, + 0.006873213220387697, + 0.003259187564253807, + 0.00300473696552217, + 0.005252042319625616, + 0.003111399943009019, + 0.009807576425373554, + 0.00344442599453032, + 0.0037616293411701918, + 0.008964642882347107, + 0.003083153860643506, + 0.007031202781945467, + 0.008722472004592419, + 0.009466872550547123, + 0.007827258668839931, + 0.01624968647956848, + 0.006751677952706814, + 0.009751356206834316, + 0.009112857282161713, + 0.01170323695987463, + 0.004732692614197731, + 0.004716843366622925, + 0.004181588068604469, + 0.004426669795066118, + 0.004848693963140249, + 0.004705819301307201, + 0.0036370372399687767, + 0.005798324942588806, + 0.0038934589829295874, + 0.00620998814702034, + 0.0034265622962266207, + 0.004644719418138266, + 0.004636496305465698, + 0.0038910391740500927, + 0.004534570034593344, + 0.0027542237658053637, + 0.005302215460687876, + 0.011483057402074337, + 0.006001822184771299, + 0.005807614419609308, + 0.0073482743464410305, + 0.005883460398763418, + 0.0065589905716478825, + 0.006463152822107077, + 0.004436612129211426, + 0.008195293135941029, + 0.005062607582658529, + 0.00581919401884079, + 0.00723467068746686, + 0.0053282976150512695, + 0.004273910541087389, + 0.008043346926569939, + 0.0979481115937233, + 0.06111341342329979, + 0.03614860400557518, + 0.11414836347103119, + 0.031406812369823456, + 0.046579841524362564, + 0.04907190799713135, + 0.06267828494310379, + 0.013369271531701088, + 0.011913234367966652, + 0.018298456445336342, + 0.013085927814245224, + 0.010748492553830147, + 0.008500324562191963, + 0.02855750545859337, + 0.009161515161395073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.012582560069859028 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013510395074263215 + ], + "opposite_logprob_tokens": [ + -6.6263508796691895 + ], + "logprob_reward": -0.0013510395074263215, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013510395074263215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11656118184328079, + 0.05212928727269173, + 0.13742120563983917, + 0.12928524613380432, + 0.08264075219631195, + 0.045774489641189575, + 0.04398630931973457, + 0.06100470945239067, + 0.11388517916202545, + 0.21731163561344147 + ], + "credit_map": [ + 0.01538052223622799, + 0.012643471360206604, + 0.013812322169542313, + 0.01593366451561451, + 0.011326094157993793, + 0.02045409567654133, + 0.014498313888907433, + 0.012512698769569397, + 0.004285970702767372, + 0.004990261513739824, + 0.010858070105314255, + 0.005575861316174269, + 0.004832654260098934, + 0.007875009439885616, + 0.006531652994453907, + 0.007179807405918837, + 0.018834037706255913, + 0.013789596036076546, + 0.0290392879396677, + 0.0104256272315979, + 0.011765425093472004, + 0.023764025419950485, + 0.01064472459256649, + 0.019158480688929558, + 0.014574892818927765, + 0.016765404492616653, + 0.01293010264635086, + 0.016700387001037598, + 0.009058813564479351, + 0.02300160750746727, + 0.01378593873232603, + 0.02246810495853424, + 0.006719190627336502, + 0.014344302006065845, + 0.006985219195485115, + 0.011653083376586437, + 0.012885899282991886, + 0.011537183076143265, + 0.008618243969976902, + 0.009897629730403423, + 0.006334437057375908, + 0.006424773018807173, + 0.004359451588243246, + 0.010295183397829533, + 0.004083076026290655, + 0.0044695353135466576, + 0.0055434913374483585, + 0.004264542367309332, + 0.004831430036574602, + 0.009292706847190857, + 0.004837200976908207, + 0.00403417507186532, + 0.004986622836440802, + 0.005403298418968916, + 0.004282196518033743, + 0.006318682339042425, + 0.006424865685403347, + 0.006953518372029066, + 0.008415094576776028, + 0.008106549270451069, + 0.01045815460383892, + 0.0064217024482786655, + 0.00621929531916976, + 0.008005533367395401, + 0.0197586752474308, + 0.011465011164546013, + 0.010078699328005314, + 0.02527964487671852, + 0.00881847646087408, + 0.012454953044652939, + 0.010103404521942139, + 0.015926308929920197, + 0.026167528703808784, + 0.018617691472172737, + 0.03826166316866875, + 0.022879498079419136, + 0.01888135075569153, + 0.013039235025644302, + 0.06347803771495819, + 0.01598665490746498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.0013510395074263215 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020582424476742744 + ], + "opposite_logprob_tokens": [ + -3.895582437515259 + ], + "logprob_reward": -0.020582424476742744, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020582424476742744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07369531691074371, + 0.05352535471320152, + 0.08403122425079346, + 0.09416712820529938, + 0.030789708718657494, + 0.026741383597254753, + 0.03016071766614914, + 0.033855851739645004, + 0.1820371150970459, + 0.39099612832069397 + ], + "credit_map": [ + 0.0074373711831867695, + 0.009031142108142376, + 0.00850795116275549, + 0.008924063295125961, + 0.007744364440441132, + 0.009864225052297115, + 0.009905771352350712, + 0.012280427850782871, + 0.004163634963333607, + 0.0035836438182741404, + 0.012807579711079597, + 0.0047607808373868465, + 0.004351623356342316, + 0.010667271912097931, + 0.007154082879424095, + 0.006036742124706507, + 0.010524740442633629, + 0.006043149624019861, + 0.018906015902757645, + 0.006182024721056223, + 0.006727232597768307, + 0.017677681520581245, + 0.006261080969125032, + 0.011709300801157951, + 0.011504042893648148, + 0.00799508672207594, + 0.007750129327178001, + 0.020336642861366272, + 0.00558817433193326, + 0.013565214350819588, + 0.01086380984634161, + 0.01656402461230755, + 0.003839001525193453, + 0.005297623109072447, + 0.0027314708568155766, + 0.0035777443554252386, + 0.004355099517852068, + 0.0037766911555081606, + 0.0023898654617369175, + 0.004822212737053633, + 0.003411069978028536, + 0.003584235440939665, + 0.0024107960052788258, + 0.004111751448363066, + 0.003910517320036888, + 0.003225580556318164, + 0.0025716321542859077, + 0.003515800228342414, + 0.002763093216344714, + 0.00460401177406311, + 0.002865393180400133, + 0.0029425732791423798, + 0.005047754850238562, + 0.004211885388940573, + 0.0031231625471264124, + 0.004602842964231968, + 0.003294466296210885, + 0.0056498367339372635, + 0.0035346182994544506, + 0.002969177905470133, + 0.005081880837678909, + 0.005757920444011688, + 0.002230579499155283, + 0.005337372422218323, + 0.04044558107852936, + 0.020559338852763176, + 0.01412112545222044, + 0.037746768444776535, + 0.009370296262204647, + 0.018324458971619606, + 0.016000976786017418, + 0.02546856366097927, + 0.03745041787624359, + 0.041301194578409195, + 0.05361602082848549, + 0.04956643283367157, + 0.03357458859682083, + 0.022192547097802162, + 0.12707242369651794, + 0.026222556829452515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.020582424476742744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029807578772306442 + ], + "opposite_logprob_tokens": [ + -3.5298075675964355 + ], + "logprob_reward": -0.029807578772306442, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029807578772306442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11782506853342056, + 0.07479041069746017, + 0.14087003469467163, + 0.07870624959468842, + 0.08921507745981216, + 0.032157253473997116, + 0.055151861160993576, + 0.07743344455957413, + 0.1510087549686432, + 0.18284183740615845 + ], + "credit_map": [ + 0.011127324774861336, + 0.012207424268126488, + 0.016863929107785225, + 0.0167926624417305, + 0.010009362362325191, + 0.01720689982175827, + 0.014993170276284218, + 0.01862429827451706, + 0.008734908886253834, + 0.0063601271249353886, + 0.012184255756437778, + 0.007005241699516773, + 0.007153376471251249, + 0.010601076297461987, + 0.010428646579384804, + 0.012322772294282913, + 0.018118618056178093, + 0.007824106141924858, + 0.02656722255051136, + 0.01416933536529541, + 0.010730654001235962, + 0.03320466727018356, + 0.011916243471205235, + 0.01833919808268547, + 0.007374389562755823, + 0.008827847428619862, + 0.009089105762541294, + 0.011933998204767704, + 0.007686660625040531, + 0.010536796413362026, + 0.008792577311396599, + 0.014464880339801311, + 0.007416641805320978, + 0.01258774846792221, + 0.006875708699226379, + 0.018154900521039963, + 0.011272275820374489, + 0.012897289358079433, + 0.010489148087799549, + 0.009521367028355598, + 0.004118418321013451, + 0.004279745742678642, + 0.0022356826812028885, + 0.005114695522934198, + 0.004456492606550455, + 0.004178709350526333, + 0.0041115800850093365, + 0.0036619282327592373, + 0.004805519711226225, + 0.015572184696793556, + 0.005565895698964596, + 0.005084230564534664, + 0.00795030314475298, + 0.004767483565956354, + 0.0064250570721924305, + 0.004981185309588909, + 0.01142759621143341, + 0.011208425275981426, + 0.005751911550760269, + 0.008046071045100689, + 0.013660435564815998, + 0.008176697418093681, + 0.006410371977835894, + 0.012751936912536621, + 0.02397923730313778, + 0.02296518161892891, + 0.014415125362575054, + 0.03418418765068054, + 0.012092288583517075, + 0.012342341244220734, + 0.015039021149277687, + 0.015991363674402237, + 0.023786203935742378, + 0.011645661666989326, + 0.030086621642112732, + 0.016557898372411728, + 0.015442587435245514, + 0.012689698487520218, + 0.055401019752025604, + 0.0172321405261755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.029807578772306442 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot captures the young hero Nezha plummeting through a swirling, inky black cloud. His signature twin, jet-black buns, bound tightly by a red headband, tremble slightly in the swift currents of air. His face is defined by a furrowed brow and a resolute jawline. His lips are pursed, and his eyes burn with an unyielding fire. He rides upon a blazing Wind Fire Wheel, emitting blazing orange-red flames and visible swirling currents, the only constant source of warm light in the gloomy sky. The famous red Hun Tian Ling, like a living, shimmering silk dragon, violently twirls, coils, and stretches behind him, its dancing path tracing a flowing light trail through the air. The backdrop is a depressing apocalyptic scene, with deep blue, nearly black clouds forming a vast vortex. Dazzling purple lightning periodically rips through the sky, each flash instantly illuminating his red armor, emblazoned with golden lotus patterns, and his resolute face, casting a sharp, shifting shadow across him. Like a flaming meteorite, his entire body streaks steadily and decisively through the sky at astonishing speed, emitting a sharp whistling sound as he hurtles towards an unknown land. The entire scene is sharply contrasted by the cool tones of the sky and the fiery warm tones of Nezha himself, creating an epic sense of doom and a tense atmosphere filled with power and defiance.", + "seed": 31236, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_8503_seed_31236_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": -0.04331608788343146, + "vlm_logprob": -0.04331608788343146, + "vlm_yes_no_margin": 4.3249999523162845, + "vlm_token_credit": -0.04331608788343146, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1542752981185913 + ], + "opposite_logprob_tokens": [ + -2.279275417327881 + ], + "logprob_reward": -0.1542752981185913, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1542752981185913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2561037838459015, + 0.0895238071680069, + 0.20016853511333466, + 0.09846290946006775, + 0.08371269702911377, + 0.0892779529094696, + 0.07188598066568375, + 0.029159342870116234, + 0.024946054443717003, + 0.056758951395750046 + ], + "credit_map": [ + 0.058813683688640594, + 0.05597692355513573, + 0.04691056162118912, + 0.01609567366540432, + 0.020026208832859993, + 0.026733748614788055, + 0.012880118563771248, + 0.01866685412824154, + 0.007030821871012449, + 0.0169157013297081, + 0.012023004703223705, + 0.010550254955887794, + 0.009034874849021435, + 0.010642588138580322, + 0.007578262127935886, + 0.01574830524623394, + 0.02365422435104847, + 0.024076802656054497, + 0.03911951556801796, + 0.03258330747485161, + 0.015393280424177647, + 0.02822737954556942, + 0.015788821503520012, + 0.02132520079612732, + 0.010313351638615131, + 0.013590180315077305, + 0.013511893339455128, + 0.00804039929062128, + 0.008691594004631042, + 0.012644375674426556, + 0.008750632405281067, + 0.022920478135347366, + 0.011978955939412117, + 0.0064811985939741135, + 0.013418372720479965, + 0.011973385699093342, + 0.007633363828063011, + 0.010718275792896748, + 0.008339294232428074, + 0.013169853948056698, + 0.018543820828199387, + 0.01019346248358488, + 0.009591745212674141, + 0.006483905483037233, + 0.009804501198232174, + 0.013777878135442734, + 0.00922484789043665, + 0.011657792143523693, + 0.010114893317222595, + 0.007026066537946463, + 0.006437328178435564, + 0.007729160133749247, + 0.019152624532580376, + 0.006101940758526325, + 0.005832152906805277, + 0.009491817094385624, + 0.0037504013162106276, + 0.003342034062370658, + 0.003246401669457555, + 0.003095622407272458, + 0.005390798673033714, + 0.0026717823930084705, + 0.0032779043540358543, + 0.004384398926049471, + 0.0031184249091893435, + 0.0030955737456679344, + 0.002773377113044262, + 0.00255625881254673, + 0.004362230654805899, + 0.002947294618934393, + 0.003240083809942007, + 0.0028528105467557907, + 0.00561520317569375, + 0.006131987553089857, + 0.007715034298598766, + 0.009899597615003586, + 0.005556415766477585, + 0.0037044340278953314, + 0.011233851313591003, + 0.006902423687279224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.1542752981185913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018229680135846138 + ], + "opposite_logprob_tokens": [ + -4.0182294845581055 + ], + "logprob_reward": -0.018229680135846138, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018229680135846138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6607766151428223, + 0.040726590901613235, + 0.024273082613945007, + 0.025723641738295555, + 0.043133169412612915, + 0.03162643685936928, + 0.04585663974285126, + 0.028879055753350258, + 0.02000335045158863, + 0.07900141179561615 + ], + "credit_map": [ + 0.04791206493973732, + 0.2779098451137543, + 0.19673532247543335, + 0.014309250749647617, + 0.019785957410931587, + 0.021114202216267586, + 0.011250135488808155, + 0.07175981998443604, + 0.004523000214248896, + 0.005826220847666264, + 0.006102471146732569, + 0.004200646188110113, + 0.004379482474178076, + 0.004832552745938301, + 0.0034065558575093746, + 0.007455663755536079, + 0.002498926594853401, + 0.0026433018501847982, + 0.004473142325878143, + 0.0032966453582048416, + 0.002563179237768054, + 0.0030127193313091993, + 0.0031054893042892218, + 0.0026796774473041296, + 0.0025816624984145164, + 0.004195602610707283, + 0.003286461578682065, + 0.0024706136900931597, + 0.0030152620747685432, + 0.0029018481727689505, + 0.002592018572613597, + 0.00468017254024744, + 0.007630687206983566, + 0.004892874974757433, + 0.00729160662740469, + 0.0036530576180666685, + 0.003038141643628478, + 0.006593543570488691, + 0.004119790159165859, + 0.005913469009101391, + 0.0035349696408957243, + 0.003508712165057659, + 0.0032935976050794125, + 0.0030640214681625366, + 0.005707062780857086, + 0.005153736565262079, + 0.003608501283451915, + 0.0037558353506028652, + 0.006267227698117495, + 0.00404419144615531, + 0.004675819538533688, + 0.004769825376570225, + 0.00971144624054432, + 0.006198691204190254, + 0.003961754497140646, + 0.0062276823446154594, + 0.004277661442756653, + 0.004297146573662758, + 0.002995031652972102, + 0.002269776538014412, + 0.006659282371401787, + 0.0028010495007038116, + 0.0026101020630449057, + 0.0029690053779631853, + 0.0022141484078019857, + 0.0022484755609184504, + 0.0019263193244114518, + 0.0028074360452592373, + 0.002135905437171459, + 0.003256171243265271, + 0.002824996365234256, + 0.002589897718280554, + 0.00927676074206829, + 0.0104953208938241, + 0.009168566204607487, + 0.010321531444787979, + 0.008444567210972309, + 0.009302644990384579, + 0.011045781895518303, + 0.010946234688162804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.018229680135846138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005723983631469309 + ], + "opposite_logprob_tokens": [ + -7.500572204589844 + ], + "logprob_reward": -0.0005723983631469309, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005723983631469309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.691421389579773, + 0.03595704585313797, + 0.03699755668640137, + 0.020445145666599274, + 0.04068616405129433, + 0.027917491272091866, + 0.037934694439172745, + 0.025781894102692604, + 0.023710357025265694, + 0.059148166328668594 + ], + "credit_map": [ + 0.044387273490428925, + 0.29283204674720764, + 0.20745110511779785, + 0.015299536287784576, + 0.021640852093696594, + 0.02088378556072712, + 0.011910846456885338, + 0.07701599597930908, + 0.00361440796405077, + 0.005544291343539953, + 0.005225608125329018, + 0.003954854793846607, + 0.005249392706900835, + 0.004630992654711008, + 0.0030641360208392143, + 0.004673367831856012, + 0.00490765692666173, + 0.0039012720808386803, + 0.004927738569676876, + 0.005931686144322157, + 0.004086112137883902, + 0.005318031180649996, + 0.0034630594309419394, + 0.004462003242224455, + 0.0022251796908676624, + 0.0025321515277028084, + 0.0035984378773719072, + 0.0022520828060805798, + 0.002170498948544264, + 0.0025769199710339308, + 0.0015366183361038566, + 0.003553258255124092, + 0.006734682247042656, + 0.005127296783030033, + 0.006109011359512806, + 0.003969680983573198, + 0.003035120666027069, + 0.0057298364117741585, + 0.003896303242072463, + 0.0060842325910925865, + 0.00349913677200675, + 0.0027599921450018883, + 0.0024442505091428757, + 0.0030740993097424507, + 0.005143573973327875, + 0.004596319980919361, + 0.002897523809224367, + 0.0035025980323553085, + 0.0057785832323133945, + 0.003421894507482648, + 0.0024754798505455256, + 0.0032814268488436937, + 0.00999931525439024, + 0.003682636423036456, + 0.00367946270853281, + 0.005615898407995701, + 0.0033556981943547726, + 0.002843053312972188, + 0.0040686242282390594, + 0.0019419352756813169, + 0.0043636709451675415, + 0.002479714574292302, + 0.002047196961939335, + 0.004682001192122698, + 0.0022546646650880575, + 0.0033474257215857506, + 0.002162241842597723, + 0.0028925989754498005, + 0.0036389220040291548, + 0.00348176178522408, + 0.0029856241308152676, + 0.0029471185989677906, + 0.00598911615088582, + 0.007695815525949001, + 0.0077618141658604145, + 0.007597580086439848, + 0.007411697879433632, + 0.00807960145175457, + 0.005998415872454643, + 0.00861413311213255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0005723983631469309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371020406484604 + ], + "opposite_logprob_tokens": [ + -3.408710241317749 + ], + "logprob_reward": -0.03371020406484604, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371020406484604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7104663848876953, + 0.028346113860607147, + 0.025361476466059685, + 0.03476950526237488, + 0.03682822734117508, + 0.029217800125479698, + 0.06218015030026436, + 0.01753481663763523, + 0.02381913736462593, + 0.03147643432021141 + ], + "credit_map": [ + 0.05244183912873268, + 0.2964005172252655, + 0.21065205335617065, + 0.01700824499130249, + 0.02058337815105915, + 0.02341674268245697, + 0.013103699311614037, + 0.07685992121696472, + 0.003103727474808693, + 0.0047807455994188786, + 0.004271642304956913, + 0.0028517532628029585, + 0.0035139592364430428, + 0.00370130967348814, + 0.00265887682326138, + 0.0034641013480722904, + 0.0031875059939920902, + 0.0027419552206993103, + 0.004094610456377268, + 0.002956212032586336, + 0.003133172169327736, + 0.0039034998044371605, + 0.0030911294743418694, + 0.0022533924784511328, + 0.004213009029626846, + 0.0054181236773729324, + 0.005181781481951475, + 0.003320359857752919, + 0.004164326470345259, + 0.003617331385612488, + 0.0038236661348491907, + 0.005030907224863768, + 0.006438783835619688, + 0.003909613937139511, + 0.005359722301363945, + 0.004447274375706911, + 0.00265280669555068, + 0.004695284646004438, + 0.003105826210230589, + 0.006218914873898029, + 0.00350409047678113, + 0.003686655545607209, + 0.003342693205922842, + 0.0032586497254669666, + 0.004927818663418293, + 0.004370733629912138, + 0.0024912762455642223, + 0.0036358849611133337, + 0.008874633349478245, + 0.005827667191624641, + 0.0032333997078239918, + 0.006022435612976551, + 0.017839159816503525, + 0.006061785388737917, + 0.005286652594804764, + 0.009034418500959873, + 0.0027769117150455713, + 0.0020157143007963896, + 0.0016575694316998124, + 0.0013287117471918464, + 0.004409991204738617, + 0.0017256044084206223, + 0.001505799125880003, + 0.002114514820277691, + 0.003030328080058098, + 0.0022530686110258102, + 0.002794156316667795, + 0.0038251541554927826, + 0.0025245745200663805, + 0.0029638041742146015, + 0.003581267548725009, + 0.002846783958375454, + 0.005045605823397636, + 0.004041635897010565, + 0.002671495545655489, + 0.004342399071902037, + 0.0029333122074604034, + 0.004223974421620369, + 0.0039270236156880856, + 0.004290987271815538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.03371020406484604 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792858734726906 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792858734726906, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792858734726906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6266863346099854, + 0.04776628687977791, + 0.034970469772815704, + 0.02204221859574318, + 0.07082680612802505, + 0.030066844075918198, + 0.053341273218393326, + 0.01739409752190113, + 0.02847018837928772, + 0.06843548268079758 + ], + "credit_map": [ + 0.04712025821208954, + 0.2603503465652466, + 0.18356716632843018, + 0.01567061059176922, + 0.019215278327465057, + 0.020734569057822227, + 0.011619355529546738, + 0.0684087797999382, + 0.005040821619331837, + 0.007968463934957981, + 0.0074068233370780945, + 0.005713296122848988, + 0.004018404986709356, + 0.007108015473932028, + 0.003486579516902566, + 0.007023891434073448, + 0.004579138942062855, + 0.004292604047805071, + 0.004747565370053053, + 0.004114373121410608, + 0.0041163768619298935, + 0.005929715931415558, + 0.00368093722499907, + 0.003509764326736331, + 0.0026221738662570715, + 0.002999591873958707, + 0.002989170141518116, + 0.002691510831937194, + 0.0027525753248482943, + 0.0022574542090296745, + 0.0031942485366016626, + 0.0025354980025440454, + 0.01340030413120985, + 0.006354207172989845, + 0.012412303127348423, + 0.006135190837085247, + 0.004830411169677973, + 0.010111122392117977, + 0.006880870554596186, + 0.0107024060562253, + 0.0040768892504274845, + 0.004035722464323044, + 0.002662781858816743, + 0.0032103846315294504, + 0.004435126204043627, + 0.004873456433415413, + 0.002769193146377802, + 0.004003292880952358, + 0.008297385647892952, + 0.005099147092550993, + 0.0027933705132454634, + 0.003930575214326382, + 0.015604489482939243, + 0.005024113692343235, + 0.0039803278632462025, + 0.008611868135631084, + 0.002270217053592205, + 0.0016773785464465618, + 0.0012965926434844732, + 0.0018337011570110917, + 0.0036437958478927612, + 0.0018957379506900907, + 0.0018713375320658088, + 0.0029053394682705402, + 0.0024721946101635695, + 0.0035314266569912434, + 0.002374030416831374, + 0.004390308633446693, + 0.0036378735676407814, + 0.0041418978944420815, + 0.004242601804435253, + 0.0036798580549657345, + 0.008093367330729961, + 0.008588897995650768, + 0.009174440056085587, + 0.010280035436153412, + 0.007018552161753178, + 0.009241068735718727, + 0.006496460177004337, + 0.009542673826217651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.009792858734726906 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, overhead tracking shot captures the young hero Nezha plummeting through a swirling, inky black cloud. His signature twin, jet-black buns, bound tightly by a red headband, tremble slightly in the swift currents of air. His face is defined by a furrowed brow and a resolute jawline. His lips are pursed, and his eyes burn with an unyielding fire. He rides upon a blazing Wind Fire Wheel, emitting blazing orange-red flames and visible swirling currents, the only constant source of warm light in the gloomy sky. The famous red Hun Tian Ling, like a living, shimmering silk dragon, violently twirls, coils, and stretches behind him, its dancing path tracing a flowing light trail through the air. The backdrop is a depressing apocalyptic scene, with deep blue, nearly black clouds forming a vast vortex. Dazzling purple lightning periodically rips through the sky, each flash instantly illuminating his red armor, emblazoned with golden lotus patterns, and his resolute face, casting a sharp, shifting shadow across him. Like a flaming meteorite, his entire body streaks steadily and decisively through the sky at astonishing speed, emitting a sharp whistling sound as he hurtles towards an unknown land. The entire scene is sharply contrasted by the cool tones of the sky and the fiery warm tones of Nezha himself, creating an epic sense of doom and a tense atmosphere filled with power and defiance.", + "seed": 31237, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_15/index_8503_seed_31237_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": -0.10356569973519072, + "vlm_logprob": -0.10356569973519072, + "vlm_yes_no_margin": 3.874999988079071, + "vlm_token_credit": -0.10356569973519072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4286354184150696 + ], + "opposite_logprob_tokens": [ + -1.1786353588104248 + ], + "logprob_reward": -0.4286354184150696, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.4286354184150696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1816190481185913, + 0.11200600862503052, + 0.0494966059923172, + 0.06530801951885223, + 0.1021212488412857, + 0.058110009878873825, + 0.11016148328781128, + 0.1670176237821579, + 0.050184957683086395, + 0.10397506505250931 + ], + "credit_map": [ + 0.0348103828728199, + 0.016519924625754356, + 0.01883913017809391, + 0.013019440695643425, + 0.019415350630879402, + 0.026677070185542107, + 0.020131317898631096, + 0.03220644220709801, + 0.02033533900976181, + 0.006417478434741497, + 0.018162326887249947, + 0.005675449967384338, + 0.005992130842059851, + 0.021048102527856827, + 0.008988477289676666, + 0.02538670226931572, + 0.007094076834619045, + 0.007372946944087744, + 0.00518010975793004, + 0.009357122704386711, + 0.004748891107738018, + 0.004522872623056173, + 0.004663194064050913, + 0.006557391490787268, + 0.005499233957380056, + 0.006813657470047474, + 0.009142317809164524, + 0.01266007125377655, + 0.004485130310058594, + 0.006663383450359106, + 0.007640770170837641, + 0.012403454631567001, + 0.0077165174297988415, + 0.011213401332497597, + 0.01922323927283287, + 0.014671148732304573, + 0.015521086752414703, + 0.01088308822363615, + 0.0145280035212636, + 0.008364755660295486, + 0.0065719387494027615, + 0.006322053261101246, + 0.00674417894333601, + 0.008067842572927475, + 0.012057519517838955, + 0.006497131194919348, + 0.007328114006668329, + 0.004521232098340988, + 0.014913852326571941, + 0.012242184951901436, + 0.018190303817391396, + 0.0161187332123518, + 0.01202664989978075, + 0.011828639544546604, + 0.011374623514711857, + 0.013466489501297474, + 0.024752695113420486, + 0.02565905638039112, + 0.04509382322430611, + 0.020784949883818626, + 0.013275111094117165, + 0.014719889499247074, + 0.011391332373023033, + 0.011340768076479435, + 0.0064818356186151505, + 0.0076642027124762535, + 0.011909290216863155, + 0.004406913183629513, + 0.0036336674820631742, + 0.007078994531184435, + 0.0031810228247195482, + 0.005829032510519028, + 0.02388554997742176, + 0.012387876398861408, + 0.01795930787920952, + 0.008267018012702465, + 0.010438743978738785, + 0.014338644221425056, + 0.009502827189862728, + 0.00719509506598115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.4286354184150696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871052876114845 + ], + "opposite_logprob_tokens": [ + -3.048710584640503 + ], + "logprob_reward": -0.04871052876114845, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871052876114845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06620370596647263, + 0.215748593211174, + 0.021175237372517586, + 0.026568200439214706, + 0.09468799084424973, + 0.24011074006557465, + 0.16050654649734497, + 0.02525361068546772, + 0.07004685699939728, + 0.07969850301742554 + ], + "credit_map": [ + 0.007891026325523853, + 0.0068786959163844585, + 0.00892704539000988, + 0.004269556142389774, + 0.0072874585166573524, + 0.00879011768847704, + 0.00814864318817854, + 0.01401116419583559, + 0.07611031085252762, + 0.009181669913232327, + 0.015208957716822624, + 0.004672293551266193, + 0.005978853441774845, + 0.06401973962783813, + 0.008130301721394062, + 0.032446470111608505, + 0.0023765810765326023, + 0.003312972839921713, + 0.0017497255466878414, + 0.00299832527525723, + 0.0027555127162486315, + 0.0022581464145332575, + 0.002228484023362398, + 0.0034954885486513376, + 0.0023155445232987404, + 0.003293193643912673, + 0.0038809196557849646, + 0.007090822793543339, + 0.002023715991526842, + 0.002998404437676072, + 0.002764540258795023, + 0.002201058203354478, + 0.007421466056257486, + 0.010398458689451218, + 0.02053551748394966, + 0.013471221551299095, + 0.014854095876216888, + 0.010004094801843166, + 0.012524746358394623, + 0.005478390026837587, + 0.021269718185067177, + 0.021283280104398727, + 0.030197927728295326, + 0.03701937943696976, + 0.043473295867443085, + 0.02644110843539238, + 0.0362674817442894, + 0.024158544838428497, + 0.014411620795726776, + 0.011903629638254642, + 0.025673266500234604, + 0.02466500550508499, + 0.02460293285548687, + 0.021633898839354515, + 0.022200224921107292, + 0.015415966510772705, + 0.0036791195161640644, + 0.004123106598854065, + 0.006820838898420334, + 0.0020362744107842445, + 0.0023759114556014538, + 0.0022984023671597242, + 0.002256543841212988, + 0.0016634149942547083, + 0.00917820818722248, + 0.009961810894310474, + 0.0169526319950819, + 0.005950086284428835, + 0.006015603430569172, + 0.011198836378753185, + 0.005367283709347248, + 0.005422396119683981, + 0.015654442831873894, + 0.006710932590067387, + 0.013267314992845058, + 0.007523269392549992, + 0.009276074357330799, + 0.00979555957019329, + 0.009252700954675674, + 0.008218209259212017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.04871052876114845 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000825898430775851 + ], + "opposite_logprob_tokens": [ + -7.125825881958008 + ], + "logprob_reward": -0.000825898430775851, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.000825898430775851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11279567331075668, + 0.2060963362455368, + 0.05094529688358307, + 0.04829973354935646, + 0.09569546580314636, + 0.08020627498626709, + 0.08232592791318893, + 0.11175363510847092, + 0.07193002104759216, + 0.1399516612291336 + ], + "credit_map": [ + 0.014768208377063274, + 0.010067093186080456, + 0.011335832066833973, + 0.010058625601232052, + 0.012825839221477509, + 0.015585139393806458, + 0.012362193316221237, + 0.025792745873332024, + 0.026081493124365807, + 0.009291034191846848, + 0.031394489109516144, + 0.009792301803827286, + 0.007585453800857067, + 0.06045025587081909, + 0.008077074773609638, + 0.053424227982759476, + 0.004836126230657101, + 0.0060156360268592834, + 0.006418740842491388, + 0.007749765180051327, + 0.003730274038389325, + 0.0031753694638609886, + 0.005566301755607128, + 0.013453084975481033, + 0.0038761727046221495, + 0.006277857813984156, + 0.005659573711454868, + 0.0066616348922252655, + 0.005734873004257679, + 0.005606499966233969, + 0.005481744650751352, + 0.009001375176012516, + 0.008675864897668362, + 0.011759408749639988, + 0.013238904997706413, + 0.010614540427923203, + 0.01774837262928486, + 0.010534984990954399, + 0.01341888029128313, + 0.009704514406621456, + 0.01428903080523014, + 0.008893822319805622, + 0.013139883056282997, + 0.013580907136201859, + 0.009447350166738033, + 0.006414341274648905, + 0.009302016347646713, + 0.005138925742357969, + 0.008971362374722958, + 0.006803522817790508, + 0.011168869212269783, + 0.012195875868201256, + 0.008396905846893787, + 0.009362513199448586, + 0.010177601128816605, + 0.015249276533722878, + 0.016331730410456657, + 0.017407063394784927, + 0.032182104885578156, + 0.013425758108496666, + 0.008200637996196747, + 0.00860828347504139, + 0.0076858876273036, + 0.007912172004580498, + 0.010465971194207668, + 0.009316840209066868, + 0.02024903893470764, + 0.005966661497950554, + 0.005578489508479834, + 0.010621621273458004, + 0.004647465888410807, + 0.005083925556391478, + 0.03539431095123291, + 0.018775759264826775, + 0.019399788230657578, + 0.012511972337961197, + 0.01248660497367382, + 0.017436254769563675, + 0.01247840654104948, + 0.011468568816781044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.000825898430775851 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370155766606331 + ], + "opposite_logprob_tokens": [ + -3.4087016582489014 + ], + "logprob_reward": -0.03370155766606331, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370155766606331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1094847321510315, + 0.11752448976039886, + 0.032660361379384995, + 0.04104330763220787, + 0.0483163557946682, + 0.08550146967172623, + 0.07430577278137207, + 0.04797990247607231, + 0.1250438541173935, + 0.3181397616863251 + ], + "credit_map": [ + 0.01405540481209755, + 0.010293271392583847, + 0.008919995278120041, + 0.007938770577311516, + 0.010745731182396412, + 0.01674104854464531, + 0.01418977975845337, + 0.026600735262036324, + 0.040430400520563126, + 0.006268518976867199, + 0.011168576776981354, + 0.005685234908014536, + 0.0042428369633853436, + 0.01170012354850769, + 0.005752196069806814, + 0.03227660059928894, + 0.0028659687377512455, + 0.0043386491015553474, + 0.004129210952669382, + 0.003997220657765865, + 0.0032370020635426044, + 0.0029634113889187574, + 0.003598320996388793, + 0.007530576549470425, + 0.0034082229249179363, + 0.0060713752172887325, + 0.005174499470740557, + 0.007092795800417662, + 0.005148076917976141, + 0.004753817338496447, + 0.004282643087208271, + 0.005111874546855688, + 0.0038916589692234993, + 0.0048318603076040745, + 0.013513647951185703, + 0.008189518935978413, + 0.004240137990564108, + 0.004664616193622351, + 0.005926401820033789, + 0.0030585124623030424, + 0.010512725450098515, + 0.007245090324431658, + 0.010808252729475498, + 0.013465866446495056, + 0.015561634674668312, + 0.007109969388693571, + 0.012684742920100689, + 0.008113193325698376, + 0.021548928692936897, + 0.015364003367722034, + 0.010938440449535847, + 0.010269656777381897, + 0.004638234153389931, + 0.004189702682197094, + 0.0033010488841682673, + 0.004055758472532034, + 0.007834723219275475, + 0.007227463647723198, + 0.013224304653704166, + 0.004883536137640476, + 0.0031670499593019485, + 0.004039038438349962, + 0.0036390763707458973, + 0.003964709583669901, + 0.014627889730036259, + 0.02280268259346485, + 0.024843057617545128, + 0.010961574502289295, + 0.011036657728254795, + 0.02127009816467762, + 0.009338537231087685, + 0.01016334816813469, + 0.05107921361923218, + 0.032183192670345306, + 0.04085148125886917, + 0.03124833293259144, + 0.04665172100067139, + 0.03933953121304512, + 0.03788359463214874, + 0.03890271857380867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.03370155766606331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned overhead and tracking the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character Nezha visible with twin jet-black buns bound by a red headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is Nezha riding a blazing Wind Fire Wheel emitting orange-red flames and swirling currents?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the red Hun Tian Ling visible behind him, twirling and coiling like a shimmering silk dragon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark apocalyptic scene with deep blue/black clouds, purple lightning, and Nezha wearing red armor with golden lotus patterns?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005955095402896404 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.005955095402896404, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005955095402896404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08640249818563461, + 0.311384916305542, + 0.1556519716978073, + 0.1408669650554657, + 0.04607662558555603, + 0.04262089729309082, + 0.07151496410369873, + 0.02655165269970894, + 0.05173950642347336, + 0.0671900138258934 + ], + "credit_map": [ + 0.008537582121789455, + 0.005288354121148586, + 0.007494782097637653, + 0.005295914132148027, + 0.007091325707733631, + 0.014432060532271862, + 0.011358709074556828, + 0.026903782039880753, + 0.042543306946754456, + 0.009440891444683075, + 0.039279595017433167, + 0.00881831906735897, + 0.006354256067425013, + 0.09650809317827225, + 0.01117545086890459, + 0.09726499766111374, + 0.012652427889406681, + 0.015798144042491913, + 0.02979954704642296, + 0.01411393377929926, + 0.005585084669291973, + 0.006585957482457161, + 0.018856676295399666, + 0.052260201424360275, + 0.013798295520246029, + 0.02100127562880516, + 0.02062201127409935, + 0.03532367944717407, + 0.009066295810043812, + 0.00857458170503378, + 0.015349211171269417, + 0.017131607979536057, + 0.0037284372374415398, + 0.004540069494396448, + 0.013832409866154194, + 0.007693325635045767, + 0.005487855989485979, + 0.0035386725794523954, + 0.004945919383317232, + 0.0023099412210285664, + 0.005113806575536728, + 0.004557131789624691, + 0.007202798966318369, + 0.006993051618337631, + 0.005884524434804916, + 0.003597881644964218, + 0.0061692432500422, + 0.003102461574599147, + 0.013131573796272278, + 0.0108573567122221, + 0.01092802919447422, + 0.009913044050335884, + 0.005899262614548206, + 0.007243157364428043, + 0.00576816126704216, + 0.007774381432682276, + 0.004589125979691744, + 0.003760209307074547, + 0.006563912611454725, + 0.0025528804399073124, + 0.0018683780217543244, + 0.003090988378971815, + 0.0018082110909745097, + 0.0023179480340331793, + 0.006679981015622616, + 0.01154356636106968, + 0.007634490262717009, + 0.004522596951574087, + 0.004145415034145117, + 0.00922602042555809, + 0.003568807616829872, + 0.004418633878231049, + 0.012571539729833603, + 0.006106688641011715, + 0.010453385300934315, + 0.005271172150969505, + 0.011323463171720505, + 0.008563258685171604, + 0.00661100260913372, + 0.0062895044684410095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.005955095402896404 + } + ], + "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/0000158/group_2/index_tensor([4159, 4159])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_tensor([4159, 4159])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c4c77101f62b1fc2412ce59074e6955e2a1973e8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_tensor([4159, 4159])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, low-angle, medium-long shot slowly pans to reveal a young prince in his early twenties, standing defiantly amid the smoldering, smoke-filled ruins of a magnificent stone hall. The scene is rendered in the dramatic, romantic style of John William Waterhouse and Wadim Kashin, but entirely in stark, high-contrast black and white. A single, harsh light source from outside the camera pierces the darkness, creating a powerful contrast. This light captures the sharp planes of his face, highlighting a prominent, jagged white scar on his left cheekbone. His dark, wavy, shoulder-length hair gleams in the gentle draft. His expression is sullen, filled with sorrow and unwavering determination, his jaw set in a rigid position. He wears a high-collared, soot-stained tunic, once rich in fabric and intricately embroidered, now barely discernible. A tattered cape hangs from one shoulder. His right hand rests grimly on the hilt of an ornate longsword, his knuckles pale as he grips it. Around him, thick clouds of white and grey smoke swirl, catching the light and partially obscuring a backdrop of broken arches and piles of textured, shattered stone. A stark, monochrome palette emphasizes the gritty textures of stone, ash, and fabric, while high-contrast lighting carves out deep, inky shadows, creating an atmosphere of epic, heroic melancholy.", + "seed": 31489, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31489_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": -0.03924002195708454, + "vlm_logprob": -0.03924002195708454, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03924002195708454, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12128292769193649 + ], + "opposite_logprob_tokens": [ + -2.2462828159332275 + ], + "logprob_reward": -0.12128292769193649, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12128292769193649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12794987857341766, + 0.1027267575263977, + 0.08463326096534729, + 0.03425681218504906, + 0.3083030581474304, + 0.03538970649242401, + 0.020588137209415436, + 0.06560812145471573, + 0.058130159974098206, + 0.16241402924060822 + ], + "credit_map": [ + 0.015309016220271587, + 0.018056299537420273, + 0.024706633761525154, + 0.01308812852948904, + 0.012127495370805264, + 0.016486531123518944, + 0.015818657353520393, + 0.012357133440673351, + 0.01591574400663376, + 0.01277799904346466, + 0.010077095590531826, + 0.009955711662769318, + 0.00793106947094202, + 0.023801220580935478, + 0.007790359202772379, + 0.014477564953267574, + 0.012021022848784924, + 0.014672026969492435, + 0.010472425259649754, + 0.005849303677678108, + 0.010258696042001247, + 0.014591989107429981, + 0.007359332870692015, + 0.00940847210586071, + 0.004521435592323542, + 0.003927913028746843, + 0.0029841249343007803, + 0.004254057537764311, + 0.005016434472054243, + 0.005623478442430496, + 0.005351178348064423, + 0.0025781940203160048, + 0.06559400260448456, + 0.02850922755897045, + 0.03896919637918472, + 0.021737748757004738, + 0.028175735846161842, + 0.06576254218816757, + 0.03668932989239693, + 0.022865332663059235, + 0.004972189664840698, + 0.004545655567198992, + 0.003971736878156662, + 0.0028454556595534086, + 0.00368889095261693, + 0.005483552813529968, + 0.004039703402668238, + 0.005842526908963919, + 0.0017530254554003477, + 0.0023301932960748672, + 0.0020468418952077627, + 0.0026668140199035406, + 0.003454263787716627, + 0.0020252647809684277, + 0.0034964049700647593, + 0.0028153304010629654, + 0.008858684450387955, + 0.005488615948706865, + 0.012539833784103394, + 0.006114594172686338, + 0.007221003528684378, + 0.007789181545376778, + 0.007652938831597567, + 0.009943271055817604, + 0.005749114323407412, + 0.006049331743270159, + 0.007957019843161106, + 0.006244373042136431, + 0.008851747028529644, + 0.01009275484830141, + 0.0062927561812102795, + 0.006893065292388201, + 0.01889193058013916, + 0.01602678932249546, + 0.02783072181046009, + 0.022437656298279762, + 0.01747835986316204, + 0.01837538555264473, + 0.0205826535820961, + 0.020790545269846916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.12128292769193649 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598890259861946 + ], + "opposite_logprob_tokens": [ + -3.895598888397217 + ], + "logprob_reward": -0.020598890259861946, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598890259861946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18181449174880981, + 0.056844186037778854, + 0.05760670825839043, + 0.02913092076778412, + 0.0616285540163517, + 0.07160035520792007, + 0.0677667185664177, + 0.15294687449932098, + 0.06838055700063705, + 0.25228065252304077 + ], + "credit_map": [ + 0.025283850729465485, + 0.025238653644919395, + 0.015258165076375008, + 0.019188981503248215, + 0.029126925393939018, + 0.01683598756790161, + 0.03517267480492592, + 0.01570923998951912, + 0.007975985296070576, + 0.0087244538590312, + 0.005625713616609573, + 0.007906635291874409, + 0.006538089830428362, + 0.0056497505865991116, + 0.010171551257371902, + 0.004251998849213123, + 0.007828549481928349, + 0.009038802236318588, + 0.006522578652948141, + 0.007384655997157097, + 0.00692996010184288, + 0.006039966829121113, + 0.007302503101527691, + 0.00655968626961112, + 0.003711187746375799, + 0.0033622640185058117, + 0.0027113582473248243, + 0.0038369789253920317, + 0.004099028185009956, + 0.003426486859098077, + 0.005688849370926619, + 0.002294763922691345, + 0.014522708021104336, + 0.0065730563364923, + 0.008111078292131424, + 0.004895662888884544, + 0.00841489527374506, + 0.008867859840393066, + 0.006444533355534077, + 0.003798755118623376, + 0.0065900953486561775, + 0.006390431895852089, + 0.008875430561602116, + 0.007277155295014381, + 0.009489278309047222, + 0.008789344690740108, + 0.010074063204228878, + 0.014114553108811378, + 0.006584705784916878, + 0.006406078115105629, + 0.011401869356632233, + 0.007414852734655142, + 0.013321654871106148, + 0.005717724096029997, + 0.008411860093474388, + 0.008507971651852131, + 0.015994852408766747, + 0.012151436880230904, + 0.01724911481142044, + 0.01708785630762577, + 0.02269069105386734, + 0.030550289899110794, + 0.019205540418624878, + 0.01801707223057747, + 0.004315286874771118, + 0.008805249817669392, + 0.004485876299440861, + 0.012305879034101963, + 0.014396472834050655, + 0.005927157122641802, + 0.009965783916413784, + 0.00817884597927332, + 0.024702386930584908, + 0.03463233262300491, + 0.033947527408599854, + 0.027550872415304184, + 0.0373118631541729, + 0.03761589154601097, + 0.029272854328155518, + 0.027246907353401184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.020598890259861946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031967272516340017 + ], + "opposite_logprob_tokens": [ + -5.753196716308594 + ], + "logprob_reward": -0.0031967272516340017, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031967272516340017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16374222934246063, + 0.034161485731601715, + 0.030564557760953903, + 0.015864213928580284, + 0.04065176844596863, + 0.061737190932035446, + 0.054528914391994476, + 0.13039815425872803, + 0.10163941979408264, + 0.3667120933532715 + ], + "credit_map": [ + 0.01892491616308689, + 0.03403850272297859, + 0.015453126281499863, + 0.01773248240351677, + 0.01933644898235798, + 0.016249731183052063, + 0.028690636157989502, + 0.013316365890204906, + 0.004240855574607849, + 0.005906613077968359, + 0.004729931708425283, + 0.004184807185083628, + 0.0037237582728266716, + 0.0027489750646054745, + 0.005656743887811899, + 0.0029697977006435394, + 0.005256450269371271, + 0.0035110278986394405, + 0.0026942400727421045, + 0.004231947008520365, + 0.0031316911336034536, + 0.00429141242057085, + 0.005421149544417858, + 0.0020266363862901926, + 0.0022409758530557156, + 0.0018899498973041773, + 0.0016416989965364337, + 0.001660962705500424, + 0.0020391824655234814, + 0.0030636033043265343, + 0.0020748311653733253, + 0.001253008027561009, + 0.006230674218386412, + 0.003140864660963416, + 0.004696681164205074, + 0.0037085663061589003, + 0.005082332994788885, + 0.00653033796697855, + 0.00749353738501668, + 0.0037687711883336306, + 0.007672179955989122, + 0.0048599583096802235, + 0.015129432082176208, + 0.006093777250498533, + 0.0059282430447638035, + 0.008817018941044807, + 0.005544548854231834, + 0.007692031096667051, + 0.006361682433634996, + 0.00442494684830308, + 0.010695911943912506, + 0.007076072972267866, + 0.007660039700567722, + 0.005671106744557619, + 0.005550012458115816, + 0.007089137099683285, + 0.017395464703440666, + 0.008759749121963978, + 0.017448393628001213, + 0.011031052097678185, + 0.015131831169128418, + 0.02885143645107746, + 0.014511002227663994, + 0.017269205302000046, + 0.007616442162543535, + 0.012691160663962364, + 0.010709713213145733, + 0.015282428823411465, + 0.018893154338002205, + 0.010754363611340523, + 0.013492690399289131, + 0.012199457734823227, + 0.030033404007554054, + 0.054970864206552505, + 0.041007984429597855, + 0.04093972593545914, + 0.06080492585897446, + 0.05165060982108116, + 0.04742896556854248, + 0.03987560793757439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0031967272516340017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024880189448595047 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.0024880189448595047, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024880189448595047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17827679216861725, + 0.1477501541376114, + 0.15140631794929504, + 0.052899811416864395, + 0.09657295048236847, + 0.051271840929985046, + 0.038448721170425415, + 0.1029665544629097, + 0.10002706944942474, + 0.08037979900836945 + ], + "credit_map": [ + 0.020178522914648056, + 0.03614473342895508, + 0.02262994833290577, + 0.017903391271829605, + 0.014899092726409435, + 0.018163084983825684, + 0.02930835261940956, + 0.01904968172311783, + 0.023978812620043755, + 0.03349066153168678, + 0.018182585015892982, + 0.013188769109547138, + 0.01320022065192461, + 0.022470861673355103, + 0.011398631148040295, + 0.011839617975056171, + 0.02351972460746765, + 0.01663029193878174, + 0.02412654645740986, + 0.018037594854831696, + 0.016214704141020775, + 0.018725402653217316, + 0.019565779715776443, + 0.014586275443434715, + 0.006633646320551634, + 0.00558528071269393, + 0.007147712167352438, + 0.006859481334686279, + 0.005123816430568695, + 0.007497007492929697, + 0.00939260609447956, + 0.0046602641232311726, + 0.019950583577156067, + 0.008934549987316132, + 0.01401819009333849, + 0.00843028724193573, + 0.007376955822110176, + 0.01862274669110775, + 0.012662095949053764, + 0.006577544379979372, + 0.0057587879709899426, + 0.004626102279871702, + 0.007548558991402388, + 0.004756289068609476, + 0.004994198679924011, + 0.007302498910576105, + 0.007026337552815676, + 0.00925907026976347, + 0.0044378861784935, + 0.00389347062446177, + 0.0048586297780275345, + 0.004003002308309078, + 0.005597284529358149, + 0.005100938957184553, + 0.005107904318720102, + 0.005449604243040085, + 0.011781557463109493, + 0.010332915931940079, + 0.012758071534335613, + 0.0112895043566823, + 0.018902327865362167, + 0.013654478825628757, + 0.0120796924456954, + 0.01216799858957529, + 0.008610326796770096, + 0.011281928047537804, + 0.011454936116933823, + 0.014526495710015297, + 0.01743962988257408, + 0.010579490102827549, + 0.013560712337493896, + 0.012573556043207645, + 0.008016128093004227, + 0.01112905889749527, + 0.011427398771047592, + 0.011411002837121487, + 0.009820378385484219, + 0.008419686928391457, + 0.011479285545647144, + 0.008676856756210327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0024880189448595047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863354563713074 + ], + "opposite_logprob_tokens": [ + -3.048633575439453 + ], + "logprob_reward": -0.04863354563713074, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863354563713074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05706492438912392, + 0.05049267038702965, + 0.05522657185792923, + 0.041882824152708054, + 0.0689472034573555, + 0.04397400841116905, + 0.0578099749982357, + 0.11638196557760239, + 0.39484986662864685, + 0.11336992681026459 + ], + "credit_map": [ + 0.007269350811839104, + 0.010109060443937778, + 0.006400671787559986, + 0.004864091519266367, + 0.008125648833811283, + 0.007837440818548203, + 0.007649880833923817, + 0.004808782134205103, + 0.0073376805521547794, + 0.006918248254805803, + 0.006035594269633293, + 0.0045795137993991375, + 0.005159229505807161, + 0.010996616445481777, + 0.005130284931510687, + 0.004335507750511169, + 0.007292466703802347, + 0.0076914564706385136, + 0.00559245515614748, + 0.006235317327082157, + 0.008240488357841969, + 0.007598807103931904, + 0.00800828356295824, + 0.004567303229123354, + 0.005607869476079941, + 0.0033471835777163506, + 0.0036118007265031338, + 0.006093891337513924, + 0.006085368804633617, + 0.005547387525439262, + 0.007614839822053909, + 0.00397448567673564, + 0.009377392940223217, + 0.0039510768838226795, + 0.005517987534403801, + 0.0032296860590577126, + 0.005504359491169453, + 0.020969640463590622, + 0.014191937632858753, + 0.006205134093761444, + 0.006378263235092163, + 0.004448732361197472, + 0.005278168246150017, + 0.003313460387289524, + 0.004468715749680996, + 0.005852126516401768, + 0.005090857855975628, + 0.009143687784671783, + 0.006214811932295561, + 0.006366819608956575, + 0.00568812433630228, + 0.007476632483303547, + 0.006726288236677647, + 0.009311416186392307, + 0.008265646174550056, + 0.007760243024677038, + 0.01128893718123436, + 0.010274681262671947, + 0.015561731532216072, + 0.01307580154389143, + 0.021189428865909576, + 0.013542138040065765, + 0.01589522883296013, + 0.01555402297526598, + 0.02959080971777439, + 0.04314495250582695, + 0.03496883437037468, + 0.06305216997861862, + 0.0767691433429718, + 0.05237375572323799, + 0.04939578101038933, + 0.045554421842098236, + 0.011345676146447659, + 0.013126658275723457, + 0.016049090772867203, + 0.014436987228691578, + 0.014828692190349102, + 0.015072420239448547, + 0.01547691784799099, + 0.013033496215939522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.04863354563713074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, medium-long shot slowly pans to reveal a young prince in his early twenties, standing defiantly amid the smoldering, smoke-filled ruins of a magnificent stone hall. The scene is rendered in the dramatic, romantic style of John William Waterhouse and Wadim Kashin, but entirely in stark, high-contrast black and white. A single, harsh light source from outside the camera pierces the darkness, creating a powerful contrast. This light captures the sharp planes of his face, highlighting a prominent, jagged white scar on his left cheekbone. His dark, wavy, shoulder-length hair gleams in the gentle draft. His expression is sullen, filled with sorrow and unwavering determination, his jaw set in a rigid position. He wears a high-collared, soot-stained tunic, once rich in fabric and intricately embroidered, now barely discernible. A tattered cape hangs from one shoulder. His right hand rests grimly on the hilt of an ornate longsword, his knuckles pale as he grips it. Around him, thick clouds of white and grey smoke swirl, catching the light and partially obscuring a backdrop of broken arches and piles of textured, shattered stone. A stark, monochrome palette emphasizes the gritty textures of stone, ash, and fabric, while high-contrast lighting carves out deep, inky shadows, creating an atmosphere of epic, heroic melancholy.", + "seed": 31490, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31490_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": -0.04282596716657281, + "vlm_logprob": -0.04282596716657281, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.04282596716657281, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10539703816175461 + ], + "opposite_logprob_tokens": [ + -2.3553969860076904 + ], + "logprob_reward": -0.10539703816175461, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10539703816175461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41999030113220215, + 0.1111128181219101, + 0.15707914531230927, + 0.03192564472556114, + 0.03569016605615616, + 0.044535353779792786, + 0.025286497548222542, + 0.030538329854607582, + 0.05926261097192764, + 0.08457912504673004 + ], + "credit_map": [ + 0.036637865006923676, + 0.0760146975517273, + 0.05237184837460518, + 0.024178408086299896, + 0.031013809144496918, + 0.14131008088588715, + 0.03434351086616516, + 0.024120120331645012, + 0.012183543294668198, + 0.011933666653931141, + 0.011509360745549202, + 0.008791033178567886, + 0.012049411423504353, + 0.031589239835739136, + 0.01414510142058134, + 0.008911464363336563, + 0.03205322101712227, + 0.02737063355743885, + 0.00934090930968523, + 0.014108697883784771, + 0.01750931330025196, + 0.019497014582157135, + 0.017407214269042015, + 0.01979212649166584, + 0.0021239423658698797, + 0.004933079704642296, + 0.0028626322746276855, + 0.0027345262933522463, + 0.005233959294855595, + 0.004960338585078716, + 0.006055805366486311, + 0.003021357348188758, + 0.003475453471764922, + 0.006345958448946476, + 0.0040796296671032906, + 0.003107686061412096, + 0.004510100930929184, + 0.00508160050958395, + 0.00406882306560874, + 0.005020910874009132, + 0.0066514029167592525, + 0.006983762141317129, + 0.005505299661308527, + 0.00416430551558733, + 0.005149109289050102, + 0.00593006843701005, + 0.006126809399574995, + 0.0040245987474918365, + 0.0027879055123776197, + 0.0038413950242102146, + 0.0034334470983594656, + 0.002269065473228693, + 0.00306724407710135, + 0.0037545987870544195, + 0.0034403339959681034, + 0.002692508278414607, + 0.004597388673573732, + 0.004988276399672031, + 0.005091329105198383, + 0.002524813637137413, + 0.003927676007151604, + 0.0038175724912434816, + 0.003138815751299262, + 0.002452456159517169, + 0.005741730332374573, + 0.012213102541863918, + 0.009428361430764198, + 0.003192046657204628, + 0.005196076352149248, + 0.01335129328072071, + 0.005032057873904705, + 0.0051079411059618, + 0.009774254634976387, + 0.011941341683268547, + 0.016262350603938103, + 0.009471447207033634, + 0.011513412930071354, + 0.008922893553972244, + 0.010481753386557102, + 0.00621166592463851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.10539703816175461 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496896430850029 + ], + "opposite_logprob_tokens": [ + -2.929969072341919 + ], + "logprob_reward": -0.05496896430850029, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496896430850029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3474104404449463, + 0.20074670016765594, + 0.09594163298606873, + 0.041828323155641556, + 0.018364720046520233, + 0.07197055220603943, + 0.0372987724840641, + 0.021263904869556427, + 0.08441624790430069, + 0.0807587131857872 + ], + "credit_map": [ + 0.017233718186616898, + 0.14621832966804504, + 0.020570285618305206, + 0.009480643086135387, + 0.019751545041799545, + 0.10116581618785858, + 0.024243025109171867, + 0.008747062645852566, + 0.013919462449848652, + 0.024846266955137253, + 0.024086978286504745, + 0.011839981190860271, + 0.014981436543166637, + 0.07592593878507614, + 0.024782627820968628, + 0.01036399882286787, + 0.024411601945757866, + 0.021671460941433907, + 0.006248911377042532, + 0.005960689391940832, + 0.00999325979501009, + 0.011623273603618145, + 0.008832174353301525, + 0.007200251333415508, + 0.0024253663141280413, + 0.005812403745949268, + 0.0022889513056725264, + 0.00390451867133379, + 0.008702111430466175, + 0.006413607858121395, + 0.0067089395597577095, + 0.005572419613599777, + 0.0012092149117961526, + 0.0023697963915765285, + 0.0015882218722254038, + 0.0015497102867811918, + 0.004541453439742327, + 0.0027840323746204376, + 0.0023921343963593245, + 0.0019301552092656493, + 0.007964110001921654, + 0.009150958620011806, + 0.006172423250973225, + 0.0077257342636585236, + 0.014339771121740341, + 0.00966369453817606, + 0.010709373280405998, + 0.006244487129151821, + 0.003341896226629615, + 0.004096069373190403, + 0.0029471369925886393, + 0.0030321909580379725, + 0.00983202364295721, + 0.005332141648977995, + 0.004803456831723452, + 0.003913854714483023, + 0.0016515539027750492, + 0.0027698769699782133, + 0.0012418836122378707, + 0.0018415977247059345, + 0.0052365572191774845, + 0.003282204270362854, + 0.003698222804814577, + 0.0015420090639963746, + 0.0046510109677910805, + 0.01903812400996685, + 0.0056870863772928715, + 0.006451815366744995, + 0.012657348066568375, + 0.01881936937570572, + 0.0117160864174366, + 0.005395400337874889, + 0.00617178762331605, + 0.011235401965677738, + 0.007003379985690117, + 0.008330154232680798, + 0.020958568900823593, + 0.01212377566844225, + 0.009899057447910309, + 0.005036586429923773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05496896430850029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31008055806159973, + 0.16519704461097717, + 0.07111644744873047, + 0.03466188162565231, + 0.035026226192712784, + 0.09905912727117538, + 0.06733950972557068, + 0.032938066869974136, + 0.08935746550559998, + 0.09522370249032974 + ], + "credit_map": [ + 0.017038894817233086, + 0.12465046346187592, + 0.013072292320430279, + 0.012290802784264088, + 0.016751574352383614, + 0.09716201573610306, + 0.019384706392884254, + 0.009729823097586632, + 0.011475997976958752, + 0.021160027012228966, + 0.01919240690767765, + 0.00992124155163765, + 0.013146130368113518, + 0.06108547002077103, + 0.020019061863422394, + 0.009196704253554344, + 0.01821018010377884, + 0.011795446276664734, + 0.003610382555052638, + 0.004719994030892849, + 0.009055500850081444, + 0.009319968521595001, + 0.00806325301527977, + 0.006341723259538412, + 0.0028572436422109604, + 0.004833560902625322, + 0.00226034689694643, + 0.0030097553972154856, + 0.007228373549878597, + 0.005354446358978748, + 0.005564506631344557, + 0.0035536508075892925, + 0.0022933166474103928, + 0.004782607778906822, + 0.002790529979392886, + 0.0033225303050130606, + 0.00700563145801425, + 0.005270171444863081, + 0.0055459062568843365, + 0.004015529993921518, + 0.011536122299730778, + 0.016626177355647087, + 0.012249428778886795, + 0.009961114265024662, + 0.013104530982673168, + 0.013794089667499065, + 0.012353683821856976, + 0.009433976374566555, + 0.005422517191618681, + 0.007625994738191366, + 0.004782159347087145, + 0.005247895140200853, + 0.01742246188223362, + 0.008593627251684666, + 0.010558363981544971, + 0.0076864901930093765, + 0.0024273418821394444, + 0.003947885241359472, + 0.003001806791871786, + 0.00330824451521039, + 0.008351225405931473, + 0.003927518613636494, + 0.005832960829138756, + 0.002141082426533103, + 0.004803075920790434, + 0.022785572335124016, + 0.005970031954348087, + 0.006259472109377384, + 0.016160331666469574, + 0.017414579167962074, + 0.0107443667948246, + 0.0052200304344296455, + 0.006037262734025717, + 0.014371765777468681, + 0.008811290375888348, + 0.00934511236846447, + 0.024598808959126472, + 0.014292505569756031, + 0.011831753887236118, + 0.005935206543654203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.003193875541910529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019377756398171186 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019377756398171186, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019377756398171186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1568928062915802, + 0.31441134214401245, + 0.05881299823522568, + 0.05120759829878807, + 0.021664177998900414, + 0.06928012520074844, + 0.05155821889638901, + 0.0207703597843647, + 0.22402779757976532, + 0.031374599784612656 + ], + "credit_map": [ + 0.01297052949666977, + 0.031482987105846405, + 0.011449401266872883, + 0.015482394024729729, + 0.01839764602482319, + 0.0394669845700264, + 0.01707036979496479, + 0.010572483763098717, + 0.023649681359529495, + 0.04106578603386879, + 0.03762313351035118, + 0.01771005056798458, + 0.02359941229224205, + 0.11805319041013718, + 0.03685684874653816, + 0.015853239223361015, + 0.014115622267127037, + 0.011873910203576088, + 0.0034847897477447987, + 0.004729586187750101, + 0.0046919058077037334, + 0.008141499012708664, + 0.006151926703751087, + 0.005623750854283571, + 0.004396411590278149, + 0.009747269563376904, + 0.004766599740833044, + 0.00528852641582489, + 0.007102263625711203, + 0.008075159043073654, + 0.006316324230283499, + 0.005515038035809994, + 0.0016822870820760727, + 0.003497362369671464, + 0.00240695895627141, + 0.002559447893872857, + 0.002569225849583745, + 0.0031100697815418243, + 0.0035722695756703615, + 0.0022665527649223804, + 0.008322943933308125, + 0.012899769470095634, + 0.009659565053880215, + 0.006637086160480976, + 0.007516378536820412, + 0.010427648201584816, + 0.007451867684721947, + 0.00636485917493701, + 0.007311755791306496, + 0.008068207651376724, + 0.007096376270055771, + 0.00407382333651185, + 0.008393554016947746, + 0.006177227012813091, + 0.006083829794079065, + 0.0043534389697015285, + 0.0030019402038306, + 0.0031082313507795334, + 0.0029532990884035826, + 0.0013623615959659219, + 0.0025846275966614485, + 0.0030260994099080563, + 0.002322012558579445, + 0.0024117871653288603, + 0.015124890021979809, + 0.06305089592933655, + 0.02677544206380844, + 0.010762007907032967, + 0.011860297992825508, + 0.06727901101112366, + 0.015355303883552551, + 0.013819935731589794, + 0.003844442544505, + 0.005244826432317495, + 0.005788683425635099, + 0.002896067453548312, + 0.0034564072266221046, + 0.0038959148805588484, + 0.003390850033611059, + 0.0028574056923389435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0019377756398171186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486321821808815 + ], + "opposite_logprob_tokens": [ + -3.0486321449279785 + ], + "logprob_reward": -0.0486321821808815, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486321821808815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5047295689582825, + 0.10430478304624557, + 0.04012960568070412, + 0.041300103068351746, + 0.017002219334244728, + 0.09439059346914291, + 0.03008418157696724, + 0.018280763179063797, + 0.1268961876630783, + 0.02288200333714485 + ], + "credit_map": [ + 0.027517937123775482, + 0.19796660542488098, + 0.01796337217092514, + 0.013077390380203724, + 0.029520107433199883, + 0.17489126324653625, + 0.027377396821975708, + 0.016415534541010857, + 0.00854093860834837, + 0.013748189434409142, + 0.010820557363331318, + 0.007330657914280891, + 0.013953624293208122, + 0.030300555750727654, + 0.013064436614513397, + 0.00654581980779767, + 0.0041312831453979015, + 0.005645186174660921, + 0.0024974015541374683, + 0.0035409340634942055, + 0.007543767802417278, + 0.00587351992726326, + 0.006022033281624317, + 0.004875478334724903, + 0.00412176176905632, + 0.004499065689742565, + 0.0038215783424675465, + 0.003431848017498851, + 0.007457766216248274, + 0.005846918560564518, + 0.005765935406088829, + 0.0063552288338541985, + 0.0013668994652107358, + 0.0023389868438243866, + 0.001387899974361062, + 0.0015256815822795033, + 0.0028652800247073174, + 0.002452564425766468, + 0.0025182818062603474, + 0.0025466245133429766, + 0.01008650753647089, + 0.017782039940357208, + 0.012244906276464462, + 0.010296016000211239, + 0.009811275638639927, + 0.013454679399728775, + 0.01137174665927887, + 0.009343422017991543, + 0.0025807905476540327, + 0.004152658861130476, + 0.00388407614082098, + 0.002724585123360157, + 0.006401839666068554, + 0.00354430777952075, + 0.004259936977177858, + 0.002535988111048937, + 0.0024232184514403343, + 0.003525385633111, + 0.0016642117407172918, + 0.001191551797091961, + 0.0017188771162182093, + 0.004323954228311777, + 0.0017370317364111543, + 0.001696532010100782, + 0.00822236854583025, + 0.03261851891875267, + 0.010472684167325497, + 0.007791541051119566, + 0.016171706840395927, + 0.030574975535273552, + 0.014828150160610676, + 0.006216233130544424, + 0.003016059286892414, + 0.0031834100373089314, + 0.0031698073726147413, + 0.0018587742233648896, + 0.004325428511947393, + 0.0029999222606420517, + 0.0029970151372253895, + 0.0013315863907337189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0486321821808815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, medium-long shot slowly pans to reveal a young prince in his early twenties, standing defiantly amid the smoldering, smoke-filled ruins of a magnificent stone hall. The scene is rendered in the dramatic, romantic style of John William Waterhouse and Wadim Kashin, but entirely in stark, high-contrast black and white. A single, harsh light source from outside the camera pierces the darkness, creating a powerful contrast. This light captures the sharp planes of his face, highlighting a prominent, jagged white scar on his left cheekbone. His dark, wavy, shoulder-length hair gleams in the gentle draft. His expression is sullen, filled with sorrow and unwavering determination, his jaw set in a rigid position. He wears a high-collared, soot-stained tunic, once rich in fabric and intricately embroidered, now barely discernible. A tattered cape hangs from one shoulder. His right hand rests grimly on the hilt of an ornate longsword, his knuckles pale as he grips it. Around him, thick clouds of white and grey smoke swirl, catching the light and partially obscuring a backdrop of broken arches and piles of textured, shattered stone. A stark, monochrome palette emphasizes the gritty textures of stone, ash, and fabric, while high-contrast lighting carves out deep, inky shadows, creating an atmosphere of epic, heroic melancholy.", + "seed": 31491, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31491_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": -0.05098597288597375, + "vlm_logprob": -0.05098597288597375, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.05098597288597375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20789797604084015 + ], + "opposite_logprob_tokens": [ + -1.7078980207443237 + ], + "logprob_reward": -0.20789797604084015, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20789797604084015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7626872062683105, + 0.044229865074157715, + 0.049712829291820526, + 0.041903525590896606, + 0.018140282481908798, + 0.025096356868743896, + 0.010217364877462387, + 0.007906198501586914, + 0.015197076834738255, + 0.024909349158406258 + ], + "credit_map": [ + 0.2509783208370209, + 0.04229312762618065, + 0.3047518730163574, + 0.041281599551439285, + 0.03290025144815445, + 0.03305817022919655, + 0.03263223543763161, + 0.024791641160845757, + 0.00370118604041636, + 0.010692496784031391, + 0.003959733992815018, + 0.003722795285284519, + 0.005370992235839367, + 0.003383395029231906, + 0.003499734913930297, + 0.009899532422423363, + 0.007716283202171326, + 0.00426639337092638, + 0.003804786130785942, + 0.003183049848303199, + 0.0034991579595953226, + 0.01225341111421585, + 0.003431792138144374, + 0.011557955294847488, + 0.00292095891200006, + 0.012139725498855114, + 0.002737676724791527, + 0.0031879176385700703, + 0.003430793061852455, + 0.003816056763753295, + 0.004245685413479805, + 0.009424708783626556, + 0.0014605853939428926, + 0.0028731082566082478, + 0.0019940859638154507, + 0.002333121607080102, + 0.0029197493568062782, + 0.0017688162624835968, + 0.0030453824438154697, + 0.0017454336630180478, + 0.002293587662279606, + 0.003846315434202552, + 0.003736702958121896, + 0.003545175539329648, + 0.0033479854464530945, + 0.0028315347153693438, + 0.0026802693028002977, + 0.0028147853445261717, + 0.001448969473131001, + 0.0011202138848602772, + 0.0009658823255449533, + 0.0013176221400499344, + 0.0010159261291846633, + 0.0008582001319155097, + 0.001127442461438477, + 0.002363108564168215, + 0.0011145665775984526, + 0.0008908024174161255, + 0.0010597710497677326, + 0.0010728600900620222, + 0.0011247305665165186, + 0.000753312255255878, + 0.0010024060029536486, + 0.0008877500658854842, + 0.0019024356734007597, + 0.0018049507634714246, + 0.0017844844842329621, + 0.0018187720561400056, + 0.0019107712432742119, + 0.002313558477908373, + 0.0017429230501875281, + 0.001919182133860886, + 0.0022112291771918535, + 0.0019180761883035302, + 0.0009579432662576437, + 0.00617279252037406, + 0.003740232205018401, + 0.0028429729864001274, + 0.0021544815972447395, + 0.004911620635539293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.20789797604084015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647258393466473 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647258393466473, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647258393466473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3362196683883667, + 0.05702706053853035, + 0.1926022171974182, + 0.09455903619527817, + 0.0666624903678894, + 0.05033491924405098, + 0.11328969150781631, + 0.012782737612724304, + 0.01998330093920231, + 0.05653892084956169 + ], + "credit_map": [ + 0.11086626350879669, + 0.01729004643857479, + 0.1338554322719574, + 0.018400685861706734, + 0.016254374757409096, + 0.01235227845609188, + 0.01519727148115635, + 0.012003292329609394, + 0.0038474202156066895, + 0.011639365926384926, + 0.00450329901650548, + 0.007165675517171621, + 0.009457866661250591, + 0.005637063179165125, + 0.005479022394865751, + 0.00929735042154789, + 0.014812403358519077, + 0.03117264248430729, + 0.01805858314037323, + 0.013952902518212795, + 0.02364007942378521, + 0.04770146310329437, + 0.016010578721761703, + 0.02725355140864849, + 0.007683299481868744, + 0.018455229699611664, + 0.004216078668832779, + 0.0047220648266375065, + 0.021596388891339302, + 0.00698079215362668, + 0.015741998329758644, + 0.015163191594183445, + 0.006332330405712128, + 0.014018822461366653, + 0.009815608151257038, + 0.004569138865917921, + 0.011599920690059662, + 0.006149228196591139, + 0.009612356312572956, + 0.004565082956105471, + 0.005714327562600374, + 0.009152581915259361, + 0.0048102582804858685, + 0.0039324951358139515, + 0.008363381028175354, + 0.007132241502404213, + 0.004950912203639746, + 0.006278720684349537, + 0.017299212515354156, + 0.011078676208853722, + 0.011931059882044792, + 0.010302720591425896, + 0.011580836027860641, + 0.009179575368762016, + 0.009943725541234016, + 0.031973883509635925, + 0.0013026217930018902, + 0.0016789415385574102, + 0.0012246286496520042, + 0.0017825778340920806, + 0.0021981848403811455, + 0.0011792182922363281, + 0.0018516071140766144, + 0.0015649573178961873, + 0.002674624789506197, + 0.0018912989180535078, + 0.0031848729122430086, + 0.001887466642074287, + 0.0027486577164381742, + 0.002094672992825508, + 0.0032010702416300774, + 0.002300635678693652, + 0.0043268087320029736, + 0.005126672796905041, + 0.0022632726468145847, + 0.014403051696717739, + 0.005647419020533562, + 0.007865243591368198, + 0.005328682716935873, + 0.011577770113945007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.008647258393466473 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193400101736188 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.003193400101736188, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193400101736188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660171091556549, + 0.08568157255649567, + 0.09567669779062271, + 0.11987534165382385, + 0.10227272659540176, + 0.0695604458451271, + 0.08557955920696259, + 0.04061907157301903, + 0.0239711981266737, + 0.11074626445770264 + ], + "credit_map": [ + 0.08006884157657623, + 0.016055908054113388, + 0.09059792011976242, + 0.01399902906268835, + 0.018872566521167755, + 0.018089866265654564, + 0.016246480867266655, + 0.01208649855107069, + 0.005433228798210621, + 0.014776543714106083, + 0.007471045944839716, + 0.0102176982909441, + 0.014513581059873104, + 0.010762172751128674, + 0.010617385618388653, + 0.011889917775988579, + 0.0073531316593289375, + 0.014119752682745457, + 0.010788622312247753, + 0.008140778169035912, + 0.010260512121021748, + 0.022119026631116867, + 0.008403042331337929, + 0.01449184026569128, + 0.009534935466945171, + 0.0236886665225029, + 0.007286426145583391, + 0.009140673093497753, + 0.023348219692707062, + 0.010992984287440777, + 0.01569576933979988, + 0.02018766477704048, + 0.009455389343202114, + 0.021512268111109734, + 0.010151940397918224, + 0.007874814793467522, + 0.020285656675696373, + 0.012091733515262604, + 0.013670322485268116, + 0.007230618502944708, + 0.007944299839437008, + 0.008127325214445591, + 0.011671128682792187, + 0.009396040812134743, + 0.009206139482557774, + 0.0087240906432271, + 0.00916195847094059, + 0.005329470615833998, + 0.01371487695723772, + 0.0091189444065094, + 0.007452891208231449, + 0.00702672591432929, + 0.0066430820152163506, + 0.006993778515607119, + 0.006944545079022646, + 0.02768472023308277, + 0.005103675182908773, + 0.005534617230296135, + 0.005516536999493837, + 0.00443639699369669, + 0.006137324497103691, + 0.005053033120930195, + 0.004160856828093529, + 0.00467663025483489, + 0.0024813388008624315, + 0.0032941836398094893, + 0.0022020081523805857, + 0.0027146139182150364, + 0.0027892838697880507, + 0.0024747177958488464, + 0.00454450910910964, + 0.0034705461002886295, + 0.00824978481978178, + 0.0071333725936710835, + 0.004094269592314959, + 0.03657931461930275, + 0.009510939940810204, + 0.007285977713763714, + 0.007283677347004414, + 0.03060894086956978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.003193400101736188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015099564334377646 + ], + "opposite_logprob_tokens": [ + -6.501510143280029 + ], + "logprob_reward": -0.0015099564334377646, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015099564334377646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17831674218177795, + 0.06557602435350418, + 0.17535270750522614, + 0.14425504207611084, + 0.08916281163692474, + 0.07724706083536148, + 0.052505768835544586, + 0.035169195383787155, + 0.05046774446964264, + 0.13194693624973297 + ], + "credit_map": [ + 0.05211444944143295, + 0.010821635834872723, + 0.05916726216673851, + 0.010028154589235783, + 0.011610880494117737, + 0.01659591682255268, + 0.00954380538314581, + 0.008434614166617393, + 0.00525319017469883, + 0.01289864256978035, + 0.005914994981139898, + 0.005460856948047876, + 0.0107779735699296, + 0.007258173078298569, + 0.005173827521502972, + 0.012838358990848064, + 0.013983390294015408, + 0.025674154981970787, + 0.015076249837875366, + 0.01326757948845625, + 0.016935359686613083, + 0.04246087372303009, + 0.013439597561955452, + 0.03451547399163246, + 0.01482261624187231, + 0.04007123038172722, + 0.009693720377981663, + 0.013505238108336926, + 0.01153271272778511, + 0.015664895996451378, + 0.013052037917077541, + 0.025912584736943245, + 0.012254386208951473, + 0.02459551952779293, + 0.0072446102276444435, + 0.009959724731743336, + 0.010309228673577309, + 0.0074517205357551575, + 0.011295760050415993, + 0.006051856093108654, + 0.007986987940967083, + 0.015036219730973244, + 0.010570165701210499, + 0.006981661077588797, + 0.010366182774305344, + 0.0100138820707798, + 0.008428218774497509, + 0.00786373857408762, + 0.0058739627711474895, + 0.007048565428704023, + 0.006692881230264902, + 0.006628667004406452, + 0.006227097474038601, + 0.006705702282488346, + 0.0064969733357429504, + 0.006831916980445385, + 0.00334145943634212, + 0.0033742585219442844, + 0.0038027120754122734, + 0.004737412091344595, + 0.006726156920194626, + 0.0039464980363845825, + 0.005400916561484337, + 0.003839778481051326, + 0.006565715651959181, + 0.006442677229642868, + 0.008015201427042484, + 0.006319418083876371, + 0.006134468130767345, + 0.005310668144375086, + 0.00632160808891058, + 0.005357986316084862, + 0.009032719768583775, + 0.008724366314709187, + 0.005025544203817844, + 0.04311998561024666, + 0.011691595427691936, + 0.00821920856833458, + 0.00933522917330265, + 0.036798279732465744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0015099564334377646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033681273460388184 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.033681273460388184, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033681273460388184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41792532801628113, + 0.0811949223279953, + 0.1361699402332306, + 0.09720048308372498, + 0.0654490739107132, + 0.06080539897084236, + 0.04018845036625862, + 0.01967283897101879, + 0.029997732490301132, + 0.05139581486582756 + ], + "credit_map": [ + 0.13984738290309906, + 0.02476678416132927, + 0.16291120648384094, + 0.02116337977349758, + 0.019245006144046783, + 0.014987309463322163, + 0.02000652439892292, + 0.01499771885573864, + 0.005570617504417896, + 0.013389702886343002, + 0.006584614980965853, + 0.008887398056685925, + 0.017299894243478775, + 0.009611189365386963, + 0.00662652775645256, + 0.013224977999925613, + 0.009386491030454636, + 0.017501791939139366, + 0.013472383841872215, + 0.01062864251434803, + 0.012543419376015663, + 0.034193627536296844, + 0.011969359591603279, + 0.026474231854081154, + 0.009076191112399101, + 0.0176401287317276, + 0.0053389957174658775, + 0.009246738627552986, + 0.019096262753009796, + 0.010180855169892311, + 0.013970892876386642, + 0.01265041809529066, + 0.007154021877795458, + 0.013866993598639965, + 0.005606417544186115, + 0.008783189579844475, + 0.01061181165277958, + 0.004846241790801287, + 0.009965687990188599, + 0.004614708945155144, + 0.004938141908496618, + 0.007495695259422064, + 0.008534670807421207, + 0.007198049686849117, + 0.009930901229381561, + 0.007978549227118492, + 0.009214409627020359, + 0.0055149816907942295, + 0.005128378048539162, + 0.005006170831620693, + 0.0035820601042360067, + 0.004908697213977575, + 0.005708727054297924, + 0.003962615970522165, + 0.004587353672832251, + 0.007304446771740913, + 0.0024806479923427105, + 0.002080813981592655, + 0.0018646139651536942, + 0.002491073915734887, + 0.0023229580838233232, + 0.003573604393750429, + 0.002645868808031082, + 0.002213257597759366, + 0.003369109472259879, + 0.0031457378063350916, + 0.004077224992215633, + 0.0034386091865599155, + 0.004129019100219011, + 0.0031187962740659714, + 0.005536166951060295, + 0.003183070570230484, + 0.003230164060369134, + 0.003356087254360318, + 0.0021381385158747435, + 0.014594164676964283, + 0.007218477316200733, + 0.003635221626609564, + 0.0046004969626665115, + 0.012623063288629055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.033681273460388184 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, low-angle, medium-long shot slowly pans to reveal a young prince in his early twenties, standing defiantly amid the smoldering, smoke-filled ruins of a magnificent stone hall. The scene is rendered in the dramatic, romantic style of John William Waterhouse and Wadim Kashin, but entirely in stark, high-contrast black and white. A single, harsh light source from outside the camera pierces the darkness, creating a powerful contrast. This light captures the sharp planes of his face, highlighting a prominent, jagged white scar on his left cheekbone. His dark, wavy, shoulder-length hair gleams in the gentle draft. His expression is sullen, filled with sorrow and unwavering determination, his jaw set in a rigid position. He wears a high-collared, soot-stained tunic, once rich in fabric and intricately embroidered, now barely discernible. A tattered cape hangs from one shoulder. His right hand rests grimly on the hilt of an ornate longsword, his knuckles pale as he grips it. Around him, thick clouds of white and grey smoke swirl, catching the light and partially obscuring a backdrop of broken arches and piles of textured, shattered stone. A stark, monochrome palette emphasizes the gritty textures of stone, ash, and fabric, while high-contrast lighting carves out deep, inky shadows, creating an atmosphere of epic, heroic melancholy.", + "seed": 31492, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_2/index_4159_seed_31492_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": -0.05099411071278155, + "vlm_logprob": -0.05099411071278155, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.05099411071278155, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1842605471611023 + ], + "opposite_logprob_tokens": [ + -1.809260606765747 + ], + "logprob_reward": -0.1842605471611023, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1842605471611023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05704927071928978, + 0.010781257413327694, + 0.013749012723565102, + 0.01911952719092369, + 0.018431179225444794, + 0.040906790643930435, + 0.037817973643541336, + 0.7098200917243958, + 0.0847630426287651, + 0.0075617567636072636 + ], + "credit_map": [ + 0.009708262048661709, + 0.007974417880177498, + 0.008350088261067867, + 0.0041946303099393845, + 0.006500440649688244, + 0.007237069308757782, + 0.0060723512433469296, + 0.007012012414634228, + 0.0021583205088973045, + 0.0010579570662230253, + 0.0014810241991654038, + 0.0006058111321181059, + 0.0011559297563508153, + 0.0014617140404880047, + 0.0008914894424378872, + 0.001969012198969722, + 0.0018565044738352299, + 0.0012970682000741363, + 0.0018288977444171906, + 0.0014127298491075635, + 0.001963167916983366, + 0.002406278159469366, + 0.0014899210073053837, + 0.001494445838034153, + 0.0022840106394141912, + 0.00352874887175858, + 0.002337116515263915, + 0.002060540020465851, + 0.0019157768692821264, + 0.0038933795876801014, + 0.0016882057534530759, + 0.001411748700775206, + 0.0022662843111902475, + 0.0015839991392567754, + 0.0027699305210262537, + 0.0016437218291684985, + 0.002825845032930374, + 0.003165874630212784, + 0.0023697789292782545, + 0.0018057443667203188, + 0.0056088389828801155, + 0.005678871646523476, + 0.009331141598522663, + 0.003020161297172308, + 0.003581920638680458, + 0.007851608097553253, + 0.0024999461602419615, + 0.003334302455186844, + 0.00864478200674057, + 0.0034924233332276344, + 0.003902899567037821, + 0.0013534992467612028, + 0.002353855175897479, + 0.01439436711370945, + 0.0011466549476608634, + 0.002529492834582925, + 0.1881060153245926, + 0.054332781583070755, + 0.061585795134305954, + 0.016120603308081627, + 0.03930945694446564, + 0.27967920899391174, + 0.01998206600546837, + 0.05070418864488602, + 0.022778231650590897, + 0.002043373417109251, + 0.03589874133467674, + 0.002505393698811531, + 0.001424952526576817, + 0.016086848452687263, + 0.002097264863550663, + 0.001928240293636918, + 0.0009457449777983129, + 0.0009381560375913978, + 0.0010882749920710921, + 0.000701198645401746, + 0.0013130728621035814, + 0.0008126156171783805, + 0.0009986946824938059, + 0.0007639984833076596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.1842605471611023 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651395328342915 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651395328342915, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651395328342915, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10019618272781372, + 0.03850223869085312, + 0.023708244785666466, + 0.018902191892266273, + 0.039450980722904205, + 0.023515811190009117, + 0.43984100222587585, + 0.2326335906982422, + 0.04723697900772095, + 0.036012839525938034 + ], + "credit_map": [ + 0.0131838945671916, + 0.014016510918736458, + 0.009032956324517727, + 0.009137257002294064, + 0.02019672654569149, + 0.012645930983126163, + 0.012533723376691341, + 0.009449182078242302, + 0.006896945647895336, + 0.005114647559821606, + 0.004882134031504393, + 0.004071800038218498, + 0.0057196952402591705, + 0.0044387453235685825, + 0.0045432280749082565, + 0.00283504044637084, + 0.003098356770351529, + 0.0034469023812562227, + 0.0015854784287512302, + 0.002036496764048934, + 0.00426088273525238, + 0.00430840440094471, + 0.002321813954040408, + 0.0026499079540371895, + 0.0019658643286675215, + 0.003157015424221754, + 0.0033036107197403908, + 0.0030388724990189075, + 0.0024485730100423098, + 0.0018660696223378181, + 0.0018596748122945428, + 0.0012625115923583508, + 0.004405403044074774, + 0.005786714144051075, + 0.005404024384915829, + 0.0034435929264873266, + 0.006716521922498941, + 0.006834644358605146, + 0.003232919843867421, + 0.0036271612625569105, + 0.0029003943782299757, + 0.003335272893309593, + 0.005121478345245123, + 0.002205502474680543, + 0.002412285190075636, + 0.003172161988914013, + 0.001919458620250225, + 0.002449258929118514, + 0.11131545156240463, + 0.03592848777770996, + 0.0425805039703846, + 0.011417283676564693, + 0.022097423672676086, + 0.1825886219739914, + 0.008530765771865845, + 0.02538243494927883, + 0.06166943535208702, + 0.017712987959384918, + 0.020184723660349846, + 0.005119091831147671, + 0.012966901063919067, + 0.09155257791280746, + 0.006601287052035332, + 0.016826586797833443, + 0.016413863748311996, + 0.0022213810589164495, + 0.012480733916163445, + 0.00285837659612298, + 0.0021537039428949356, + 0.005543729290366173, + 0.0024587621446698904, + 0.003106429474428296, + 0.00428741192445159, + 0.005015030037611723, + 0.006142543628811836, + 0.0032879046630114317, + 0.004823523107916117, + 0.0038507916033267975, + 0.00391143886372447, + 0.0046941968612372875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.008651395328342915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00249087275005877 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.00249087275005877, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00249087275005877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06878717988729477, + 0.035238735377788544, + 0.03191395476460457, + 0.04628703370690346, + 0.053199853748083115, + 0.02663360722362995, + 0.18961603939533234, + 0.4148876368999481, + 0.09047364443540573, + 0.04296228289604187 + ], + "credit_map": [ + 0.010126299224793911, + 0.009701347909867764, + 0.008771245367825031, + 0.005991901271045208, + 0.010531899519264698, + 0.008656499907374382, + 0.009110857732594013, + 0.005897127091884613, + 0.006389237474650145, + 0.004718526266515255, + 0.006361976731568575, + 0.0032413932494819164, + 0.004197352100163698, + 0.003921958617866039, + 0.0036444838624447584, + 0.002763804281130433, + 0.0038866675458848476, + 0.0048063513822853565, + 0.0034094101283699274, + 0.003031850093975663, + 0.004904420115053654, + 0.0049502127803862095, + 0.0035380488261580467, + 0.0033869908656924963, + 0.005348960869014263, + 0.009876616299152374, + 0.0045558009296655655, + 0.005535102915018797, + 0.007118972484022379, + 0.005795984994620085, + 0.005570949520915747, + 0.002484643366187811, + 0.006857194472104311, + 0.007162542548030615, + 0.007022462785243988, + 0.005127741955220699, + 0.007522509433329105, + 0.010029774159193039, + 0.004496958572417498, + 0.004980669356882572, + 0.004585335496813059, + 0.0024242254439741373, + 0.00458136573433876, + 0.0015651335706934333, + 0.002397658070549369, + 0.007355273235589266, + 0.0016318353591486812, + 0.0020927798468619585, + 0.04619792848825455, + 0.01632552593946457, + 0.01938476599752903, + 0.005457556806504726, + 0.010625742375850677, + 0.07614065706729889, + 0.004614662379026413, + 0.010869189165532589, + 0.11005163192749023, + 0.03243958577513695, + 0.03669894486665726, + 0.009453612379729748, + 0.02286466211080551, + 0.1632312387228012, + 0.011287938803434372, + 0.028860028833150864, + 0.026583276689052582, + 0.0028630574233829975, + 0.03493370860815048, + 0.004717027302831411, + 0.0035596780944615602, + 0.011584323830902576, + 0.0036116524133831263, + 0.002620916347950697, + 0.004716557450592518, + 0.0035113892517983913, + 0.0041904752142727375, + 0.005421338137239218, + 0.007916951552033424, + 0.004299194552004337, + 0.00719466945156455, + 0.005711707752197981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.00249087275005877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635896533727646 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635896533727646, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635896533727646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04496914520859718, + 0.02259226143360138, + 0.060050416737794876, + 0.027581464499235153, + 0.028540173545479774, + 0.04884064942598343, + 0.41008371114730835, + 0.15486560761928558, + 0.14439773559570312, + 0.058078840374946594 + ], + "credit_map": [ + 0.006997946184128523, + 0.007616726215928793, + 0.005321886390447617, + 0.004204181022942066, + 0.00576437683776021, + 0.00527965510264039, + 0.005880588199943304, + 0.003903784789144993, + 0.003759461920708418, + 0.0027261804789304733, + 0.004397445358335972, + 0.0017341547645628452, + 0.0031421962194144726, + 0.0024638299364596605, + 0.0020797590259462595, + 0.002289232797920704, + 0.007076002657413483, + 0.009513967670500278, + 0.005359039641916752, + 0.005548737943172455, + 0.0075786239467561245, + 0.012502760626375675, + 0.006648830603808165, + 0.005822456441819668, + 0.0024755659978836775, + 0.0061027128249406815, + 0.0036210946273058653, + 0.002995838178321719, + 0.0033243796788156033, + 0.004063054919242859, + 0.002704059239476919, + 0.0022947601974010468, + 0.004468968603760004, + 0.0035112304612994194, + 0.004605684895068407, + 0.002468497259542346, + 0.0037272265180945396, + 0.004394570365548134, + 0.0029912069439888, + 0.0023727877996861935, + 0.0062683504074811935, + 0.007530801929533482, + 0.01148140337318182, + 0.0033822543919086456, + 0.004471885040402412, + 0.00717228464782238, + 0.0031349139753729105, + 0.005398754496127367, + 0.09916052967309952, + 0.03512771800160408, + 0.04268081486225128, + 0.011211706325411797, + 0.024552708491683006, + 0.16206857562065125, + 0.011256380006670952, + 0.02402527816593647, + 0.04065351188182831, + 0.012030268087983131, + 0.013781270012259483, + 0.003554221475496888, + 0.00860693957656622, + 0.06124948710203171, + 0.004159549251198769, + 0.010830353014171124, + 0.03967549279332161, + 0.003779178485274315, + 0.059434402734041214, + 0.005544449668377638, + 0.0034900344908237457, + 0.024436920881271362, + 0.004103144630789757, + 0.003934108652174473, + 0.007508453447371721, + 0.0049422564916312695, + 0.008875693194568157, + 0.005809057969599962, + 0.009164937771856785, + 0.005828907247632742, + 0.00854572094976902, + 0.007403814699500799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.004635896533727646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in stark, high-contrast black and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young prince with dark wavy shoulder-length hair and a prominent jagged white scar on his left cheekbone visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a high-collared soot-stained tunic and a tattered cape hanging from one shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character standing in the ruins of a stone hall with broken arches and piles of shattered stone, surrounded by swirling smoke?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera use a low-angle medium-long shot that slowly pans to reveal the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493184179067612 + ], + "opposite_logprob_tokens": [ + -2.929931879043579 + ], + "logprob_reward": -0.05493184179067612, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493184179067612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08554390072822571, + 0.08502313494682312, + 0.12247346341609955, + 0.15334928035736084, + 0.07862100750207901, + 0.04296276718378067, + 0.10418274998664856, + 0.09585284441709518, + 0.1814914345741272, + 0.05049940571188927 + ], + "credit_map": [ + 0.007820895873010159, + 0.014864981174468994, + 0.006266805808991194, + 0.008525165729224682, + 0.012297800742089748, + 0.01222473755478859, + 0.014672641642391682, + 0.00887086521834135, + 0.019568877294659615, + 0.012020859867334366, + 0.008267479948699474, + 0.006055921781808138, + 0.011020012199878693, + 0.011140814051032066, + 0.008890680968761444, + 0.008058481849730015, + 0.01190161146223545, + 0.02287554368376732, + 0.011382820084691048, + 0.01192308496683836, + 0.02019808068871498, + 0.0172396432608366, + 0.01505832839757204, + 0.011894354596734047, + 0.012636172585189342, + 0.03965326398611069, + 0.033045507967472076, + 0.013463546521961689, + 0.019544364884495735, + 0.013853510841727257, + 0.014391965232789516, + 0.006760937627404928, + 0.008016171865165234, + 0.009691678918898106, + 0.0100382249802351, + 0.01293682586401701, + 0.009112933650612831, + 0.013294387608766556, + 0.007644005585461855, + 0.007886779494583607, + 0.0057399761863052845, + 0.004528550431132317, + 0.005774228367954493, + 0.0035706365015357733, + 0.006779403425753117, + 0.008884409442543983, + 0.003612163010984659, + 0.0040733967907726765, + 0.024121461436152458, + 0.009322719648480415, + 0.010382812470197678, + 0.0037125691305845976, + 0.006359391380101442, + 0.040926493704319, + 0.003164741676300764, + 0.006192553322762251, + 0.02463969960808754, + 0.007531068287789822, + 0.008286993019282818, + 0.002723667537793517, + 0.005674798507243395, + 0.03629603981971741, + 0.0035694327671080828, + 0.007131142541766167, + 0.03666549548506737, + 0.00776659045368433, + 0.06331146508455276, + 0.011055690236389637, + 0.01157292164862156, + 0.03372936323285103, + 0.009903263300657272, + 0.007486650720238686, + 0.006775483023375273, + 0.007768871262669563, + 0.00975992064923048, + 0.0040775784291327, + 0.0052834427915513515, + 0.006153113208711147, + 0.005166586022824049, + 0.005514409858733416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.05493184179067612 + } + ], + "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/0000158/group_3/index_tensor([3963, 3963])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_tensor([3963, 3963])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..23bd9ff1acf687d0fc6908973b8fef08f77e37f7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_tensor([3963, 3963])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, right-moving wide shot sweeps across the towering, dark wood shelves of a magnificent library. Thousands of heavy books are neatly arranged on the shelves, forming a long, narrow aisle. A soft, dusty beam of sunlight slants in from a high window just outside the frame, casting a clear path through the air and glazing this ocean of knowledge with a warm, golden glow. As the camera pans, a slender, white hand slowly enters the frame. Its fingertips gently trace the spines of several books, as if tactilely reading their history, finally resting on one. It's a hardcover book with a deep red leather cover, its delicate, gilded, vintage pattern gleaming subtly in the light. Then, firmly grasping the top of the book's spine with its index finger and thumb, the hand carefully yet firmly pulls it free from the tightly packed stack. The book makes a barely audible sound as it slides out, leaving a noticeable gap in the shelf. The whole scene, under the warm light and shadow, creates a tranquil, solemn and exploratory academic atmosphere.", + "seed": 31395, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31395_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": -0.1082874158397317, + "vlm_logprob": -0.1082874158397317, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.1082874158397317, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012256543152034283 + ], + "opposite_logprob_tokens": [ + -5.637256622314453 + ], + "logprob_reward": -0.012256543152034283, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012256543152034283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18826477229595184, + 0.11463730782270432, + 0.058373481035232544, + 0.1336958259344101, + 0.10475703328847885, + 0.07284312695264816, + 0.11176583915948868, + 0.07131443917751312, + 0.053960658609867096, + 0.09038756787776947 + ], + "credit_map": [ + 0.021402843296527863, + 0.027130665257573128, + 0.02463652193546295, + 0.028616691008210182, + 0.014145610854029655, + 0.013213966973125935, + 0.03671451285481453, + 0.02240394987165928, + 0.012282210402190685, + 0.020737970247864723, + 0.012726341374218464, + 0.02928442321717739, + 0.00685609458014369, + 0.0068568573333323, + 0.010674729943275452, + 0.01521867886185646, + 0.005308008287101984, + 0.005295579321682453, + 0.009701043367385864, + 0.009466328658163548, + 0.0051796892657876015, + 0.006751210428774357, + 0.008198640309274197, + 0.00847297441214323, + 0.011137311346828938, + 0.009777415543794632, + 0.01644890196621418, + 0.019201509654521942, + 0.01489954348653555, + 0.028047043830156326, + 0.012610428035259247, + 0.021573666483163834, + 0.009632214903831482, + 0.008809992112219334, + 0.014433846808969975, + 0.010374611243605614, + 0.02034367248415947, + 0.013432277366518974, + 0.01565999537706375, + 0.01207041647285223, + 0.009358441457152367, + 0.00888937059789896, + 0.012865760363638401, + 0.006463301833719015, + 0.012447886168956757, + 0.007927680388092995, + 0.0063420734368264675, + 0.008548608981072903, + 0.01779855228960514, + 0.013859132304787636, + 0.010507858358323574, + 0.007486448157578707, + 0.013119000941514969, + 0.018387066200375557, + 0.012827125377953053, + 0.01778065413236618, + 0.006531530525535345, + 0.007588732521981001, + 0.006961412262171507, + 0.007052486762404442, + 0.013768029399216175, + 0.010957235470414162, + 0.0091904541477561, + 0.00926455482840538, + 0.0070588719099760056, + 0.005352731328457594, + 0.004861431196331978, + 0.006738527212291956, + 0.0053504654206335545, + 0.009562867693603039, + 0.005444165784865618, + 0.009591598995029926, + 0.009024124592542648, + 0.0056105200201272964, + 0.00687890499830246, + 0.0084676593542099, + 0.011684532277286053, + 0.02420760691165924, + 0.008259033784270287, + 0.016255179420113564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012256543152034283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330775558948517 + ], + "opposite_logprob_tokens": [ + -3.7733078002929688 + ], + "logprob_reward": -0.02330775558948517, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330775558948517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14275027811527252, + 0.2781306803226471, + 0.07374787330627441, + 0.0790395736694336, + 0.14152975380420685, + 0.057293493300676346, + 0.10217601805925369, + 0.03278464451432228, + 0.029459478333592415, + 0.06308828294277191 + ], + "credit_map": [ + 0.017523836344480515, + 0.01844342052936554, + 0.018975308164954185, + 0.012542420998215675, + 0.012329385615885258, + 0.02146957255899906, + 0.020497523248195648, + 0.020968792960047722, + 0.03386369347572327, + 0.056281350553035736, + 0.02668061852455139, + 0.06447243690490723, + 0.014955470338463783, + 0.015644285827875137, + 0.029768019914627075, + 0.03646476939320564, + 0.009344000369310379, + 0.005207212176173925, + 0.012663440778851509, + 0.011415351182222366, + 0.0047895582392811775, + 0.009617641568183899, + 0.009260679595172405, + 0.01144998800009489, + 0.005724187940359116, + 0.006691369228065014, + 0.010246388614177704, + 0.010227621532976627, + 0.01047708373516798, + 0.014850311912596226, + 0.007562704849988222, + 0.013259903527796268, + 0.013162101618945599, + 0.011384217068552971, + 0.016655737534165382, + 0.012067764066159725, + 0.02584393136203289, + 0.020173275843262672, + 0.021906830370426178, + 0.020335901528596878, + 0.008163345977663994, + 0.005528551060706377, + 0.007341708522289991, + 0.004486437421292067, + 0.009637446142733097, + 0.006541126873344183, + 0.005157394800335169, + 0.010437481105327606, + 0.018078412860631943, + 0.013223781250417233, + 0.0176099706441164, + 0.0063307699747383595, + 0.00983487255871296, + 0.01390443928539753, + 0.009827856905758381, + 0.013365911319851875, + 0.003069949569180608, + 0.004327600356191397, + 0.002713401336222887, + 0.002521619200706482, + 0.006943314801901579, + 0.004386447370052338, + 0.0036312094889581203, + 0.005191100295633078, + 0.0043646544218063354, + 0.003081230679526925, + 0.003769387025386095, + 0.003162233391776681, + 0.00359692657366395, + 0.0046712253242731094, + 0.0034167850390076637, + 0.00339703680947423, + 0.003943333867937326, + 0.005252505652606487, + 0.004851818084716797, + 0.005995160900056362, + 0.0071208057925105095, + 0.01378592662513256, + 0.00814916007220745, + 0.013989569619297981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.02330775558948517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254032090306282 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254032090306282, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254032090306282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06796319782733917, + 0.043753188103437424, + 0.06363625824451447, + 0.07371419668197632, + 0.037292808294296265, + 0.10013975203037262, + 0.3954413831233978, + 0.07911629974842072, + 0.056906558573246, + 0.08203637599945068 + ], + "credit_map": [ + 0.006364620756357908, + 0.007457176223397255, + 0.006589552853256464, + 0.006717316806316376, + 0.00614698464050889, + 0.00878970418125391, + 0.015527501702308655, + 0.010370333679020405, + 0.006185152567923069, + 0.007717652712017298, + 0.004299876280128956, + 0.005621382035315037, + 0.004600192420184612, + 0.0037207475397735834, + 0.00630038371309638, + 0.005307796411216259, + 0.006074759643524885, + 0.005828098859637976, + 0.009284141473472118, + 0.011633028276264668, + 0.004596652463078499, + 0.008749788627028465, + 0.0074650561437010765, + 0.010004731826484203, + 0.006513850297778845, + 0.005567851942032576, + 0.00916743278503418, + 0.010042263194918633, + 0.008527792058885098, + 0.017262179404497147, + 0.005950062070041895, + 0.010682760737836361, + 0.004002690780907869, + 0.003989384975284338, + 0.0063432916067540646, + 0.004124064929783344, + 0.0064755831845104694, + 0.004905918147414923, + 0.0030225820373743773, + 0.004429290071129799, + 0.015935365110635757, + 0.00833996944129467, + 0.01294554490596056, + 0.0073597184382379055, + 0.009340649470686913, + 0.014581907540559769, + 0.005230297334492207, + 0.02640629932284355, + 0.08743471652269363, + 0.040827471762895584, + 0.06254912912845612, + 0.030991531908512115, + 0.038172926753759384, + 0.050111718475818634, + 0.038165733218193054, + 0.047188129276037216, + 0.006469774525612593, + 0.009349781088531017, + 0.006074805278331041, + 0.005379965994507074, + 0.010484733618795872, + 0.016393419355154037, + 0.0075235567055642605, + 0.01744025945663452, + 0.0067444429732859135, + 0.007304849568754435, + 0.007818207144737244, + 0.006535140797495842, + 0.00643046572804451, + 0.008970093913376331, + 0.005819004960358143, + 0.007284353021532297, + 0.006359864957630634, + 0.0049955821596086025, + 0.007057197857648134, + 0.009612062945961952, + 0.009608243592083454, + 0.01701292395591736, + 0.011632212437689304, + 0.015758275985717773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.005254032090306282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.05372923985123634, + 0.01765800267457962, + 0.03854024410247803, + 0.057564664632081985, + 0.03148540109395981, + 0.041860587894916534, + 0.16334353387355804, + 0.058521416038274765, + 0.09932274371385574, + 0.43797412514686584 + ], + "credit_map": [ + 0.004477865993976593, + 0.005109110847115517, + 0.006942586041986942, + 0.006659676320850849, + 0.003950636833906174, + 0.007257620804011822, + 0.010668457485735416, + 0.008663286454975605, + 0.0014297382440418005, + 0.002207608427852392, + 0.0025998663622885942, + 0.004395889118313789, + 0.0012901301961392164, + 0.0012606051750481129, + 0.0020428220741450787, + 0.002431343775242567, + 0.0034352485090494156, + 0.0031109440606087446, + 0.005249619018286467, + 0.005975813139230013, + 0.0035023505333811045, + 0.005672520026564598, + 0.006032078992575407, + 0.0055616749450564384, + 0.0060914927162230015, + 0.005285234190523624, + 0.007902518846094608, + 0.007494621444493532, + 0.005822719540446997, + 0.00924342405050993, + 0.006829320918768644, + 0.008895336650311947, + 0.003473417367786169, + 0.002470901468768716, + 0.005376943852752447, + 0.0033290742430835962, + 0.006194001529365778, + 0.0041347527876496315, + 0.002650616690516472, + 0.0038556947838515043, + 0.005286712199449539, + 0.0044108182191848755, + 0.004872591234743595, + 0.0032397257164120674, + 0.0056199925020337105, + 0.005929425824433565, + 0.0037963702343404293, + 0.008704953826963902, + 0.03725253418087959, + 0.01599321886897087, + 0.02576165832579136, + 0.012662069872021675, + 0.016774261370301247, + 0.01961224339902401, + 0.015188724733889103, + 0.020098844543099403, + 0.0066789318807423115, + 0.006020443979650736, + 0.0068086921237409115, + 0.003998962696641684, + 0.011930156499147415, + 0.007425821851938963, + 0.006695833057165146, + 0.00896257720887661, + 0.013578912243247032, + 0.010172752663493156, + 0.013389399275183678, + 0.007557660806924105, + 0.009413955733180046, + 0.021592257544398308, + 0.011705518700182438, + 0.011912294663488865, + 0.02761603519320488, + 0.03669067472219467, + 0.038700997829437256, + 0.04619605839252472, + 0.04795297607779503, + 0.08141864836215973, + 0.05747874453663826, + 0.10192001610994339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "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\": \"Is the scene a library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47426825761795044 + ], + "opposite_logprob_tokens": [ + -0.9742682576179504 + ], + "logprob_reward": -0.47426825761795044, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47426825761795044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11873801797628403, + 0.07927100360393524, + 0.18457067012786865, + 0.09851767122745514, + 0.06558043509721756, + 0.046704426407814026, + 0.30065175890922546, + 0.04055696353316307, + 0.02906552143394947, + 0.036343544721603394 + ], + "credit_map": [ + 0.009767778217792511, + 0.01522378996014595, + 0.011081342585384846, + 0.014959183521568775, + 0.012722169980406761, + 0.01665482483804226, + 0.01887102611362934, + 0.019457900896668434, + 0.008696837350726128, + 0.012794418260455132, + 0.007807800080627203, + 0.012528100050985813, + 0.007988045923411846, + 0.009351430460810661, + 0.010299194604158401, + 0.00980517640709877, + 0.01833711937069893, + 0.016032177954912186, + 0.02995017170906067, + 0.03203464671969414, + 0.012305179610848427, + 0.02382311411201954, + 0.021298160776495934, + 0.030790109187364578, + 0.00931348092854023, + 0.008951886557042599, + 0.011605279520154, + 0.012119140475988388, + 0.011274133808910847, + 0.01958637498319149, + 0.011976159177720547, + 0.013691218569874763, + 0.006497959606349468, + 0.006677979603409767, + 0.00962458923459053, + 0.006597929168492556, + 0.010185984894633293, + 0.007410967256873846, + 0.0109096709638834, + 0.007675348315387964, + 0.004460838157683611, + 0.006326436065137386, + 0.006614028476178646, + 0.004493779968470335, + 0.009291030466556549, + 0.006743613630533218, + 0.0031612745951861143, + 0.005613422486931086, + 0.06983254104852676, + 0.032220471650362015, + 0.04996205493807793, + 0.021101364865899086, + 0.03107471577823162, + 0.03363103047013283, + 0.027103878557682037, + 0.0357256643474102, + 0.004314524121582508, + 0.005575948860496283, + 0.004303052555769682, + 0.0021869027987122536, + 0.005729558411985636, + 0.005332226399332285, + 0.003784310072660446, + 0.009330440312623978, + 0.0036484617739915848, + 0.0032700502779334784, + 0.002580150030553341, + 0.002109966240823269, + 0.0037383127491921186, + 0.0060768104158341885, + 0.003620699280872941, + 0.004021069500595331, + 0.003484539920464158, + 0.0025226499419659376, + 0.003053503343835473, + 0.003354561747983098, + 0.003706492017954588, + 0.0064041088335216045, + 0.00511509133502841, + 0.008702601306140423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.47426825761795044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-moving wide shot sweeps across the towering, dark wood shelves of a magnificent library. Thousands of heavy books are neatly arranged on the shelves, forming a long, narrow aisle. A soft, dusty beam of sunlight slants in from a high window just outside the frame, casting a clear path through the air and glazing this ocean of knowledge with a warm, golden glow. As the camera pans, a slender, white hand slowly enters the frame. Its fingertips gently trace the spines of several books, as if tactilely reading their history, finally resting on one. It's a hardcover book with a deep red leather cover, its delicate, gilded, vintage pattern gleaming subtly in the light. Then, firmly grasping the top of the book's spine with its index finger and thumb, the hand carefully yet firmly pulls it free from the tightly packed stack. The book makes a barely audible sound as it slides out, leaving a noticeable gap in the shelf. The whole scene, under the warm light and shadow, creates a tranquil, solemn and exploratory academic atmosphere.", + "seed": 31396, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31396_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": -0.15801143422722816, + "vlm_logprob": -0.15801143422722816, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.15801143422722816, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028576377779245377 + ], + "opposite_logprob_tokens": [ + -4.153576374053955 + ], + "logprob_reward": -0.028576377779245377, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028576377779245377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16339229047298431, + 0.0364961102604866, + 0.06255653500556946, + 0.11959263682365417, + 0.0680144801735878, + 0.03722783923149109, + 0.06941904872655869, + 0.06587904691696167, + 0.3429971933364868, + 0.034424737095832825 + ], + "credit_map": [ + 0.014599255286157131, + 0.011119245551526546, + 0.05675233155488968, + 0.013070318847894669, + 0.010260429233312607, + 0.018170129507780075, + 0.015678049996495247, + 0.023742543533444405, + 0.0027210449334234, + 0.003411589190363884, + 0.008980287238955498, + 0.004048469476401806, + 0.003630467224866152, + 0.003987229894846678, + 0.004861125256866217, + 0.00485590286552906, + 0.003461298765614629, + 0.0038525161799043417, + 0.007590149529278278, + 0.003376156557351351, + 0.008661575615406036, + 0.007615610957145691, + 0.013399283401668072, + 0.014599950984120369, + 0.01339728944003582, + 0.005407408345490694, + 0.018659137189388275, + 0.007062489632517099, + 0.008966085501015186, + 0.02813226543366909, + 0.006066405214369297, + 0.03190156817436218, + 0.006212940905243158, + 0.008128666318953037, + 0.00878957286477089, + 0.006088936701416969, + 0.011257692240178585, + 0.008807539008557796, + 0.010997149161994457, + 0.007731994614005089, + 0.004226131364703178, + 0.006458103656768799, + 0.0067961737513542175, + 0.005687144584953785, + 0.003931623417884111, + 0.004335692618042231, + 0.0029871007427573204, + 0.0028058721218258142, + 0.006454955320805311, + 0.023143356665968895, + 0.003991162870079279, + 0.005621027667075396, + 0.004092753864824772, + 0.006190691143274307, + 0.00428552832454443, + 0.01563958078622818, + 0.005024034064263105, + 0.00402191374450922, + 0.006438487209379673, + 0.0041443416848778725, + 0.007003873120993376, + 0.02136773243546486, + 0.007402882445603609, + 0.01047579012811184, + 0.020819347351789474, + 0.03281247243285179, + 0.015875788405537605, + 0.020395707339048386, + 0.12040045857429504, + 0.014619001187384129, + 0.09657815098762512, + 0.02149631455540657, + 0.005484956782311201, + 0.0044374470598995686, + 0.0069048102013766766, + 0.005023368634283543, + 0.002753369277343154, + 0.0034550046548247337, + 0.004030153620988131, + 0.002335629193112254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.028576377779245377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023329418152570724 + ], + "opposite_logprob_tokens": [ + -3.773329496383667 + ], + "logprob_reward": -0.023329418152570724, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023329418152570724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06365188211202621, + 0.02635917067527771, + 0.042750075459480286, + 0.5344278812408447, + 0.021244149655103683, + 0.018620392307639122, + 0.06408386677503586, + 0.045647285878658295, + 0.1685529351234436, + 0.01466232817620039 + ], + "credit_map": [ + 0.008232115767896175, + 0.005230248905718327, + 0.014456549659371376, + 0.004362624138593674, + 0.004493716172873974, + 0.007489435840398073, + 0.007425033021718264, + 0.011962157674133778, + 0.0021936430130153894, + 0.0018747143913060427, + 0.008626693859696388, + 0.00203778687864542, + 0.0018059207359328866, + 0.002295086393132806, + 0.004270764533430338, + 0.00325456028804183, + 0.002949054818600416, + 0.002092896029353142, + 0.007854681462049484, + 0.0022809330839663744, + 0.0032041878439486027, + 0.004980991594493389, + 0.008452830836176872, + 0.010934502817690372, + 0.0632733628153801, + 0.02399849146604538, + 0.08619911223649979, + 0.030249224975705147, + 0.028316065669059753, + 0.12811996042728424, + 0.026857543736696243, + 0.14741414785385132, + 0.0016489489935338497, + 0.0027070885989814997, + 0.0025082756765186787, + 0.0024724516551941633, + 0.0025110808201134205, + 0.004157524090260267, + 0.002077085431665182, + 0.003161695087328553, + 0.002525553572922945, + 0.0035276957787573338, + 0.0017983817961066961, + 0.0024471809156239033, + 0.0021395201329141855, + 0.0023986578453332186, + 0.0016150454757735133, + 0.0021683559753000736, + 0.006116100586950779, + 0.02006298117339611, + 0.004821256268769503, + 0.0060157859697937965, + 0.00585379870608449, + 0.0055232266895473, + 0.006738108582794666, + 0.008952607400715351, + 0.003993329592049122, + 0.003893145825713873, + 0.005027518607676029, + 0.005631266627460718, + 0.005788083653897047, + 0.005831326358020306, + 0.009173069149255753, + 0.006309544667601585, + 0.009808522649109364, + 0.017324484884738922, + 0.007865793071687222, + 0.011076502501964569, + 0.057122476398944855, + 0.0077506848610937595, + 0.04505591467022896, + 0.01254856027662754, + 0.0025685380678623915, + 0.0014564094599336386, + 0.0015588761307299137, + 0.0017558078980073333, + 0.001312487991526723, + 0.0017605897737666965, + 0.002426763065159321, + 0.0018228556727990508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.023329418152570724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108582317829132 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.01108582317829132, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108582317829132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027484366670250893, + 0.020669687539339066, + 0.03727710247039795, + 0.033997148275375366, + 0.09686186164617538, + 0.07842887938022614, + 0.06811691075563431, + 0.2320493906736374, + 0.3849385976791382, + 0.02017604187130928 + ], + "credit_map": [ + 0.003180776722729206, + 0.003150511533021927, + 0.006430839654058218, + 0.0029496424831449986, + 0.0024351603351533413, + 0.002621573628857732, + 0.003247597021982074, + 0.003468265989795327, + 0.0016803436446934938, + 0.0015165958320721984, + 0.003501146798953414, + 0.0021060428116470575, + 0.001677067717537284, + 0.004679230973124504, + 0.0024803781416267157, + 0.00302888173609972, + 0.002289516618475318, + 0.0033305040560662746, + 0.00531216524541378, + 0.0030818418599665165, + 0.006544790230691433, + 0.0052495477721095085, + 0.005977848544716835, + 0.005490885116159916, + 0.005027275066822767, + 0.0024194347206503153, + 0.006468614563345909, + 0.0024613277055323124, + 0.004221160896122456, + 0.005120321176946163, + 0.002592536387965083, + 0.005686477292329073, + 0.008239594288170338, + 0.008272103033959866, + 0.015002247877418995, + 0.00953531265258789, + 0.017047561705112457, + 0.01296067051589489, + 0.014544560573995113, + 0.01125981193035841, + 0.007153019774705172, + 0.006895667873322964, + 0.020888857543468475, + 0.018645554780960083, + 0.006482667289674282, + 0.008382447995245457, + 0.0046409862115979195, + 0.0053396825678646564, + 0.007578165736049414, + 0.019065413624048233, + 0.00482313334941864, + 0.005599446129053831, + 0.005566966719925404, + 0.005430537275969982, + 0.006313205696642399, + 0.013740039430558681, + 0.014817319810390472, + 0.014217565767467022, + 0.016037162393331528, + 0.0150247011333704, + 0.02470659650862217, + 0.07897821068763733, + 0.029454315081238747, + 0.03881349787116051, + 0.023424215614795685, + 0.03850068151950836, + 0.0181528739631176, + 0.027568640187382698, + 0.12478185445070267, + 0.02004428207874298, + 0.1008373349905014, + 0.031628698110580444, + 0.003940627444535494, + 0.002461755881085992, + 0.003044493729248643, + 0.0022623741533607244, + 0.0014531349297612906, + 0.002304394030943513, + 0.00250986497849226, + 0.002199397888034582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.01108582317829132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712852746248245 + ], + "opposite_logprob_tokens": [ + -3.408712863922119 + ], + "logprob_reward": -0.033712852746248245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712852746248245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010273780673742294, + 0.0069111427292227745, + 0.018776951357722282, + 0.031146017834544182, + 0.012056484818458557, + 0.012001914903521538, + 0.030898824334144592, + 0.06555144488811493, + 0.8018624782562256, + 0.010520903393626213 + ], + "credit_map": [ + 0.0010961133521050215, + 0.000936764816287905, + 0.0018485273467376828, + 0.0008172490634024143, + 0.0010238997638225555, + 0.0012195738963782787, + 0.00147026008926332, + 0.0018613925203680992, + 0.0005576825351454318, + 0.00042496673995628953, + 0.0013183794217184186, + 0.0005358493071980774, + 0.0005459175445139408, + 0.001272027031518519, + 0.0009574503637850285, + 0.0012988693779334426, + 0.0011928174644708633, + 0.001373196835629642, + 0.00232519069686532, + 0.001166796195320785, + 0.0020045568235218525, + 0.0021717974450439215, + 0.003926533740013838, + 0.004616061225533485, + 0.0036887216847389936, + 0.001406494528055191, + 0.003963682800531387, + 0.0016698482213541865, + 0.0026474399492144585, + 0.007403856609016657, + 0.002066669287160039, + 0.008299303241074085, + 0.001465233275666833, + 0.0024742770474404097, + 0.001362029230222106, + 0.0012645428068935871, + 0.0015978753799572587, + 0.0013304039603099227, + 0.001371100079268217, + 0.001191021641716361, + 0.0013396881986409426, + 0.0016357294516637921, + 0.0014624809846282005, + 0.0022245794534683228, + 0.0012788198655471206, + 0.0018233907176181674, + 0.0010172916809096932, + 0.0012199326884001493, + 0.0030449391342699528, + 0.009238282218575478, + 0.0016983204986900091, + 0.0028448132798075676, + 0.001798496930859983, + 0.0027468036860227585, + 0.0023542498238384724, + 0.007172916084527969, + 0.004325066693127155, + 0.003673614701256156, + 0.0042252144776284695, + 0.004594564903527498, + 0.006089660804718733, + 0.023397112265229225, + 0.007441982626914978, + 0.01180422306060791, + 0.04558517783880234, + 0.07589112222194672, + 0.035045236349105835, + 0.053838811814785004, + 0.26318833231925964, + 0.04187728092074394, + 0.22084099054336548, + 0.06559552252292633, + 0.0022865182254463434, + 0.0013208499876782298, + 0.001981992507353425, + 0.0014506533043459058, + 0.0008225002093240619, + 0.0007743785972706974, + 0.001020920230075717, + 0.0008630889933556318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.033712852746248245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6933526992797852 + ], + "opposite_logprob_tokens": [ + -0.6933526992797852 + ], + "logprob_reward": -0.6933526992797852, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933526992797852, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09324655681848526, + 0.08362116664648056, + 0.10087396949529648, + 0.15013034641742706, + 0.0617651604115963, + 0.07439808547496796, + 0.0975140854716301, + 0.10957850515842438, + 0.17819447815418243, + 0.05067765712738037 + ], + "credit_map": [ + 0.009521771222352982, + 0.014792036265134811, + 0.02424539439380169, + 0.0073735979385674, + 0.0088806739076972, + 0.010664653964340687, + 0.009244538843631744, + 0.00852388795465231, + 0.006129767745733261, + 0.007622140925377607, + 0.01904456503689289, + 0.006420290097594261, + 0.009291982278227806, + 0.01151119265705347, + 0.012753556482493877, + 0.010847672820091248, + 0.0064393142238259315, + 0.006726503372192383, + 0.008497083559632301, + 0.00543363718315959, + 0.009448600932955742, + 0.012056665495038033, + 0.023904671892523766, + 0.02836749702692032, + 0.014092452824115753, + 0.008406401611864567, + 0.017085840925574303, + 0.010843543335795403, + 0.012725233100354671, + 0.03658691793680191, + 0.009014986455440521, + 0.04137495905160904, + 0.0046073379926383495, + 0.005797708407044411, + 0.006121272686868906, + 0.004735318478196859, + 0.011679071933031082, + 0.015302694402635098, + 0.004686072003096342, + 0.008835683576762676, + 0.010074048303067684, + 0.015221922658383846, + 0.010584529489278793, + 0.011273331008851528, + 0.007393053732812405, + 0.006748551037162542, + 0.00575589295476675, + 0.0073467642068862915, + 0.011908168904483318, + 0.022775141522288322, + 0.006465302314609289, + 0.008704430423676968, + 0.009796082973480225, + 0.010423767380416393, + 0.011445063166320324, + 0.01599612832069397, + 0.005895658396184444, + 0.00727655366063118, + 0.004542176146060228, + 0.008145620115101337, + 0.013545169495046139, + 0.036490678787231445, + 0.01483082678169012, + 0.01885182037949562, + 0.012369079515337944, + 0.02269754931330681, + 0.011312844231724739, + 0.021428385749459267, + 0.03988564759492874, + 0.0192404855042696, + 0.033029962331056595, + 0.018230529502034187, + 0.004605228081345558, + 0.0040159462951123714, + 0.0047957939095795155, + 0.004128959029912949, + 0.00740090711042285, + 0.010813227854669094, + 0.007321660406887531, + 0.007595936302095652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.6933526992797852 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, right-moving wide shot sweeps across the towering, dark wood shelves of a magnificent library. Thousands of heavy books are neatly arranged on the shelves, forming a long, narrow aisle. A soft, dusty beam of sunlight slants in from a high window just outside the frame, casting a clear path through the air and glazing this ocean of knowledge with a warm, golden glow. As the camera pans, a slender, white hand slowly enters the frame. Its fingertips gently trace the spines of several books, as if tactilely reading their history, finally resting on one. It's a hardcover book with a deep red leather cover, its delicate, gilded, vintage pattern gleaming subtly in the light. Then, firmly grasping the top of the book's spine with its index finger and thumb, the hand carefully yet firmly pulls it free from the tightly packed stack. The book makes a barely audible sound as it slides out, leaving a noticeable gap in the shelf. The whole scene, under the warm light and shadow, creates a tranquil, solemn and exploratory academic atmosphere.", + "seed": 31397, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31397_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": -0.08367023314349353, + "vlm_logprob": -0.08367023314349353, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.08367023314349353, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030331484973430634 + ], + "opposite_logprob_tokens": [ + -3.9053313732147217 + ], + "logprob_reward": -0.030331484973430634, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030331484973430634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4766536056995392, + 0.09515555202960968, + 0.04320053383708, + 0.0485866405069828, + 0.11257614195346832, + 0.03152071312069893, + 0.024564022198319435, + 0.07662976533174515, + 0.06767185032367706, + 0.023441271856427193 + ], + "credit_map": [ + 0.03847489506006241, + 0.08347658812999725, + 0.09140752255916595, + 0.0728822648525238, + 0.031427688896656036, + 0.020872248336672783, + 0.08763319253921509, + 0.05047914758324623, + 0.00677399942651391, + 0.004713049158453941, + 0.029891816899180412, + 0.015720712020993233, + 0.0060011958703398705, + 0.004364953842014074, + 0.01793317310512066, + 0.009756642393767834, + 0.003596992464736104, + 0.003176264464855194, + 0.005955340340733528, + 0.008083581924438477, + 0.00328516261652112, + 0.0024557120632380247, + 0.00900822039693594, + 0.007639257237315178, + 0.00866873562335968, + 0.004157237242907286, + 0.006460541393607855, + 0.007411189377307892, + 0.0022494622971862555, + 0.00319707952439785, + 0.011331157758831978, + 0.00511123426258564, + 0.01638968475162983, + 0.01124971266835928, + 0.01830344833433628, + 0.025232642889022827, + 0.006641815882176161, + 0.011252464726567268, + 0.009809385985136032, + 0.013696970418095589, + 0.0024093736428767443, + 0.002298681065440178, + 0.00632471265271306, + 0.004973784554749727, + 0.004581940360367298, + 0.002285099821165204, + 0.003175549441948533, + 0.00547156622633338, + 0.0016907159006223083, + 0.002170862630009651, + 0.003407413372769952, + 0.0033470981288701296, + 0.0024041682481765747, + 0.004427185282111168, + 0.003205434652045369, + 0.003911140840500593, + 0.007580948993563652, + 0.010595232248306274, + 0.009994672611355782, + 0.012549186125397682, + 0.007225495763123035, + 0.011503922753036022, + 0.007815849035978317, + 0.009364446625113487, + 0.005015687085688114, + 0.011880043894052505, + 0.007739585358649492, + 0.004847413394600153, + 0.006715104915201664, + 0.012109540402889252, + 0.007631623651832342, + 0.01173284463584423, + 0.0019429309759289026, + 0.004561941605061293, + 0.002327174646779895, + 0.0025100819766521454, + 0.0026877890340983868, + 0.003073163330554962, + 0.003791602561250329, + 0.0025465835351496935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.030331484973430634 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026389503851532936 + ], + "opposite_logprob_tokens": [ + -3.6513895988464355 + ], + "logprob_reward": -0.026389503851532936, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026389503851532936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40000271797180176, + 0.05821551755070686, + 0.02207668125629425, + 0.10009294748306274, + 0.0820869579911232, + 0.03526744619011879, + 0.0159438569098711, + 0.11887011677026749, + 0.10947511345148087, + 0.057968638837337494 + ], + "credit_map": [ + 0.03491796925663948, + 0.06556278467178345, + 0.07750152796506882, + 0.06368225067853928, + 0.021511128172278404, + 0.013795389793813229, + 0.0755637064576149, + 0.04746796190738678, + 0.004954913165420294, + 0.005725829396396875, + 0.01177666150033474, + 0.01158883422613144, + 0.0046373694203794, + 0.0033459265250712633, + 0.009595165029168129, + 0.006590819917619228, + 0.002013163175433874, + 0.002042677951976657, + 0.0034578079357743263, + 0.0036629540845751762, + 0.001359280664473772, + 0.0015732907922938466, + 0.005071546416729689, + 0.0028959610499441624, + 0.01949181593954563, + 0.008274238556623459, + 0.014909704215824604, + 0.014714094810187817, + 0.005175009835511446, + 0.0052747540175914764, + 0.021152686327695847, + 0.0111006498336792, + 0.008197706192731857, + 0.0056751384399831295, + 0.015146518126130104, + 0.01279357261955738, + 0.004153331741690636, + 0.007442721631377935, + 0.013765649870038033, + 0.0149123165756464, + 0.0026119318790733814, + 0.002986854175105691, + 0.006167186424136162, + 0.008675103075802326, + 0.003316660411655903, + 0.003427852876484394, + 0.0036538171116262674, + 0.00442803930491209, + 0.0013535235775634646, + 0.0015768351731821895, + 0.002598879858851433, + 0.002671332797035575, + 0.0015817235689610243, + 0.0019574027974158525, + 0.002211480401456356, + 0.0019926796667277813, + 0.009367103688418865, + 0.011509635485708714, + 0.011698179878294468, + 0.013287892565131187, + 0.020474232733249664, + 0.013311873190104961, + 0.023986773565411568, + 0.015234420076012611, + 0.007100878283381462, + 0.015011760406196117, + 0.010504061356186867, + 0.0076720681972801685, + 0.016277868300676346, + 0.01793673448264599, + 0.015813665464520454, + 0.019158069044351578, + 0.005569575820118189, + 0.0083558214828372, + 0.010004590265452862, + 0.006532616447657347, + 0.007675892673432827, + 0.004454534966498613, + 0.010122912935912609, + 0.005252692848443985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.026389503851532936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031985098030418158 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031985098030418158, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031985098030418158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4539121389389038, + 0.034794941544532776, + 0.03195391967892647, + 0.04035532847046852, + 0.08140814304351807, + 0.049096643924713135, + 0.12003367394208908, + 0.12314017117023468, + 0.04002455621957779, + 0.02528044767677784 + ], + "credit_map": [ + 0.030074022710323334, + 0.06614125519990921, + 0.11580657958984375, + 0.08554279804229736, + 0.024666063487529755, + 0.01585148461163044, + 0.07463353127241135, + 0.04119642823934555, + 0.004013827536255121, + 0.0034016401041299105, + 0.005366564262658358, + 0.006552823819220066, + 0.002769248792901635, + 0.0026831510476768017, + 0.005687098950147629, + 0.004320587497204542, + 0.0031859150622040033, + 0.0026550039183348417, + 0.007541500963270664, + 0.005020626820623875, + 0.001994834281504154, + 0.0019066246459260583, + 0.004896864295005798, + 0.004752549808472395, + 0.004589007701724768, + 0.0024459154810756445, + 0.005857062991708517, + 0.006167002487927675, + 0.00217102887108922, + 0.003017374547198415, + 0.012215020135045052, + 0.003892916953191161, + 0.0076669929549098015, + 0.006163756363093853, + 0.014113777317106724, + 0.01917591504752636, + 0.00577675923705101, + 0.008064193651080132, + 0.009607181884348392, + 0.010839561931788921, + 0.004427673760801554, + 0.004250594414770603, + 0.00892452709376812, + 0.009736176580190659, + 0.005144807975739241, + 0.0052529689855873585, + 0.004971983376890421, + 0.006387911271303892, + 0.009097479283809662, + 0.00798034481704235, + 0.02497764863073826, + 0.020384958013892174, + 0.008398130536079407, + 0.010382219217717648, + 0.011407441459596157, + 0.027405453845858574, + 0.010321930050849915, + 0.010865500196814537, + 0.011685756966471672, + 0.011409018188714981, + 0.019348446279764175, + 0.012322098948061466, + 0.028672538697719574, + 0.018514879047870636, + 0.0021691909059882164, + 0.004146444145590067, + 0.005278108175843954, + 0.004848298616707325, + 0.007800724357366562, + 0.005482898093760014, + 0.005665189120918512, + 0.004633702803403139, + 0.0028293633367866278, + 0.004797641653567553, + 0.0024962418247014284, + 0.0029615554958581924, + 0.0025716167874634266, + 0.001836657407693565, + 0.004739435855299234, + 0.0030479358974844217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0031985098030418158 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3153914511203766, + 0.04265211522579193, + 0.04743446037173271, + 0.06437359750270844, + 0.04433334618806839, + 0.028945304453372955, + 0.04473951458930969, + 0.12501153349876404, + 0.10182201862335205, + 0.18529662489891052 + ], + "credit_map": [ + 0.025407860055565834, + 0.056931253522634506, + 0.06835562735795975, + 0.04832065477967262, + 0.020445410162210464, + 0.013250721618533134, + 0.05459015071392059, + 0.02808975614607334, + 0.004928287584334612, + 0.003675710177049041, + 0.007625910919159651, + 0.00882770400494337, + 0.0030263580847531557, + 0.002696676179766655, + 0.006299402564764023, + 0.005572064779698849, + 0.0038887441623955965, + 0.0035180365666747093, + 0.015187878161668777, + 0.008316083811223507, + 0.0023489531595259905, + 0.002688579959794879, + 0.007369674742221832, + 0.004116511438041925, + 0.009378264658153057, + 0.0048144301399588585, + 0.008944484405219555, + 0.010118360631167889, + 0.002701680175960064, + 0.0034336326643824577, + 0.017345920205116272, + 0.007636820897459984, + 0.006946246605366468, + 0.0038893495220690966, + 0.007722816430032253, + 0.008025651797652245, + 0.0028830752708017826, + 0.0040204995311796665, + 0.0052261147648096085, + 0.005619591102004051, + 0.0017246388597413898, + 0.0023199410643428564, + 0.003876082133501768, + 0.004214013926684856, + 0.0030723887030035257, + 0.004183182958513498, + 0.004268583841621876, + 0.005286471918225288, + 0.0034056573640555143, + 0.0032462652307003736, + 0.008031778037548065, + 0.007796971593052149, + 0.003075553337112069, + 0.007220983970910311, + 0.004887057933956385, + 0.00707524549216032, + 0.011161628179252148, + 0.014211984351277351, + 0.015066481195390224, + 0.017898695543408394, + 0.014988853596150875, + 0.018851613625884056, + 0.016523266211152077, + 0.016309017315506935, + 0.007291718386113644, + 0.011945577338337898, + 0.012190207839012146, + 0.00944126769900322, + 0.018773430958390236, + 0.014031513594090939, + 0.016062606126070023, + 0.01208570133894682, + 0.017010118812322617, + 0.04512607678771019, + 0.019190974533557892, + 0.021141497418284416, + 0.020666364580392838, + 0.012732816860079765, + 0.03434888273477554, + 0.015079881064593792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486393988132477 + ], + "opposite_logprob_tokens": [ + -1.2236393690109253 + ], + "logprob_reward": -0.3486393988132477, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486393988132477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44352298974990845, + 0.07527060061693192, + 0.06133263558149338, + 0.10190039873123169, + 0.04628881439566612, + 0.03941613435745239, + 0.031355954706668854, + 0.09819687157869339, + 0.04970976337790489, + 0.05300578102469444 + ], + "credit_map": [ + 0.029060538858175278, + 0.06290727108716965, + 0.11353488266468048, + 0.08241743594408035, + 0.023800617083907127, + 0.0158877894282341, + 0.0721549391746521, + 0.043759558349847794, + 0.005952996201813221, + 0.006674964912235737, + 0.02081744559109211, + 0.014376025646924973, + 0.004721107427030802, + 0.005341278854757547, + 0.0102072823792696, + 0.007179493550211191, + 0.0037373278755694628, + 0.005277905147522688, + 0.00954603124409914, + 0.008458186872303486, + 0.005089164711534977, + 0.003685959614813328, + 0.013115842826664448, + 0.01242221612483263, + 0.015499817207455635, + 0.008096236735582352, + 0.012764927931129932, + 0.016698366031050682, + 0.004738889168947935, + 0.007969255559146404, + 0.02391858957707882, + 0.012214325368404388, + 0.005719618406146765, + 0.004302833694964647, + 0.007001269608736038, + 0.008067901246249676, + 0.003350523067638278, + 0.005297871772199869, + 0.005311235319823027, + 0.007237560581415892, + 0.004749807994812727, + 0.004085839726030827, + 0.008333985693752766, + 0.005679889116436243, + 0.004773243796080351, + 0.0037599061615765095, + 0.0036412812769412994, + 0.004392180126160383, + 0.0029586839955300093, + 0.0025287459138780832, + 0.0060443151742219925, + 0.005455541890114546, + 0.002385997911915183, + 0.003404866438359022, + 0.0046416521072387695, + 0.003936149179935455, + 0.008742546662688255, + 0.008170862682163715, + 0.009060739539563656, + 0.008477146737277508, + 0.016180112957954407, + 0.01157462503761053, + 0.021979177370667458, + 0.014011661522090435, + 0.002731613116338849, + 0.006968207657337189, + 0.004587104544043541, + 0.006433334667235613, + 0.007422605995088816, + 0.006433323957026005, + 0.006042334716767073, + 0.009091237559914589, + 0.0051305717788636684, + 0.006618451327085495, + 0.007701776456087828, + 0.0066672940738499165, + 0.0074327727779746056, + 0.0057347482070326805, + 0.006506657227873802, + 0.00721350871026516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.3486393988132477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-moving wide shot sweeps across the towering, dark wood shelves of a magnificent library. Thousands of heavy books are neatly arranged on the shelves, forming a long, narrow aisle. A soft, dusty beam of sunlight slants in from a high window just outside the frame, casting a clear path through the air and glazing this ocean of knowledge with a warm, golden glow. As the camera pans, a slender, white hand slowly enters the frame. Its fingertips gently trace the spines of several books, as if tactilely reading their history, finally resting on one. It's a hardcover book with a deep red leather cover, its delicate, gilded, vintage pattern gleaming subtly in the light. Then, firmly grasping the top of the book's spine with its index finger and thumb, the hand carefully yet firmly pulls it free from the tightly packed stack. The book makes a barely audible sound as it slides out, leaving a noticeable gap in the shelf. The whole scene, under the warm light and shadow, creates a tranquil, solemn and exploratory academic atmosphere.", + "seed": 31398, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_3/index_3963_seed_31398_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": -0.16829989063553513, + "vlm_logprob": -0.16829989063553513, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.16829989063553513, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + -3.4196860790252686 + ], + "logprob_reward": -0.04468613117933273, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04468613117933273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08953163027763367, + 0.03635982424020767, + 0.014191092923283577, + 0.08408085256814957, + 0.03882886469364166, + 0.09907549619674683, + 0.5655714869499207, + 0.023933451622724533, + 0.024434005841612816, + 0.02399330586194992 + ], + "credit_map": [ + 0.007527017500251532, + 0.014408281072974205, + 0.013799999840557575, + 0.004798225127160549, + 0.016292281448841095, + 0.016769500449299812, + 0.007244954816997051, + 0.008691372349858284, + 0.0037347774486988783, + 0.003282556077465415, + 0.0073777236975729465, + 0.0027159652672708035, + 0.006093698088079691, + 0.004559454973787069, + 0.003990362398326397, + 0.00460528489202261, + 0.0020590322092175484, + 0.0018424284644424915, + 0.00315330782905221, + 0.0011726904194802046, + 0.0020293009001761675, + 0.0009882167214527726, + 0.001267539686523378, + 0.001678577158600092, + 0.009821169078350067, + 0.010668842121958733, + 0.01703481376171112, + 0.0073259067721664906, + 0.007685162592679262, + 0.010103329084813595, + 0.011666540056467056, + 0.009775089100003242, + 0.006494452245533466, + 0.005071733612567186, + 0.005182853899896145, + 0.003922360483556986, + 0.004624583292752504, + 0.004530386533588171, + 0.0043526096269488335, + 0.004649882670491934, + 0.012497841380536556, + 0.018749387934803963, + 0.01257339771836996, + 0.009692718274891376, + 0.012654192745685577, + 0.008926913142204285, + 0.015138201415538788, + 0.008842837065458298, + 0.02171330526471138, + 0.0874994769692421, + 0.008291509002447128, + 0.016553930938243866, + 0.015679650008678436, + 0.2475961446762085, + 0.009946552105247974, + 0.15829090774059296, + 0.0025486925151199102, + 0.004582806956022978, + 0.0031219704542309046, + 0.001523429062217474, + 0.002058042911812663, + 0.00514623150229454, + 0.0020609113853424788, + 0.002891368232667446, + 0.0017154439119622111, + 0.0022572074085474014, + 0.0013238588580861688, + 0.0028855788987129927, + 0.004566024988889694, + 0.004998918157070875, + 0.0032574953511357307, + 0.00342947687022388, + 0.002167288213968277, + 0.0036773330066353083, + 0.002203803276643157, + 0.0036366714630275965, + 0.002916431287303567, + 0.0024832331109791994, + 0.003864263417199254, + 0.0030442809220403433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "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 scene a library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026387065649032593 + ], + "opposite_logprob_tokens": [ + -3.6513869762420654 + ], + "logprob_reward": -0.026387065649032593, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026387065649032593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05327250063419342, + 0.4411575198173523, + 0.04626510664820671, + 0.04533965885639191, + 0.06675275415182114, + 0.10903105139732361, + 0.0892900750041008, + 0.0392942801117897, + 0.05990064516663551, + 0.04969637468457222 + ], + "credit_map": [ + 0.005428697448223829, + 0.006981227546930313, + 0.00701015442609787, + 0.003473907010629773, + 0.010829910635948181, + 0.008632125332951546, + 0.005614045076072216, + 0.00530243432149291, + 0.06468617171049118, + 0.021649504080414772, + 0.15317009389400482, + 0.021527644246816635, + 0.07409965991973877, + 0.02806372009217739, + 0.028909817337989807, + 0.049050915986299515, + 0.0065111154690384865, + 0.00609145313501358, + 0.008613990619778633, + 0.00509184505790472, + 0.007474876008927822, + 0.00352464965544641, + 0.004403420723974705, + 0.0045537580735981464, + 0.0061869071796536446, + 0.00572814978659153, + 0.010446320287883282, + 0.004929079674184322, + 0.0044684153981506824, + 0.002882033819332719, + 0.005531252361834049, + 0.00516750430688262, + 0.009650214575231075, + 0.010017468594014645, + 0.013142107054591179, + 0.008090641349554062, + 0.00709321117028594, + 0.00677824579179287, + 0.005058927461504936, + 0.006921944674104452, + 0.012739182449877262, + 0.020451711490750313, + 0.012686667963862419, + 0.01099368929862976, + 0.013467940501868725, + 0.010055165737867355, + 0.01814710721373558, + 0.01048959232866764, + 0.003785584820434451, + 0.013300135731697083, + 0.002255039056763053, + 0.0035628341138362885, + 0.004714725539088249, + 0.03613010421395302, + 0.0031262305565178394, + 0.02241542562842369, + 0.0036492894869297743, + 0.006704641506075859, + 0.0044640181586146355, + 0.0026793782599270344, + 0.0035739359445869923, + 0.00935104675590992, + 0.0037607073318213224, + 0.00511126359924674, + 0.004591286648064852, + 0.003999712411314249, + 0.0026073402259498835, + 0.005449126474559307, + 0.018537349998950958, + 0.009141898714005947, + 0.010214715264737606, + 0.005359217990189791, + 0.008857368491590023, + 0.007438110653311014, + 0.0036961797159165144, + 0.005788084585219622, + 0.006994669325649738, + 0.004793849773705006, + 0.006435002200305462, + 0.0056931134313344955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.026387065649032593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002823535120114684 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.002823535120114684, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002823535120114684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05254120007157326, + 0.05226663872599602, + 0.0379372276365757, + 0.06724120676517487, + 0.057430021464824677, + 0.19403338432312012, + 0.2638849914073944, + 0.030633091926574707, + 0.12127351760864258, + 0.12275871634483337 + ], + "credit_map": [ + 0.003808471141383052, + 0.006119334604591131, + 0.009321329183876514, + 0.0027944212779402733, + 0.007893061265349388, + 0.011446461081504822, + 0.004819807596504688, + 0.00633830763399601, + 0.006944895256310701, + 0.0032619258854538202, + 0.014799864962697029, + 0.002695529954507947, + 0.008621685206890106, + 0.006527318619191647, + 0.003647114150226116, + 0.005768303759396076, + 0.004903306253254414, + 0.00574329262599349, + 0.006012988742440939, + 0.003175589023157954, + 0.006683854386210442, + 0.003130200318992138, + 0.004085452761501074, + 0.00420254236087203, + 0.008759235963225365, + 0.010031075216829777, + 0.014429423958063126, + 0.007350640371441841, + 0.005450894590467215, + 0.003871556604281068, + 0.009394885040819645, + 0.007953490130603313, + 0.007903087884187698, + 0.007003608159720898, + 0.0064099254086613655, + 0.006149393506348133, + 0.006825372111052275, + 0.009165099821984768, + 0.007459744345396757, + 0.006513791158795357, + 0.023830659687519073, + 0.0394718237221241, + 0.018711762502789497, + 0.018517598509788513, + 0.019495880231261253, + 0.017198581248521805, + 0.040442969650030136, + 0.016364121809601784, + 0.011805093847215176, + 0.037798214703798294, + 0.007928614504635334, + 0.009398086927831173, + 0.010650296695530415, + 0.10259751230478287, + 0.017197858542203903, + 0.06650932878255844, + 0.003291968023404479, + 0.004459354095160961, + 0.003186745336279273, + 0.0035779858008027077, + 0.004045762587338686, + 0.0042769224382936954, + 0.004176584538072348, + 0.003617769107222557, + 0.010740636847913265, + 0.01014543604105711, + 0.006211501080542803, + 0.017130229622125626, + 0.03241724893450737, + 0.014435688965022564, + 0.019166434183716774, + 0.011026336811482906, + 0.011136760003864765, + 0.025040894746780396, + 0.009864299558103085, + 0.020166877657175064, + 0.014319638721644878, + 0.014826025813817978, + 0.013939767144620419, + 0.013464451767504215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 9 + ], + "selected_reward": -0.002823535120114684 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788254275918007 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.009788254275918007, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788254275918007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06273355334997177, + 0.040277011692523956, + 0.04320048913359642, + 0.05232976749539375, + 0.045824114233255386, + 0.07817792147397995, + 0.14752207696437836, + 0.02484561875462532, + 0.12916862964630127, + 0.375920832157135 + ], + "credit_map": [ + 0.004285830073058605, + 0.0095719825476408, + 0.007928627543151379, + 0.003439107211306691, + 0.013650729320943356, + 0.012598767876625061, + 0.004933088086545467, + 0.006325422320514917, + 0.005238939542323351, + 0.002055920660495758, + 0.012778250500559807, + 0.001909715705551207, + 0.006429584231227636, + 0.0032398374751210213, + 0.0032080363016575575, + 0.005416726227849722, + 0.004369415808469057, + 0.004796503111720085, + 0.005768547300249338, + 0.002721367636695504, + 0.008545516058802605, + 0.004901491571217775, + 0.005790295545011759, + 0.006307348608970642, + 0.004923691973090172, + 0.009811817668378353, + 0.0059527806006371975, + 0.0032287517096847296, + 0.005546781700104475, + 0.01206846535205841, + 0.0046434770338237286, + 0.006154005415737629, + 0.008101380430161953, + 0.006848821882158518, + 0.005205255933105946, + 0.0044232383370399475, + 0.00450760405510664, + 0.005176729988306761, + 0.0054963803850114346, + 0.006064700894057751, + 0.009256196208298206, + 0.013592049479484558, + 0.008787307888269424, + 0.007443955633789301, + 0.009114395827054977, + 0.009404780343174934, + 0.013529260642826557, + 0.007049978710711002, + 0.0054874843917787075, + 0.02282504364848137, + 0.0023695107083767653, + 0.00430913083255291, + 0.004580841865390539, + 0.0644507110118866, + 0.003046311205253005, + 0.040453046560287476, + 0.0019441295880824327, + 0.0035060199443250895, + 0.0014202224556356668, + 0.004230597987771034, + 0.0031943267676979303, + 0.0032260860316455364, + 0.004407433792948723, + 0.0029168047476559877, + 0.009467259980738163, + 0.00948953628540039, + 0.005400761496275663, + 0.013470533303916454, + 0.042036961764097214, + 0.018350280821323395, + 0.020730141550302505, + 0.01022315677255392, + 0.027105435729026794, + 0.0662786066532135, + 0.023560330271720886, + 0.050449274480342865, + 0.05248362198472023, + 0.04807637631893158, + 0.057514820247888565, + 0.0504523403942585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.009788254275918007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 library with towering, dark wood shelves filled with thousands of books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does a soft, dusty beam of sunlight slant in from a high window outside the frame, casting a warm, golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a slender, white hand enter the frame and trace the spines of books before pulling one out?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pulled book a hardcover with a deep red leather cover and a gilded vintage pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a slow, right-moving wide shot that sweeps across the shelves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.7578144669532776 + ], + "opposite_logprob_tokens": [ + -0.6328144669532776 + ], + "logprob_reward": -0.7578144669532776, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578144669532776, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17721904814243317, + 0.12768100202083588, + 0.05257614701986313, + 0.053802549839019775, + 0.10883186757564545, + 0.07766392081975937, + 0.1821371614933014, + 0.05441085994243622, + 0.07562588155269623, + 0.0900515541434288 + ], + "credit_map": [ + 0.012306700460612774, + 0.03378445282578468, + 0.01938014291226864, + 0.010702771134674549, + 0.03329963609576225, + 0.03387969359755516, + 0.016974562779068947, + 0.01689109392464161, + 0.016813106834888458, + 0.012210813350975513, + 0.02627791278064251, + 0.009968103840947151, + 0.02107900194823742, + 0.013335207477211952, + 0.010669734328985214, + 0.01732712984085083, + 0.006291322410106659, + 0.007379862479865551, + 0.010242462158203125, + 0.0034665202256292105, + 0.007985081523656845, + 0.004194183275103569, + 0.005657303147017956, + 0.007359413430094719, + 0.00479340273886919, + 0.00910720694810152, + 0.007567772641777992, + 0.006307628005743027, + 0.004356774035841227, + 0.007332187611609697, + 0.009065900929272175, + 0.005271673668175936, + 0.015135345980525017, + 0.01680920459330082, + 0.022526629269123077, + 0.013911986723542213, + 0.009905567392706871, + 0.011285610496997833, + 0.008221474476158619, + 0.011036052368581295, + 0.008844757452607155, + 0.013672834262251854, + 0.006989254616200924, + 0.0075057800859212875, + 0.00662268465384841, + 0.01051525678485632, + 0.017410768195986748, + 0.006102587096393108, + 0.0073686158284544945, + 0.031147433444857597, + 0.005330353509634733, + 0.006048840004950762, + 0.010598842054605484, + 0.07233636826276779, + 0.007885631173849106, + 0.0414210744202137, + 0.005524190608412027, + 0.009404927492141724, + 0.006420285906642675, + 0.0059663294814527035, + 0.005698181688785553, + 0.00768635468557477, + 0.0069918022491037846, + 0.006718783639371395, + 0.005384647753089666, + 0.004815786145627499, + 0.003913880325853825, + 0.006717725656926632, + 0.023664817214012146, + 0.01338586863130331, + 0.011268122121691704, + 0.006475036032497883, + 0.007411450147628784, + 0.01392382476478815, + 0.0067534493282437325, + 0.015623250976204872, + 0.012635908089578152, + 0.010367797687649727, + 0.012432336807250977, + 0.010903534479439259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.7578144669532776 + } + ], + "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/0000158/group_4/index_tensor([204, 204])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_tensor([204, 204])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2041742497dcd7bbd810d5eb92a151b0ce701688 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_tensor([204, 204])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up shot in stop-motion style, a pair of weathered yet steady hands manipulate a traditional woodcarving knife, its handle smooth and mellow from years of use. As each frame progresses, the blade strikes with a precise and decisive rhythm, each contact sending tiny sawdust dancing through the dense, solid wood, seemingly taking on a life of its own in the still air. The scene is illuminated by a warm spotlight from above and sideways, creating a striking contrast of light and dark against the dim backdrop of the artisan's workshop, highlighting the delicate grain of the wood and the sharp, metallic quality of the blade. With each leap and vanishing of the sawdust, the Buddha's outline is miraculously shaped: from the smooth forehead to the closed, contemplative eyelids to the subtle hint of a smile at the corners of the mouth, every detail is brought to life in this rhythmic and rhythmic visual language. Finally, a compassionate and serene Buddha statue slowly reveals its complete form amidst the interweaving of light, shadow, and wood grain. The entire process is filled with concentration, tranquility, and an almost sacred sense of creation.", + "seed": 11847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_204_seed_11847_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": -0.13447760976850986, + "vlm_logprob": -0.13447760976850986, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.13447760976850986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024189457297325134 + ], + "opposite_logprob_tokens": [ + -7.024189472198486 + ], + "logprob_reward": -0.024189457297325134, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.024189457297325134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07293474674224854, + 0.06386232376098633, + 0.04918054863810539, + 0.06802385300397873, + 0.06250220537185669, + 0.04407713562250137, + 0.16164742410182953, + 0.31536683440208435, + 0.12451884895563126, + 0.037886038422584534 + ], + "credit_map": [ + 0.010381870903074741, + 0.0095728924497962, + 0.013362595811486244, + 0.006821134127676487, + 0.005824920255690813, + 0.011209316551685333, + 0.007598972413688898, + 0.008163043297827244, + 0.008411106653511524, + 0.007030594628304243, + 0.010768203996121883, + 0.00633522542193532, + 0.005561577621847391, + 0.01035765279084444, + 0.007084363140165806, + 0.008313599973917007, + 0.009935583919286728, + 0.0032558334060013294, + 0.012349957600235939, + 0.0023039630614221096, + 0.003498075995594263, + 0.0067785740830004215, + 0.003222180064767599, + 0.00783638283610344, + 0.005069542210549116, + 0.004516778979450464, + 0.005846971645951271, + 0.005802760366350412, + 0.01322237215936184, + 0.012707006186246872, + 0.01017967239022255, + 0.010678746737539768, + 0.004857991822063923, + 0.0038085924461483955, + 0.0049204519018530846, + 0.004255089443176985, + 0.014461850747466087, + 0.008080913685262203, + 0.009380584582686424, + 0.012736734934151173, + 0.003557451767846942, + 0.0043413820676505566, + 0.00475724833086133, + 0.003390898462384939, + 0.0056731803342700005, + 0.00637283967807889, + 0.01075764000415802, + 0.005226497538387775, + 0.012809033505618572, + 0.016710447147488594, + 0.019791876897215843, + 0.01805899292230606, + 0.023774875327944756, + 0.03256404027342796, + 0.016721492633223534, + 0.021216662600636482, + 0.02156270481646061, + 0.03343414515256882, + 0.028278885409235954, + 0.033918026834726334, + 0.07078154385089874, + 0.052929654717445374, + 0.04076534882187843, + 0.03369652479887009, + 0.01069042831659317, + 0.007620969787240028, + 0.013054830022156239, + 0.010559026151895523, + 0.015676483511924744, + 0.023646758869290352, + 0.020622238516807556, + 0.022648107260465622, + 0.001976656261831522, + 0.0026678917929530144, + 0.0034468218218535185, + 0.0031861194875091314, + 0.004466010257601738, + 0.0069677699357271194, + 0.0067089213989675045, + 0.008465849794447422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.024189457297325134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328456997871399 + ], + "opposite_logprob_tokens": [ + -0.7578456997871399 + ], + "logprob_reward": -0.6328456997871399, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328456997871399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1535899043083191, + 0.03504856303334236, + 0.07186494767665863, + 0.07041578739881516, + 0.17843884229660034, + 0.16841451823711395, + 0.06285415589809418, + 0.14463694393634796, + 0.06735999137163162, + 0.047376345843076706 + ], + "credit_map": [ + 0.015451469458639622, + 0.01795363239943981, + 0.03374088555574417, + 0.015261497348546982, + 0.010546543635427952, + 0.023498548194766045, + 0.01988835074007511, + 0.017248980700969696, + 0.002922272076830268, + 0.002779174828901887, + 0.004480500239878893, + 0.0024160367902368307, + 0.0035126449074596167, + 0.006651769392192364, + 0.006436476483941078, + 0.005849692039191723, + 0.010358811356127262, + 0.0035257921554148197, + 0.01135609857738018, + 0.0026311040855944157, + 0.008627709932625294, + 0.013756567612290382, + 0.009238187223672867, + 0.012370678596198559, + 0.005750959739089012, + 0.0044894651509821415, + 0.008943500928580761, + 0.005423343274742365, + 0.011884190142154694, + 0.013309662230312824, + 0.009657113812863827, + 0.010957551188766956, + 0.013628140091896057, + 0.01128124538809061, + 0.012869198806583881, + 0.012989796698093414, + 0.050381001085042953, + 0.019927868619561195, + 0.017634009942412376, + 0.039727583527565, + 0.011629201471805573, + 0.01526891253888607, + 0.013973447494208813, + 0.012151272967457771, + 0.026072993874549866, + 0.023897334933280945, + 0.03696412965655327, + 0.028457222506403923, + 0.0058983745984733105, + 0.005046100355684757, + 0.008204747922718525, + 0.006182583514600992, + 0.009961970150470734, + 0.011641315184533596, + 0.009413687512278557, + 0.006505383178591728, + 0.009460254572331905, + 0.015585171058773994, + 0.012570375576615334, + 0.014283362776041031, + 0.03298833593726158, + 0.02441459335386753, + 0.020039280876517296, + 0.015295585617423058, + 0.0056993719190359116, + 0.0037181104999035597, + 0.008133778348565102, + 0.0039892615750432014, + 0.008980723097920418, + 0.013077039271593094, + 0.010077858343720436, + 0.013683842495083809, + 0.004741517826914787, + 0.0035417904146015644, + 0.005790533497929573, + 0.004203126300126314, + 0.006943667773157358, + 0.008677544072270393, + 0.00683618476614356, + 0.00664198212325573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.6328456997871399 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104286432266235 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104286432266235, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104286432266235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12355229258537292, + 0.08124768733978271, + 0.07154633104801178, + 0.08081835508346558, + 0.09320822358131409, + 0.19915804266929626, + 0.08841271698474884, + 0.11610322445631027, + 0.08762471377849579, + 0.05832844227552414 + ], + "credit_map": [ + 0.016612987965345383, + 0.01844767853617668, + 0.026291953399777412, + 0.013915430754423141, + 0.008484547957777977, + 0.013944694772362709, + 0.011370579712092876, + 0.014484409242868423, + 0.010433217510581017, + 0.008256612345576286, + 0.010354623198509216, + 0.00565326027572155, + 0.005596363451331854, + 0.017370115965604782, + 0.01192710641771555, + 0.011656383983790874, + 0.011292003095149994, + 0.006419262383133173, + 0.013136080466210842, + 0.0075502642430365086, + 0.005724585149437189, + 0.01001782901585102, + 0.007667757570743561, + 0.009738553315401077, + 0.005078657530248165, + 0.0056833019480109215, + 0.010845662094652653, + 0.007642421405762434, + 0.012203997001051903, + 0.015363983809947968, + 0.012425518594682217, + 0.011574815958738327, + 0.00801989994943142, + 0.0071152932941913605, + 0.014348709024488926, + 0.011838316917419434, + 0.017034880816936493, + 0.012242632918059826, + 0.010587956756353378, + 0.01202053390443325, + 0.011484039016067982, + 0.011916608549654484, + 0.012020516209304333, + 0.010794646106660366, + 0.025581078603863716, + 0.040540121495723724, + 0.054579928517341614, + 0.0322410948574543, + 0.009145407006144524, + 0.006882161367684603, + 0.014150184579193592, + 0.00968118105083704, + 0.011662615463137627, + 0.013808117248117924, + 0.011424644850194454, + 0.011658405885100365, + 0.008246948011219501, + 0.010896964929997921, + 0.009139914065599442, + 0.011622204445302486, + 0.02200653962790966, + 0.02298286370933056, + 0.01614336296916008, + 0.01506443228572607, + 0.009095648303627968, + 0.0053626359440386295, + 0.01085873693227768, + 0.0067196316085755825, + 0.011022898368537426, + 0.017502935603260994, + 0.011157136410474777, + 0.01590508408844471, + 0.0067227864637970924, + 0.0036907466128468513, + 0.0071432325057685375, + 0.004613166209310293, + 0.007333829067647457, + 0.012231902219355106, + 0.00664244219660759, + 0.009950337000191212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.004104286432266235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.04606318473815918, + 0.05707494542002678, + 0.12119346857070923, + 0.03575896471738815, + 0.09067045152187347, + 0.10376792401075363, + 0.09568528831005096, + 0.2150333970785141, + 0.17176885902881622, + 0.06298355758190155 + ], + "credit_map": [ + 0.008322971872985363, + 0.005395148880779743, + 0.0041410429403185844, + 0.005030730739235878, + 0.0044691963121294975, + 0.007271117996424437, + 0.006026135757565498, + 0.005406838841736317, + 0.004931105766445398, + 0.003267651190981269, + 0.007264170795679092, + 0.0037895692512392998, + 0.00538788503035903, + 0.012475748546421528, + 0.010578588582575321, + 0.009380223229527473, + 0.019849028438329697, + 0.008198240771889687, + 0.01932682655751705, + 0.008038756437599659, + 0.009639755822718143, + 0.01867801696062088, + 0.01827545091509819, + 0.019187381491065025, + 0.0035178938414901495, + 0.0031117075122892857, + 0.004448176361620426, + 0.003778590355068445, + 0.005782362073659897, + 0.004693659022450447, + 0.00627899868413806, + 0.00414757477119565, + 0.011739209294319153, + 0.011091294698417187, + 0.014840512536466122, + 0.012669923715293407, + 0.010855546221137047, + 0.011736010201275349, + 0.008964220061898232, + 0.008773737587034702, + 0.004842967260628939, + 0.005189497489482164, + 0.006902631837874651, + 0.0065086157992482185, + 0.014235662296414375, + 0.02118895947933197, + 0.030426163226366043, + 0.014473424293100834, + 0.0043933517299592495, + 0.005629985127598047, + 0.008431521244347095, + 0.010454542934894562, + 0.016169948503375053, + 0.014850017614662647, + 0.02338751032948494, + 0.012368409894406796, + 0.014398885890841484, + 0.02228488214313984, + 0.017244113609194756, + 0.02223338931798935, + 0.049825362861156464, + 0.03925506770610809, + 0.02722257748246193, + 0.02256910875439644, + 0.009823242202401161, + 0.012113452889025211, + 0.01508014090359211, + 0.01445742230862379, + 0.02906752936542034, + 0.034519363194704056, + 0.026354286819696426, + 0.030353402718901634, + 0.006134912837296724, + 0.004495976958423853, + 0.00943068414926529, + 0.006327081937342882, + 0.008484059013426304, + 0.010734946466982365, + 0.007371548097580671, + 0.010004340671002865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.007632143795490265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616461530327797 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.003616461530327797, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616461530327797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405315399169922, + 0.051193322986364365, + 0.08094626665115356, + 0.17217019200325012, + 0.10855968296527863, + 0.13656176626682281, + 0.07786917686462402, + 0.1027672067284584, + 0.05947967246174812, + 0.06992114335298538 + ], + "credit_map": [ + 0.016080684959888458, + 0.01142845768481493, + 0.026861535385251045, + 0.014759678393602371, + 0.009933088906109333, + 0.013030139729380608, + 0.02823856845498085, + 0.02019939012825489, + 0.0058785732835531235, + 0.005161804612725973, + 0.006587312091141939, + 0.004499326925724745, + 0.004803608637303114, + 0.011342927813529968, + 0.005937288980931044, + 0.006982480175793171, + 0.011769059114158154, + 0.005338164046406746, + 0.01892806775867939, + 0.0037921660114079714, + 0.006232276558876038, + 0.012572807259857655, + 0.008584159426391125, + 0.013729567639529705, + 0.008637823164463043, + 0.0065410444512963295, + 0.010046442039310932, + 0.006753172725439072, + 0.01640603132545948, + 0.0574314221739769, + 0.02315627411007881, + 0.04319797456264496, + 0.011388935148715973, + 0.011640415526926517, + 0.013165599666535854, + 0.014956682920455933, + 0.024458499625325203, + 0.015261289663612843, + 0.009393868036568165, + 0.008294395171105862, + 0.010094466619193554, + 0.015751030296087265, + 0.009569807909429073, + 0.01240738295018673, + 0.02264142408967018, + 0.02480764128267765, + 0.016003340482711792, + 0.025286665186285973, + 0.0069930474273860455, + 0.0058534699492156506, + 0.011023099534213543, + 0.010290983133018017, + 0.01103227585554123, + 0.011612613685429096, + 0.010063686408102512, + 0.011000006459653378, + 0.006177665665745735, + 0.010209026746451855, + 0.008418708108365536, + 0.009230641648173332, + 0.023109570145606995, + 0.017199194058775902, + 0.01667371578514576, + 0.011748679913580418, + 0.004782235715538263, + 0.004022547975182533, + 0.005688054952770472, + 0.005124719347804785, + 0.010294313542544842, + 0.013119585812091827, + 0.006064733024686575, + 0.01038348488509655, + 0.006887279916554689, + 0.004846603609621525, + 0.008911795914173126, + 0.006068757269531488, + 0.008849321864545345, + 0.01462361216545105, + 0.008489674888551235, + 0.011244098655879498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.003616461530327797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in stop-motion style, a pair of weathered yet steady hands manipulate a traditional woodcarving knife, its handle smooth and mellow from years of use. As each frame progresses, the blade strikes with a precise and decisive rhythm, each contact sending tiny sawdust dancing through the dense, solid wood, seemingly taking on a life of its own in the still air. The scene is illuminated by a warm spotlight from above and sideways, creating a striking contrast of light and dark against the dim backdrop of the artisan's workshop, highlighting the delicate grain of the wood and the sharp, metallic quality of the blade. With each leap and vanishing of the sawdust, the Buddha's outline is miraculously shaped: from the smooth forehead to the closed, contemplative eyelids to the subtle hint of a smile at the corners of the mouth, every detail is brought to life in this rhythmic and rhythmic visual language. Finally, a compassionate and serene Buddha statue slowly reveals its complete form amidst the interweaving of light, shadow, and wood grain. The entire process is filled with concentration, tranquility, and an almost sacred sense of creation.", + "seed": 11848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_204_seed_11848_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": -0.22519629448652267, + "vlm_logprob": -0.22519629448652267, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.22519629448652267, + "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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03411974385380745 + ], + "opposite_logprob_tokens": [ + -7.784119606018066 + ], + "logprob_reward": -0.03411974385380745, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.03411974385380745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07575046271085739, + 0.03786163404583931, + 0.10030058771371841, + 0.05139768123626709, + 0.1337701678276062, + 0.0885813981294632, + 0.3746054172515869, + 0.06548585742712021, + 0.027586285024881363, + 0.044660598039627075 + ], + "credit_map": [ + 0.008744582533836365, + 0.008925833739340305, + 0.011027759872376919, + 0.015357335098087788, + 0.008363509550690651, + 0.007257426623255014, + 0.005862614139914513, + 0.01021140068769455, + 0.004015731159597635, + 0.0034301667474210262, + 0.005867039784789085, + 0.00639303307980299, + 0.0033617003355175257, + 0.006213850807398558, + 0.00422035763040185, + 0.00435975007712841, + 0.01003198605030775, + 0.010078568942844868, + 0.016499578952789307, + 0.010727949440479279, + 0.01081647165119648, + 0.010939526371657848, + 0.01698281429708004, + 0.01422367338091135, + 0.007880604825913906, + 0.00581852113828063, + 0.0079007213935256, + 0.0072288354858756065, + 0.005973059684038162, + 0.00420546205714345, + 0.0074894786812365055, + 0.004900992847979069, + 0.015653107315301895, + 0.00867503508925438, + 0.019769856706261635, + 0.00913348887115717, + 0.016255203634500504, + 0.02454224042594433, + 0.018210917711257935, + 0.021530305966734886, + 0.010500186122953892, + 0.008632558397948742, + 0.014588765799999237, + 0.007294987794011831, + 0.01119750551879406, + 0.016473963856697083, + 0.008239708840847015, + 0.011653720401227474, + 0.039681609719991684, + 0.03931313008069992, + 0.04700687155127525, + 0.040299851447343826, + 0.051664792001247406, + 0.061306580901145935, + 0.037119485437870026, + 0.058213066309690475, + 0.00847878772765398, + 0.004675707314163446, + 0.01022539846599102, + 0.005948633886873722, + 0.007187406532466412, + 0.010802281089127064, + 0.009260057471692562, + 0.008907576091587543, + 0.0051457746885716915, + 0.001888179685920477, + 0.005252480041235685, + 0.001666035270318389, + 0.002546357223764062, + 0.0045318217016756535, + 0.0032792328856885433, + 0.0032764007337391376, + 0.00530035886913538, + 0.004971328657120466, + 0.006041142623871565, + 0.004542558453977108, + 0.004343134351074696, + 0.009492581710219383, + 0.004756988491863012, + 0.005212501622736454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.03411974385380745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -1.05392324924469 + ], + "opposite_logprob_tokens": [ + -0.42892324924468994 + ], + "logprob_reward": -1.05392324924469, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.05392324924469, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05328129604458809, + 0.038183219730854034, + 0.0717412605881691, + 0.06209506466984749, + 0.1202450841665268, + 0.06714478880167007, + 0.47486552596092224, + 0.08437801897525787, + 0.019691338762640953, + 0.008374392986297607 + ], + "credit_map": [ + 0.004381347913295031, + 0.005233601666986942, + 0.0085027776658535, + 0.0077347480691969395, + 0.0054162126034498215, + 0.006696559023112059, + 0.0047241938300430775, + 0.010591857135295868, + 0.0052484916523098946, + 0.0035879367496818304, + 0.00521130720153451, + 0.005373458843678236, + 0.004437463358044624, + 0.006454118061810732, + 0.003862738376483321, + 0.004007706418633461, + 0.010081878863275051, + 0.00693473219871521, + 0.013251425698399544, + 0.007767173461616039, + 0.006392445415258408, + 0.01030263677239418, + 0.01001267321407795, + 0.006998289842158556, + 0.009992887265980244, + 0.006903761066496372, + 0.010661382228136063, + 0.006865703500807285, + 0.00685843825340271, + 0.004023707006126642, + 0.01132757868617773, + 0.005461606662720442, + 0.014118076302111149, + 0.009806242771446705, + 0.017788385972380638, + 0.008691289462149143, + 0.012922951020300388, + 0.021365974098443985, + 0.016582563519477844, + 0.01896960847079754, + 0.008974653668701649, + 0.005005164537578821, + 0.0118336770683527, + 0.006051754113286734, + 0.00867321528494358, + 0.011212332174181938, + 0.007146693300455809, + 0.00824730284512043, + 0.050521183758974075, + 0.052867598831653595, + 0.05963794142007828, + 0.050491202622652054, + 0.0640932023525238, + 0.07841166853904724, + 0.04623245820403099, + 0.0726102665066719, + 0.012063208967447281, + 0.006344582885503769, + 0.013459503650665283, + 0.007818061858415604, + 0.009224843233823776, + 0.01318290177732706, + 0.011288485489785671, + 0.010996431112289429, + 0.003922642674297094, + 0.0011258524609729648, + 0.003490539500489831, + 0.001719912514090538, + 0.0016650301404297352, + 0.0031614680774509907, + 0.0021604038774967194, + 0.0024454882368445396, + 0.0012843095464631915, + 0.0008582021691836417, + 0.0011436314089223742, + 0.0006532978732138872, + 0.0007568825967609882, + 0.001624083612114191, + 0.0008355448953807354, + 0.0012184412917122245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -1.05392324924469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018212003633379936 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.018212003633379936, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018212003633379936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09369964897632599, + 0.030324634164571762, + 0.12372836470603943, + 0.04531589522957802, + 0.12257624417543411, + 0.07413117587566376, + 0.40159082412719727, + 0.0633927583694458, + 0.03517042472958565, + 0.010070019401609898 + ], + "credit_map": [ + 0.0123000992462039, + 0.012549659237265587, + 0.012676655314862728, + 0.01408171746879816, + 0.012386580929160118, + 0.010194907896220684, + 0.00919563602656126, + 0.010314388200640678, + 0.004309475421905518, + 0.002351591596379876, + 0.004103055689483881, + 0.0035826482344418764, + 0.0036707548424601555, + 0.004511851351708174, + 0.0033987967763096094, + 0.004396455362439156, + 0.012834431603550911, + 0.015985598787665367, + 0.014682809822261333, + 0.018301820382475853, + 0.015339026227593422, + 0.013623994775116444, + 0.019283058121800423, + 0.013677621260285378, + 0.008136588148772717, + 0.00407774792984128, + 0.009174453094601631, + 0.00429493049159646, + 0.004746594466269016, + 0.004238287918269634, + 0.00637066038325429, + 0.004276631399989128, + 0.01356470212340355, + 0.01736719161272049, + 0.017859145998954773, + 0.009295293129980564, + 0.0118732750415802, + 0.020358791574835777, + 0.014568040147423744, + 0.017689784988760948, + 0.00896812416613102, + 0.0051881675608456135, + 0.012182505801320076, + 0.005303734447807074, + 0.009298969991505146, + 0.013402665965259075, + 0.006935387849807739, + 0.012851610779762268, + 0.042381443083286285, + 0.04315531626343727, + 0.05179690197110176, + 0.04179253801703453, + 0.05018340051174164, + 0.06988015025854111, + 0.03838512301445007, + 0.06401592493057251, + 0.00712616229429841, + 0.004051052033901215, + 0.009638680145144463, + 0.006136404816061258, + 0.006185208912938833, + 0.009783351793885231, + 0.008775020018219948, + 0.01169686857610941, + 0.004208039492368698, + 0.0032608727924525738, + 0.005050156731158495, + 0.003543696366250515, + 0.0037745407316833735, + 0.004768769256770611, + 0.004527997691184282, + 0.006036350503563881, + 0.0015390102053061128, + 0.0007502326625399292, + 0.001559320604428649, + 0.001110188546590507, + 0.001047502737492323, + 0.0013188717421144247, + 0.0014673153636977077, + 0.0012775763170793653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.018212003633379936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008645367808640003 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.008645367808640003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008645367808640003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11076555401086807, + 0.12390100955963135, + 0.28016671538352966, + 0.09649081528186798, + 0.09825000911951065, + 0.054254140704870224, + 0.09160345047712326, + 0.057658471167087555, + 0.02678041160106659, + 0.06012941896915436 + ], + "credit_map": [ + 0.011533442884683609, + 0.011909760534763336, + 0.013327040709555149, + 0.018349604681134224, + 0.016483470797538757, + 0.014618230052292347, + 0.011460172943770885, + 0.01308383233845234, + 0.022345466539263725, + 0.014432383701205254, + 0.019691890105605125, + 0.022728392854332924, + 0.010329457931220531, + 0.011216011829674244, + 0.01077632699161768, + 0.01238107867538929, + 0.03261508047580719, + 0.018123945221304893, + 0.05749156326055527, + 0.020824553444981575, + 0.026287410408258438, + 0.0621216706931591, + 0.027990393340587616, + 0.034712109714746475, + 0.02099708281457424, + 0.014433438889682293, + 0.021261565387248993, + 0.013301059603691101, + 0.004218045622110367, + 0.005217532627284527, + 0.01002349704504013, + 0.007038595620542765, + 0.014672443270683289, + 0.006844041403383017, + 0.014975274913012981, + 0.006228446029126644, + 0.00966042559593916, + 0.020648684352636337, + 0.011246961541473866, + 0.013973725959658623, + 0.011152881197631359, + 0.0034868859220296144, + 0.009477647952735424, + 0.003307180479168892, + 0.0042759994976222515, + 0.009277353063225746, + 0.005384017247706652, + 0.00789217371493578, + 0.009708442725241184, + 0.007574200630187988, + 0.012820838019251823, + 0.009426310658454895, + 0.009987222962081432, + 0.017303984612226486, + 0.009200852364301682, + 0.015581599436700344, + 0.010223912075161934, + 0.005486380308866501, + 0.009158900938928127, + 0.0061291297897696495, + 0.004980272613465786, + 0.006563567090779543, + 0.007684601936489344, + 0.007431703619658947, + 0.005903150886297226, + 0.002034252043813467, + 0.004993897397071123, + 0.002716434421017766, + 0.002171651925891638, + 0.002667259657755494, + 0.0026698531582951546, + 0.003623911878094077, + 0.012820434756577015, + 0.004818703513592482, + 0.01226956956088543, + 0.0043514808639883995, + 0.00407623453065753, + 0.006426279433071613, + 0.005685640964657068, + 0.00968107394874096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.008645367808640003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081107892096043 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011081107892096043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081107892096043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046177271753549576, + 0.0312079805880785, + 0.09355543553829193, + 0.05123366415500641, + 0.09943509101867676, + 0.08618214726448059, + 0.5041996240615845, + 0.05299548804759979, + 0.025681816041469574, + 0.009331559762358665 + ], + "credit_map": [ + 0.00606138352304697, + 0.005891374312341213, + 0.006471651140600443, + 0.008939813822507858, + 0.0036962132435292006, + 0.004491153173148632, + 0.003831924172118306, + 0.006793756037950516, + 0.0045508695766329765, + 0.003353401552885771, + 0.004926552530378103, + 0.0056971111334860325, + 0.003611029125750065, + 0.002873018616810441, + 0.0032087459694594145, + 0.0029872518498450518, + 0.010210767388343811, + 0.010781681165099144, + 0.013883065432310104, + 0.011721104383468628, + 0.010748702101409435, + 0.009287692606449127, + 0.014932945370674133, + 0.011989478953182697, + 0.008914056234061718, + 0.006998272147029638, + 0.009343171492218971, + 0.006776871625334024, + 0.0037822502199560404, + 0.00341789610683918, + 0.007150480058044195, + 0.0048506674356758595, + 0.010838769376277924, + 0.010406247340142727, + 0.014815805479884148, + 0.007166821975260973, + 0.009856333956122398, + 0.019386032596230507, + 0.010868214070796967, + 0.01609686389565468, + 0.010592115111649036, + 0.005943549331277609, + 0.01388335507363081, + 0.006007770542055368, + 0.010460696183145046, + 0.016677651554346085, + 0.007997562177479267, + 0.01461944542825222, + 0.05346091091632843, + 0.057426609098911285, + 0.06382615864276886, + 0.05396495386958122, + 0.0659644603729248, + 0.08321628719568253, + 0.04817648231983185, + 0.07816372811794281, + 0.005922696553170681, + 0.0034034892451018095, + 0.008121518418192863, + 0.005235995166003704, + 0.0054177893325686455, + 0.007762267254292965, + 0.007203700952231884, + 0.009928032755851746, + 0.003556452225893736, + 0.001998414285480976, + 0.004171541891992092, + 0.0024326364509761333, + 0.0025674847420305014, + 0.003357277950271964, + 0.00383626576513052, + 0.0037617410998791456, + 0.0012858955888077617, + 0.0007687147008255124, + 0.0014147540787234902, + 0.0006813010550104082, + 0.0008989352500066161, + 0.0018497336423024535, + 0.0010223649442195892, + 0.0014098609099164605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.011081107892096043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in stop-motion style, a pair of weathered yet steady hands manipulate a traditional woodcarving knife, its handle smooth and mellow from years of use. As each frame progresses, the blade strikes with a precise and decisive rhythm, each contact sending tiny sawdust dancing through the dense, solid wood, seemingly taking on a life of its own in the still air. The scene is illuminated by a warm spotlight from above and sideways, creating a striking contrast of light and dark against the dim backdrop of the artisan's workshop, highlighting the delicate grain of the wood and the sharp, metallic quality of the blade. With each leap and vanishing of the sawdust, the Buddha's outline is miraculously shaped: from the smooth forehead to the closed, contemplative eyelids to the subtle hint of a smile at the corners of the mouth, every detail is brought to life in this rhythmic and rhythmic visual language. Finally, a compassionate and serene Buddha statue slowly reveals its complete form amidst the interweaving of light, shadow, and wood grain. The entire process is filled with concentration, tranquility, and an almost sacred sense of creation.", + "seed": 11849, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_204_seed_11849_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": -0.13903801958076656, + "vlm_logprob": -0.13903801958076656, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.13903801958076656, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026829790323972702 + ], + "opposite_logprob_tokens": [ + -7.651829719543457 + ], + "logprob_reward": -0.026829790323972702, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.026829790323972702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3760373592376709, + 0.04258259758353233, + 0.017110981047153473, + 0.022229503840208054, + 0.025463305413722992, + 0.2761860489845276, + 0.05182787403464317, + 0.009869540110230446, + 0.13427875936031342, + 0.04441404715180397 + ], + "credit_map": [ + 0.035072505474090576, + 0.0374474972486496, + 0.01747920922935009, + 0.1108793392777443, + 0.00865141861140728, + 0.013670999556779861, + 0.02070786990225315, + 0.1321285218000412, + 0.0044152322225272655, + 0.0025063727516680956, + 0.0036940816789865494, + 0.01238197647035122, + 0.002201610244810581, + 0.0031185224652290344, + 0.003520084312185645, + 0.010744714178144932, + 0.0020232321694493294, + 0.0015702564269304276, + 0.0034437517169862986, + 0.002197149908170104, + 0.0017925356514751911, + 0.0018055738182738423, + 0.0018482503946870565, + 0.0024302315432578325, + 0.003125516464933753, + 0.0030676156748086214, + 0.002128741703927517, + 0.00285906158387661, + 0.0025445385836064816, + 0.0024074469693005085, + 0.0029586744494736195, + 0.003137907013297081, + 0.0020618338603526354, + 0.0019617753569036722, + 0.0027892738580703735, + 0.005136794876307249, + 0.002892768708989024, + 0.002681157784536481, + 0.004484293516725302, + 0.003455408615991473, + 0.01551933865994215, + 0.009743952192366123, + 0.0841621458530426, + 0.017848912626504898, + 0.01153568085283041, + 0.016830014064908028, + 0.10328087210655212, + 0.017265131697058678, + 0.0056214649230241776, + 0.004178993403911591, + 0.011350697837769985, + 0.006066254340112209, + 0.0036082302685827017, + 0.012276090681552887, + 0.004458783660084009, + 0.004267362877726555, + 0.0012729773297905922, + 0.0010631449986249208, + 0.001365019823424518, + 0.0014663487672805786, + 0.0010785144986584783, + 0.0011042814003303647, + 0.0012924526818096638, + 0.0012268006103113294, + 0.009365160018205643, + 0.005609764717519283, + 0.04368653893470764, + 0.007616687566041946, + 0.005306864157319069, + 0.052046749740839005, + 0.004231564700603485, + 0.0064154318533837795, + 0.008426554501056671, + 0.0034549382980912924, + 0.0072963354177773, + 0.0037239654920995235, + 0.0037886882200837135, + 0.006347284652292728, + 0.005117512308061123, + 0.006258766632527113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.026829790323972702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328551769256592 + ], + "opposite_logprob_tokens": [ + -0.7578551769256592 + ], + "logprob_reward": -0.6328551769256592, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328551769256592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15157267451286316, + 0.03885778412222862, + 0.015656180679798126, + 0.010605773888528347, + 0.03451631963253021, + 0.14960913360118866, + 0.03230055421590805, + 0.01366869080811739, + 0.41780492663383484, + 0.1354079395532608 + ], + "credit_map": [ + 0.034307681024074554, + 0.033641230314970016, + 0.009534886106848717, + 0.02523380145430565, + 0.00575858261436224, + 0.004775025881826878, + 0.006024924572557211, + 0.0322965607047081, + 0.004704435355961323, + 0.001838433789089322, + 0.0027957840356975794, + 0.013297810219228268, + 0.001256523304618895, + 0.002507826080545783, + 0.0025081674102693796, + 0.009948808699846268, + 0.0020527569577097893, + 0.0011200864100828767, + 0.0043285614810884, + 0.0019236596999689937, + 0.0009398250258527696, + 0.0018019455019384623, + 0.0017651808448135853, + 0.0017241673776879907, + 0.002007093746215105, + 0.0009411884821020067, + 0.0011291897390037775, + 0.0015179304173216224, + 0.0010577631182968616, + 0.001110860612243414, + 0.001431637560017407, + 0.001410111552104354, + 0.0038257851265370846, + 0.003947813995182514, + 0.0031346501782536507, + 0.007013825234025717, + 0.004126738756895065, + 0.0038170882035046816, + 0.004617746919393539, + 0.004032675642520189, + 0.008269110694527626, + 0.004928051494061947, + 0.045980896800756454, + 0.009894965216517448, + 0.006748532876372337, + 0.00965146441012621, + 0.0550822839140892, + 0.009053843095898628, + 0.0042211697436869144, + 0.0029833519365638494, + 0.005501419771462679, + 0.0054695396684110165, + 0.002343145664781332, + 0.0049307323060929775, + 0.0035905553959310055, + 0.0032606422901153564, + 0.002034870907664299, + 0.001310931402258575, + 0.0022239037789404392, + 0.0018827900057658553, + 0.0013012110721319914, + 0.0017925620777532458, + 0.0017175442771986127, + 0.0014048791490495205, + 0.02946287766098976, + 0.01750863343477249, + 0.13497720658779144, + 0.022633478045463562, + 0.015430663712322712, + 0.1609799563884735, + 0.017001966014504433, + 0.019810158759355545, + 0.027198510244488716, + 0.010702903382480145, + 0.02462933398783207, + 0.010912492871284485, + 0.011962433345615864, + 0.02022644691169262, + 0.012328602373600006, + 0.01744723506271839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.6328551769256592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362799108028412 + ], + "opposite_logprob_tokens": [ + -3.651362895965576 + ], + "logprob_reward": -0.026362799108028412, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362799108028412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27225810289382935, + 0.18267105519771576, + 0.038997527211904526, + 0.07739531993865967, + 0.0472462996840477, + 0.1238754615187645, + 0.0486743189394474, + 0.02886117435991764, + 0.11957895755767822, + 0.060441792011260986 + ], + "credit_map": [ + 0.038245491683483124, + 0.03792242705821991, + 0.01232301164418459, + 0.06992395222187042, + 0.007266915403306484, + 0.00933836493641138, + 0.013505954295396805, + 0.08373197168111801, + 0.015942079946398735, + 0.009655996225774288, + 0.012542740441858768, + 0.0583961121737957, + 0.008412553928792477, + 0.020148098468780518, + 0.011819442734122276, + 0.045754022896289825, + 0.0036102815065532923, + 0.0037490916438400745, + 0.007930107414722443, + 0.004828848876059055, + 0.0030112743843346834, + 0.005699389614164829, + 0.004833392333239317, + 0.005335142370313406, + 0.00958649069070816, + 0.013180041685700417, + 0.007789767812937498, + 0.0103447325527668, + 0.008145956322550774, + 0.007455778773874044, + 0.01053944043815136, + 0.010353111661970615, + 0.006305539980530739, + 0.005305009428411722, + 0.00523714255541563, + 0.009272517636418343, + 0.005727468524128199, + 0.005529901012778282, + 0.004975408315658569, + 0.004893312696367502, + 0.006745997350662947, + 0.005376970861107111, + 0.03626331314444542, + 0.009027642197906971, + 0.005822789389640093, + 0.007088052108883858, + 0.045750007033348083, + 0.007800684310495853, + 0.005945846438407898, + 0.005209516733884811, + 0.00892729964107275, + 0.007860440760850906, + 0.0045679002068936825, + 0.00501233758404851, + 0.0059115150943398476, + 0.005239466205239296, + 0.0038621467538177967, + 0.002828278113156557, + 0.003995790611952543, + 0.003521112957969308, + 0.0033620994072407484, + 0.004268201533704996, + 0.003495581215247512, + 0.0035279628355056047, + 0.011270523071289062, + 0.00843407865613699, + 0.030760465189814568, + 0.012167816050350666, + 0.005741445813328028, + 0.03682029992341995, + 0.008154007606208324, + 0.006230320781469345, + 0.008946781978011131, + 0.004742832388728857, + 0.010880320332944393, + 0.007462621200829744, + 0.0052626654505729675, + 0.00916125625371933, + 0.00465845363214612, + 0.009326859377324581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.026362799108028412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031928059179335833 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031928059179335833, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031928059179335833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11470495909452438, + 0.02331533655524254, + 0.03126364201307297, + 0.031535759568214417, + 0.04299728199839592, + 0.15651720762252808, + 0.05167131498456001, + 0.07178445160388947, + 0.03498296067118645, + 0.4412270784378052 + ], + "credit_map": [ + 0.01080941129475832, + 0.027159344404935837, + 0.009002971462905407, + 0.029643284156918526, + 0.0058059669099748135, + 0.00583690358325839, + 0.007563332561403513, + 0.01888374052941799, + 0.0024249849375337362, + 0.001710915588773787, + 0.0030268931295722723, + 0.005289087072014809, + 0.001820008852519095, + 0.002510310849174857, + 0.0018592729466035962, + 0.0046738628298044205, + 0.004186826758086681, + 0.0026429235003888607, + 0.003780620638281107, + 0.003915227018296719, + 0.002511079888790846, + 0.00463337916880846, + 0.003625202225521207, + 0.0059683844447135925, + 0.004749636631458998, + 0.005138315726071596, + 0.0033995304256677628, + 0.0037951290141791105, + 0.0038133931811898947, + 0.003411862999200821, + 0.004253477323800325, + 0.0029744149651378393, + 0.00395596120506525, + 0.0034039311576634645, + 0.005787973757833242, + 0.007611893117427826, + 0.00460221478715539, + 0.005686280317604542, + 0.006750581320375204, + 0.005198443308472633, + 0.00869116559624672, + 0.005657609552145004, + 0.04719824343919754, + 0.010276922956109047, + 0.00645665917545557, + 0.00925537757575512, + 0.05922873318195343, + 0.009752488695085049, + 0.006189454812556505, + 0.005118732340633869, + 0.008896628394722939, + 0.006847745273262262, + 0.004016549792140722, + 0.00797000341117382, + 0.0058836261741817, + 0.006748576182872057, + 0.018528856337070465, + 0.005467715673148632, + 0.011890541762113571, + 0.008083373308181763, + 0.004008331801742315, + 0.006733167916536331, + 0.00621333671733737, + 0.010859129950404167, + 0.004317499231547117, + 0.0025867121294140816, + 0.007413545623421669, + 0.0032523940317332745, + 0.0024395023938268423, + 0.008337137289345264, + 0.00350809283554554, + 0.0031280736438930035, + 0.08923958241939545, + 0.03727475181221962, + 0.08275172859430313, + 0.03941419720649719, + 0.04069143533706665, + 0.0656929612159729, + 0.035752467811107635, + 0.0504099540412426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0031928059179335833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005949525628238916 + ], + "opposite_logprob_tokens": [ + -5.1309494972229 + ], + "logprob_reward": -0.005949525628238916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005949525628238916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2280983328819275, + 0.0495152473449707, + 0.0721377432346344, + 0.060127682983875275, + 0.06198718398809433, + 0.27455106377601624, + 0.05644794553518295, + 0.039060357958078384, + 0.10379578173160553, + 0.0542786568403244 + ], + "credit_map": [ + 0.04026508703827858, + 0.01663162186741829, + 0.015447442419826984, + 0.05061310902237892, + 0.005745891481637955, + 0.007824036292731762, + 0.014076505787670612, + 0.07749464362859726, + 0.006200722884386778, + 0.0037475742865353823, + 0.006114875432103872, + 0.013435819186270237, + 0.003101990558207035, + 0.0027393654454499483, + 0.004129126202315092, + 0.010045773349702358, + 0.00832944456487894, + 0.006647966336458921, + 0.01591608300805092, + 0.007817394100129604, + 0.006140111479908228, + 0.00806948821991682, + 0.009534230455756187, + 0.009683026000857353, + 0.00659202691167593, + 0.005174144636839628, + 0.004877118393778801, + 0.006400928366929293, + 0.005119690205901861, + 0.007780463434755802, + 0.012986114248633385, + 0.011197197251021862, + 0.006622780580073595, + 0.006124499719589949, + 0.007673950400203466, + 0.009425169788300991, + 0.008593705482780933, + 0.008116589859127998, + 0.007812032476067543, + 0.007618463132530451, + 0.015552549622952938, + 0.01062313187867403, + 0.08275219798088074, + 0.01866341009736061, + 0.011449919082224369, + 0.017578745260834694, + 0.10072799026966095, + 0.017203116789460182, + 0.004293175879865885, + 0.003973482176661491, + 0.010861538350582123, + 0.006655553821474314, + 0.002805845346301794, + 0.015016917139291763, + 0.006251837592571974, + 0.006589592434465885, + 0.004925953224301338, + 0.0038198453839868307, + 0.005646788980811834, + 0.008784785866737366, + 0.0033506816253066063, + 0.003211814211681485, + 0.004867729730904102, + 0.004452758003026247, + 0.007665970828384161, + 0.007487169001251459, + 0.027437694370746613, + 0.00940576009452343, + 0.004842107184231281, + 0.03263198211789131, + 0.008044508285820484, + 0.006280585657805204, + 0.0071084932424128056, + 0.004590104799717665, + 0.008113354444503784, + 0.006150936242192984, + 0.004662730731070042, + 0.007573236711323261, + 0.006085127126425505, + 0.009994671680033207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.005949525628238916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up shot in stop-motion style, a pair of weathered yet steady hands manipulate a traditional woodcarving knife, its handle smooth and mellow from years of use. As each frame progresses, the blade strikes with a precise and decisive rhythm, each contact sending tiny sawdust dancing through the dense, solid wood, seemingly taking on a life of its own in the still air. The scene is illuminated by a warm spotlight from above and sideways, creating a striking contrast of light and dark against the dim backdrop of the artisan's workshop, highlighting the delicate grain of the wood and the sharp, metallic quality of the blade. With each leap and vanishing of the sawdust, the Buddha's outline is miraculously shaped: from the smooth forehead to the closed, contemplative eyelids to the subtle hint of a smile at the corners of the mouth, every detail is brought to life in this rhythmic and rhythmic visual language. Finally, a compassionate and serene Buddha statue slowly reveals its complete form amidst the interweaving of light, shadow, and wood grain. The entire process is filled with concentration, tranquility, and an almost sacred sense of creation.", + "seed": 11850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_4/index_204_seed_11850_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": -0.09033302674070001, + "vlm_logprob": -0.09033302674070001, + "vlm_yes_no_margin": 5.024999988079071, + "vlm_token_credit": -0.09033302674070001, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04882099851965904 + ], + "opposite_logprob_tokens": [ + -8.673821449279785 + ], + "logprob_reward": -0.04882099851965904, + "yes_no_margin_reward": 8.625, + "token_credit_reward": -0.04882099851965904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19856248795986176, + 0.13118942081928253, + 0.07077605277299881, + 0.09141114354133606, + 0.07365216314792633, + 0.049170978367328644, + 0.12217481434345245, + 0.14844854176044464, + 0.0742233395576477, + 0.040390998125076294 + ], + "credit_map": [ + 0.012522637844085693, + 0.010470266453921795, + 0.016605501994490623, + 0.04244301840662956, + 0.014168539084494114, + 0.01868627406656742, + 0.03885819390416145, + 0.04480806738138199, + 0.011630547232925892, + 0.009646906517446041, + 0.01280525978654623, + 0.023549912497401237, + 0.017127295956015587, + 0.01728806272149086, + 0.017770593985915184, + 0.021370861679315567, + 0.015367267653346062, + 0.0048582167364656925, + 0.0075801219791173935, + 0.007745309267193079, + 0.009932903572916985, + 0.008130364120006561, + 0.006592407822608948, + 0.010569466277956963, + 0.011464069597423077, + 0.005780935287475586, + 0.014689975418150425, + 0.019506458193063736, + 0.00760268047451973, + 0.005708169657737017, + 0.013924583792686462, + 0.012734278105199337, + 0.012872661463916302, + 0.004327363334596157, + 0.01159245427697897, + 0.007754168007522821, + 0.010479492135345936, + 0.0034639770165085793, + 0.011016619391739368, + 0.012145429849624634, + 0.00893423892557621, + 0.0029017168562859297, + 0.007111990824341774, + 0.0058769723400473595, + 0.00600777193903923, + 0.003451770404353738, + 0.005975379608571529, + 0.008911140263080597, + 0.016063008457422256, + 0.009344255551695824, + 0.01765255257487297, + 0.010246659629046917, + 0.021847300231456757, + 0.014915295876562595, + 0.013330569490790367, + 0.01877516880631447, + 0.023622168228030205, + 0.008765432983636856, + 0.030834970995783806, + 0.011108213104307652, + 0.022241801023483276, + 0.007829869166016579, + 0.012416268698871136, + 0.03162982687354088, + 0.011898528784513474, + 0.007413278333842754, + 0.010793336667120457, + 0.00873597338795662, + 0.012645432725548744, + 0.006859231274574995, + 0.009430560283362865, + 0.006446999032050371, + 0.006710059475153685, + 0.0034724846482276917, + 0.00542594725266099, + 0.004425801802426577, + 0.005831733345985413, + 0.005081069190055132, + 0.004595968406647444, + 0.004847939591854811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.04882099851965904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708192110061646 + ], + "opposite_logprob_tokens": [ + -1.1370818614959717 + ], + "logprob_reward": -0.38708192110061646, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38708192110061646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15222695469856262, + 0.13669194281101227, + 0.05434669926762581, + 0.07554470747709274, + 0.09135258197784424, + 0.07968740165233612, + 0.12360164523124695, + 0.10997269302606583, + 0.14118030667304993, + 0.03539501503109932 + ], + "credit_map": [ + 0.0134107805788517, + 0.011729946359992027, + 0.021228648722171783, + 0.021034706383943558, + 0.03087066300213337, + 0.013711231760680676, + 0.018110018223524094, + 0.022130977362394333, + 0.01713791862130165, + 0.009177178144454956, + 0.01815025880932808, + 0.024927912279963493, + 0.01168456394225359, + 0.009991093538701534, + 0.019137805327773094, + 0.026485225185751915, + 0.013057813048362732, + 0.003963402472436428, + 0.005508422385901213, + 0.008831450715661049, + 0.004675146657973528, + 0.0034713237546384335, + 0.005837820470333099, + 0.00900132954120636, + 0.008942135609686375, + 0.005485710687935352, + 0.013804377056658268, + 0.015687672421336174, + 0.006158982869237661, + 0.005415130406618118, + 0.009789833799004555, + 0.010260873474180698, + 0.019033027812838554, + 0.003402836387977004, + 0.010678528808057308, + 0.011542248539626598, + 0.00802958570420742, + 0.004358616657555103, + 0.014960830099880695, + 0.01934691146016121, + 0.01289275847375393, + 0.005228342954069376, + 0.011230278760194778, + 0.008025120012462139, + 0.007567822001874447, + 0.008686321787536144, + 0.011119908653199673, + 0.014936858788132668, + 0.024698060005903244, + 0.011436696164309978, + 0.01826806552708149, + 0.013085857033729553, + 0.016827881336212158, + 0.014783325605094433, + 0.01288242731243372, + 0.011619343422353268, + 0.017452195286750793, + 0.006137929856777191, + 0.020542696118354797, + 0.00848514772951603, + 0.009837225079536438, + 0.006889825686812401, + 0.011543654836714268, + 0.02908403053879738, + 0.017777327448129654, + 0.01009827945381403, + 0.018020065501332283, + 0.015956981107592583, + 0.030850611627101898, + 0.016725953668355942, + 0.01867762766778469, + 0.013073476031422615, + 0.007582186255604029, + 0.002641098340973258, + 0.0063905129209160805, + 0.002698287833482027, + 0.005701879970729351, + 0.0036746144760400057, + 0.003783502848818898, + 0.0029229363426566124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.38708192110061646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13734474778175354, + 0.10848333686590195, + 0.11133702099323273, + 0.1260012984275818, + 0.10798191279172897, + 0.10079433768987656, + 0.12010091543197632, + 0.10659239441156387, + 0.03569219261407852, + 0.04567188397049904 + ], + "credit_map": [ + 0.007430818397551775, + 0.01050658244639635, + 0.013561785221099854, + 0.02974737621843815, + 0.006884389091283083, + 0.011935475282371044, + 0.02619933895766735, + 0.03107897937297821, + 0.00847588386386633, + 0.00830873567610979, + 0.015421661548316479, + 0.02815096266567707, + 0.009111454710364342, + 0.006868049502372742, + 0.013921418227255344, + 0.018225161358714104, + 0.017475098371505737, + 0.00849207118153572, + 0.008432759903371334, + 0.019237754866480827, + 0.016404105350375175, + 0.011976579204201698, + 0.012541298754513264, + 0.016777340322732925, + 0.015549633651971817, + 0.0091028967872262, + 0.02053152024745941, + 0.025057142600417137, + 0.011917877942323685, + 0.00942760519683361, + 0.017642050981521606, + 0.016772562637925148, + 0.017008380964398384, + 0.004238057415932417, + 0.02184673584997654, + 0.012465923093259335, + 0.006699713878333569, + 0.007621748372912407, + 0.019058307632803917, + 0.019043033942580223, + 0.016540350392460823, + 0.007507766596972942, + 0.014828404411673546, + 0.011453975923359394, + 0.010281603783369064, + 0.008150827139616013, + 0.00982233788818121, + 0.022209065034985542, + 0.022219691425561905, + 0.009524968452751637, + 0.018153222277760506, + 0.013157028704881668, + 0.018146665766835213, + 0.01295732893049717, + 0.011752050369977951, + 0.014189948327839375, + 0.015310492366552353, + 0.007213569711893797, + 0.0227844025939703, + 0.008723931387066841, + 0.01631341688334942, + 0.006291151978075504, + 0.010008147917687893, + 0.019947271794080734, + 0.005887117236852646, + 0.0026062142569571733, + 0.0051683904603123665, + 0.0026656189002096653, + 0.007590047083795071, + 0.003167380578815937, + 0.004564431495964527, + 0.0040429881773889065, + 0.011004152707755566, + 0.002733407309278846, + 0.007698451168835163, + 0.003538927063345909, + 0.006219319999217987, + 0.005955468397587538, + 0.005659677088260651, + 0.002862473949790001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.005258656572550535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" 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 close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13546222448349, + 0.0701831579208374, + 0.021338868886232376, + 0.06166541576385498, + 0.07287545502185822, + 0.0584002360701561, + 0.10676370561122894, + 0.1562480479478836, + 0.16202905774116516, + 0.15503384172916412 + ], + "credit_map": [ + 0.012026126496493816, + 0.008729905821383, + 0.018005888909101486, + 0.02985311672091484, + 0.01012763474136591, + 0.013676602393388748, + 0.020738160237669945, + 0.022304784506559372, + 0.006979241035878658, + 0.004727385938167572, + 0.010614342987537384, + 0.013237987644970417, + 0.0057759471237659454, + 0.004318094812333584, + 0.009693527594208717, + 0.01483662985265255, + 0.002936337376013398, + 0.0021158799063414335, + 0.001899206661619246, + 0.003708101576194167, + 0.0024772300384938717, + 0.00204584002494812, + 0.0028243432752788067, + 0.0033319296780973673, + 0.007458205800503492, + 0.0041975597850978374, + 0.010491237044334412, + 0.0146502535790205, + 0.004820303060114384, + 0.0031556400936096907, + 0.008195503614842892, + 0.008696716278791428, + 0.01599782332777977, + 0.0030134115368127823, + 0.012558928690850735, + 0.008112543262541294, + 0.005239236634224653, + 0.002760525094345212, + 0.008046274073421955, + 0.01714671589434147, + 0.008417206816375256, + 0.004076501354575157, + 0.00975598394870758, + 0.007153105456382036, + 0.010183441452682018, + 0.003758564591407776, + 0.006961769890040159, + 0.008093657903373241, + 0.020590171217918396, + 0.008196692913770676, + 0.01640445366501808, + 0.009180714376270771, + 0.01558763813227415, + 0.01251128688454628, + 0.010874446481466293, + 0.013418292626738548, + 0.024845508858561516, + 0.00966197531670332, + 0.025436580181121826, + 0.014628357253968716, + 0.016966281458735466, + 0.011826105415821075, + 0.015557056292891502, + 0.037326179444789886, + 0.021192871034145355, + 0.014521009288728237, + 0.02215585485100746, + 0.018367430195212364, + 0.029796648770570755, + 0.018558062613010406, + 0.021770300343632698, + 0.015666883438825607, + 0.0379800982773304, + 0.011136399582028389, + 0.025204436853528023, + 0.01554821152240038, + 0.017536770552396774, + 0.018291166052222252, + 0.01739407144486904, + 0.011942680925130844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "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 video shot in a close-up perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style stop-motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are weathered hands manipulating a woodcarving knife to shape wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm spotlight from above and sideways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the gradual carving of a Buddha statue's face and form from wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252134520560503 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252134520560503, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252134520560503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15141622722148895, + 0.12721136212348938, + 0.0570598728954792, + 0.04163110628724098, + 0.11111855506896973, + 0.08965534716844559, + 0.11046215891838074, + 0.09529091417789459, + 0.17656870186328888, + 0.03958573937416077 + ], + "credit_map": [ + 0.009437210857868195, + 0.009446509182453156, + 0.0194438137114048, + 0.028357433155179024, + 0.021957160905003548, + 0.015429332852363586, + 0.022723650559782982, + 0.024621110409498215, + 0.013677526265382767, + 0.009458253160119057, + 0.013665427453815937, + 0.020072203129529953, + 0.01561661995947361, + 0.012886152602732182, + 0.019306277856230736, + 0.022528894245624542, + 0.005916683468967676, + 0.0048941317945718765, + 0.006451789755374193, + 0.011422848328948021, + 0.009867198765277863, + 0.006030735559761524, + 0.004974734038114548, + 0.007501750253140926, + 0.004279322922229767, + 0.0031968785915523767, + 0.0048247454687952995, + 0.006460396107286215, + 0.005665673408657312, + 0.0048100706189870834, + 0.006382242776453495, + 0.006011779885739088, + 0.018638955429196358, + 0.004351368639618158, + 0.010099979117512703, + 0.013283446431159973, + 0.013351023197174072, + 0.006669601425528526, + 0.024752236902713776, + 0.01997194066643715, + 0.014380360953509808, + 0.006184399127960205, + 0.009550275281071663, + 0.011381261982023716, + 0.017294621095061302, + 0.006425431929528713, + 0.010809194296598434, + 0.013629796914756298, + 0.024655945599079132, + 0.008100174367427826, + 0.019962530583143234, + 0.01331596914678812, + 0.012473085895180702, + 0.008444484323263168, + 0.014362159185111523, + 0.009147809818387032, + 0.013833915814757347, + 0.005969715770334005, + 0.0187683068215847, + 0.007068997714668512, + 0.00955494400113821, + 0.006231608334928751, + 0.009493228979408741, + 0.024370193481445312, + 0.0229213684797287, + 0.010144036263227463, + 0.023564156144857407, + 0.016722701489925385, + 0.040894027799367905, + 0.02203514240682125, + 0.022895002737641335, + 0.01739225909113884, + 0.009590310044586658, + 0.002596447244286537, + 0.0075567602179944515, + 0.00289422576315701, + 0.005998449865728617, + 0.004836451727896929, + 0.004247051663696766, + 0.0018660433124750853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.005252134520560503 + } + ], + "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/0000158/group_5/index_tensor([8696, 8696])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_tensor([8696, 8696])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0c8a491bd881a91723db8cc063bd59ffc48ded59 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_tensor([8696, 8696])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene begins with a slow, graceful crane shot descending from a wide angle to a medium shot, intimately depicting a young man and woman. They glide gracefully across a vast outdoor ice rink, bathed in the warm, golden glow of the afternoon sun. The man, with windswept black hair, wears a dark gray cardigan and dark pants; a gentle, adoring smile graces his face as he gazes upon his partner. The woman, her long blond hair flowing freely behind her, wears a cream turtleneck sweater and a flowing red skating skirt. Her cheeks are flushed by the cold, and her bright, warm smile mirrors his. Hand in hand, their movements perfectly synchronized, they carve a smooth, sweeping arc across the pristine ice, their skates stirring up a fine, shimmering mist. The background is a soft-focus winter wonderland, with snow-capped pine trees and the distant, blurred movements of the other skaters. The low-angle sunlight creates long, soft shadows and beautiful, cinematic lens flares, enriching the scene’s warm tones of gold, red, and dark gray, and contributing to an atmosphere of profound romance, tranquility, and joy.", + "seed": 18455, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_8696_seed_18455_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": -0.14380498258396984, + "vlm_logprob": -0.14380498258396984, + "vlm_yes_no_margin": 2.8500000119209288, + "vlm_token_credit": -0.14380498258396984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5172972083091736 + ], + "opposite_logprob_tokens": [ + -1.0172972679138184 + ], + "logprob_reward": -0.5172972083091736, + "yes_no_margin_reward": 0.5000000596046448, + "token_credit_reward": -0.5172972083091736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05564466118812561, + 0.08828837424516678, + 0.0314575657248497, + 0.3174114525318146, + 0.020574234426021576, + 0.0415063239634037, + 0.01313826348632574, + 0.029397279024124146, + 0.018801890313625336, + 0.38377994298934937 + ], + "credit_map": [ + 0.008973479270935059, + 0.007038244977593422, + 0.006328336428850889, + 0.006103835068643093, + 0.006721521262079477, + 0.007845458574593067, + 0.006914610043168068, + 0.005719176027923822, + 0.0073830196633934975, + 0.0049064019694924355, + 0.018045896664261818, + 0.004444709047675133, + 0.011555655859410763, + 0.03078480437397957, + 0.0031005332712084055, + 0.008067360147833824, + 0.0021433010697364807, + 0.003621008712798357, + 0.004030853975564241, + 0.006366155110299587, + 0.002351724775508046, + 0.0027625449001789093, + 0.00528847798705101, + 0.004893499892205, + 0.010186940431594849, + 0.006474536377936602, + 0.13553357124328613, + 0.007420169655233622, + 0.006119659636169672, + 0.01003221794962883, + 0.009300424717366695, + 0.13234394788742065, + 0.0017720648320391774, + 0.0029677681159228086, + 0.0028029426466673613, + 0.001939839101396501, + 0.0023541145492345095, + 0.003868269268423319, + 0.0021371524780988693, + 0.0027320829685777426, + 0.004406309686601162, + 0.00633940240368247, + 0.0037248253356665373, + 0.0025383292231708765, + 0.005894117522984743, + 0.008770942687988281, + 0.004421353340148926, + 0.005411047022789717, + 0.0016228508902713656, + 0.0019492597784847021, + 0.001651693950407207, + 0.0015880910214036703, + 0.0014532682253047824, + 0.0022659210953861475, + 0.0012441826984286308, + 0.0013629964087158442, + 0.0020099827088415623, + 0.003429666394367814, + 0.0023984971921890974, + 0.0022800681181252003, + 0.0040672714821994305, + 0.006786196492612362, + 0.003030266845598817, + 0.0053953323513269424, + 0.0014000603696331382, + 0.002825628500431776, + 0.0009131520055234432, + 0.001181798754259944, + 0.005091895814985037, + 0.004422467201948166, + 0.001560383359901607, + 0.0014065041905269027, + 0.007683813106268644, + 0.008691065944731236, + 0.005657609552145004, + 0.006494331173598766, + 0.17072607576847076, + 0.1722419261932373, + 0.0045966580510139465, + 0.007688465528190136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.5172972083091736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983766235411167 + ], + "opposite_logprob_tokens": [ + -3.5298376083374023 + ], + "logprob_reward": -0.02983766235411167, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983766235411167, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04950307309627533, + 0.10920664668083191, + 0.020553739741444588, + 0.2116507738828659, + 0.03513961657881737, + 0.022047782316803932, + 0.017984073609113693, + 0.020180732011795044, + 0.11531533300876617, + 0.39841824769973755 + ], + "credit_map": [ + 0.008870785124599934, + 0.0068476032465696335, + 0.005479516461491585, + 0.005230999551713467, + 0.005137807689607143, + 0.007798590697348118, + 0.004614741075783968, + 0.0055230227299034595, + 0.008476504124701023, + 0.00570164667442441, + 0.02434864453971386, + 0.0051786755211651325, + 0.013489705510437489, + 0.03850061073899269, + 0.0031202337704598904, + 0.010390612296760082, + 0.0017445621779188514, + 0.003063495270907879, + 0.0024539181031286716, + 0.003742244327440858, + 0.002076726173982024, + 0.0019059398910030723, + 0.0027264126110821962, + 0.002840441884472966, + 0.007559764198958874, + 0.0024881570134311914, + 0.09335368871688843, + 0.004250465426594019, + 0.003042468335479498, + 0.00604424811899662, + 0.005539822392165661, + 0.08937214314937592, + 0.002573449397459626, + 0.006288805976510048, + 0.00490386551246047, + 0.005159285385161638, + 0.004401266574859619, + 0.0038198912516236305, + 0.003817135002464056, + 0.004175918642431498, + 0.0016305786557495594, + 0.0036504194140434265, + 0.0019065956585109234, + 0.0015626634703949094, + 0.0033313254825770855, + 0.0038598040118813515, + 0.003147275885567069, + 0.002959115896373987, + 0.0025789893697947264, + 0.0015782421687617898, + 0.001444946858100593, + 0.0011347875697538257, + 0.0021708491258323193, + 0.005038552917540073, + 0.0020504354033619165, + 0.0019872700795531273, + 0.0022760755382478237, + 0.0033284123055636883, + 0.0018544821068644524, + 0.0026792227290570736, + 0.0034071949776262045, + 0.0025295258965343237, + 0.0023128448519855738, + 0.0017929731402546167, + 0.011382781900465488, + 0.0159435011446476, + 0.005202779546380043, + 0.005290849134325981, + 0.036351196467876434, + 0.0322151854634285, + 0.004128224216401577, + 0.0048008086159825325, + 0.007729162462055683, + 0.008629770018160343, + 0.005845916457474232, + 0.006484867539256811, + 0.17807556688785553, + 0.17907847464084625, + 0.00466543436050415, + 0.007909055799245834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.02983766235411167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674506975337863 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.00674506975337863, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674506975337863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013906020671129227, + 0.09932319074869156, + 0.011756743304431438, + 0.16706500947475433, + 0.02374030277132988, + 0.04048469290137291, + 0.020377377048134804, + 0.004942839965224266, + 0.028759116306900978, + 0.5896446704864502 + ], + "credit_map": [ + 0.001515033538453281, + 0.0019475173903629184, + 0.0014501026598736644, + 0.0012913004029542208, + 0.00212219194509089, + 0.001754612778313458, + 0.00237872707657516, + 0.0014465367421507835, + 0.007634758949279785, + 0.005006405059248209, + 0.020839620381593704, + 0.0048608360812067986, + 0.013064377941191196, + 0.03575383126735687, + 0.0035050716251134872, + 0.00865829735994339, + 0.0009389646002091467, + 0.0015039894497022033, + 0.0016841819742694497, + 0.001764247310347855, + 0.001296744798310101, + 0.0010794708505272865, + 0.0014211757807061076, + 0.002067968947812915, + 0.006151865236461163, + 0.004176445305347443, + 0.06803804636001587, + 0.0069783153012394905, + 0.0032703455071896315, + 0.0065780943259596825, + 0.008436346426606178, + 0.06343556195497513, + 0.0020483024418354034, + 0.004878439474850893, + 0.0028269789181649685, + 0.0026488900184631348, + 0.0029314991552382708, + 0.002590151969343424, + 0.0026468164287507534, + 0.0031692280899733305, + 0.004382535349577665, + 0.0071662538684904575, + 0.0035975798964500427, + 0.003308316459879279, + 0.005349007435142994, + 0.007063155993819237, + 0.004668448120355606, + 0.004949401598423719, + 0.0029465407133102417, + 0.001997164683416486, + 0.0016378097934648395, + 0.0012737840879708529, + 0.0024861423298716545, + 0.0058131651021540165, + 0.0020969538018107414, + 0.0021258206106722355, + 0.0004163913836237043, + 0.001224909326992929, + 0.00044234225060790777, + 0.0005947806057520211, + 0.0006927839131094515, + 0.0005856331554241478, + 0.00045451143523678184, + 0.0005314887384884059, + 0.003259099554270506, + 0.004184812307357788, + 0.0011333785951137543, + 0.0014971864875406027, + 0.008633151650428772, + 0.00719171529635787, + 0.0014369271229952574, + 0.0014228473883122206, + 0.011545508168637753, + 0.012584549374878407, + 0.008620657958090305, + 0.009613651782274246, + 0.26358869671821594, + 0.2649158239364624, + 0.006951156537979841, + 0.011824688874185085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.00674506975337863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810269385576248 + ], + "opposite_logprob_tokens": [ + -3.288102626800537 + ], + "logprob_reward": -0.03810269385576248, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810269385576248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0083595160394907, + 0.020013434812426567, + 0.005855529569089413, + 0.12412097305059433, + 0.008797583170235157, + 0.004864063113927841, + 0.017323791980743408, + 0.006170026492327452, + 0.011202448979020119, + 0.7932925820350647 + ], + "credit_map": [ + 0.0010415446013212204, + 0.0009812370408326387, + 0.0008477247320115566, + 0.0009368696482852101, + 0.0010332847014069557, + 0.001335388282313943, + 0.0009441135916858912, + 0.0012393539072945714, + 0.0018095043487846851, + 0.0011963233118876815, + 0.0036994211841374636, + 0.001148870913311839, + 0.002471694489941001, + 0.006568776909261942, + 0.0007258394034579396, + 0.002393006347119808, + 0.0005270784022286534, + 0.0007222808781079948, + 0.0007367837242782116, + 0.0009890841320157051, + 0.0006018235580995679, + 0.00038081096136011183, + 0.0010419911704957485, + 0.0008556775283068419, + 0.004165145568549633, + 0.0036021245177835226, + 0.04825199395418167, + 0.0063733975403010845, + 0.0027206677477806807, + 0.0039313738234341145, + 0.008834611624479294, + 0.046241678297519684, + 0.0006808201433159411, + 0.000844498339574784, + 0.0011932200286537409, + 0.0007962643867358565, + 0.0013068822445347905, + 0.001949440105818212, + 0.000846910523250699, + 0.0011795483296737075, + 0.0005109080229885876, + 0.0007681921124458313, + 0.0004966083215549588, + 0.00036007349262945354, + 0.0007347855716943741, + 0.0008843637187965214, + 0.0005156331462785602, + 0.0005934992805123329, + 0.0024656441528350115, + 0.0017460271483287215, + 0.001322814729064703, + 0.0011584735475480556, + 0.0019250974291935563, + 0.004986098036170006, + 0.001757261692546308, + 0.0019623786211013794, + 0.000505021889694035, + 0.0003885194309987128, + 0.00045296933967620134, + 0.00035306657082401216, + 0.0010404445929452777, + 0.0017643477767705917, + 0.0005318004405125976, + 0.0011338574113324285, + 0.001079982379451394, + 0.0012098188744857907, + 0.0005899661919102073, + 0.0004400477046146989, + 0.003359554335474968, + 0.002558420179411769, + 0.0005546047468669713, + 0.0014100557891651988, + 0.014946992509067059, + 0.016947083175182343, + 0.01154700294137001, + 0.012910045683383942, + 0.35481858253479004, + 0.3576870858669281, + 0.008836890570819378, + 0.015598977915942669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.03810269385576248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270422786474228 + ], + "opposite_logprob_tokens": [ + -2.127042293548584 + ], + "logprob_reward": -0.1270422786474228, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270422786474228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013319623656570911, + 0.04333491623401642, + 0.0046952860429883, + 0.12817230820655823, + 0.012744497507810593, + 0.0051357033662498, + 0.03132389485836029, + 0.0016112488228827715, + 0.005819976329803467, + 0.7538425922393799 + ], + "credit_map": [ + 0.0027852621860802174, + 0.0017649130895733833, + 0.0014656304847449064, + 0.0011224789777770638, + 0.0013500835048034787, + 0.0023928543087095022, + 0.001034916378557682, + 0.0014034826308488846, + 0.0034212409518659115, + 0.0023067379370331764, + 0.009288662113249302, + 0.002195084700360894, + 0.005438145250082016, + 0.014950649812817574, + 0.0012726738350465894, + 0.004461721982806921, + 0.0005835123010911047, + 0.0004901976208202541, + 0.0005608838400803506, + 0.0005905235884711146, + 0.0005104536539874971, + 0.0009492794051766396, + 0.0003707555297296494, + 0.0006396797834895551, + 0.004503046162426472, + 0.004129544831812382, + 0.04535036161541939, + 0.008003400638699532, + 0.003921201918274164, + 0.0069551486521959305, + 0.011161294765770435, + 0.044148292392492294, + 0.0010477150790393353, + 0.0015573033597320318, + 0.002413250273093581, + 0.0007990541053004563, + 0.0012260625371709466, + 0.0025495034642517567, + 0.0011836182093247771, + 0.001967989606782794, + 0.0007242846768349409, + 0.0006101434119045734, + 0.0005726288654841483, + 0.00032962855766527355, + 0.0005726111121475697, + 0.0012184672523289919, + 0.00043151338468305767, + 0.0006764254067093134, + 0.004624412395060062, + 0.0025793779641389847, + 0.0025968262925744057, + 0.0018358236411586404, + 0.003935301210731268, + 0.008984727784991264, + 0.0034803736489266157, + 0.003287051571533084, + 0.00020071244216524065, + 0.000284401437966153, + 0.00013982139353174716, + 0.0001546093844808638, + 0.00024056233814917505, + 0.00020821801444981247, + 0.00019372435053810477, + 0.0001891992724267766, + 0.0004569216107483953, + 0.0005552961956709623, + 0.00043753552017733455, + 0.0003754313802346587, + 0.0014329078840091825, + 0.0009893181268125772, + 0.00046606961404904723, + 0.0011064955033361912, + 0.014653478749096394, + 0.016251960769295692, + 0.010988309979438782, + 0.012328362092375755, + 0.3371163308620453, + 0.33864113688468933, + 0.008824140764772892, + 0.01503883395344019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.1270422786474228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, graceful crane shot descending from a wide angle to a medium shot, intimately depicting a young man and woman. They glide gracefully across a vast outdoor ice rink, bathed in the warm, golden glow of the afternoon sun. The man, with windswept black hair, wears a dark gray cardigan and dark pants; a gentle, adoring smile graces his face as he gazes upon his partner. The woman, her long blond hair flowing freely behind her, wears a cream turtleneck sweater and a flowing red skating skirt. Her cheeks are flushed by the cold, and her bright, warm smile mirrors his. Hand in hand, their movements perfectly synchronized, they carve a smooth, sweeping arc across the pristine ice, their skates stirring up a fine, shimmering mist. The background is a soft-focus winter wonderland, with snow-capped pine trees and the distant, blurred movements of the other skaters. The low-angle sunlight creates long, soft shadows and beautiful, cinematic lens flares, enriching the scene’s warm tones of gold, red, and dark gray, and contributing to an atmosphere of profound romance, tranquility, and joy.", + "seed": 18456, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_8696_seed_18456_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": -0.18316640509292484, + "vlm_logprob": -0.18316640509292484, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.18316640509292484, + "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\": \"Does the camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7964567542076111 + ], + "opposite_logprob_tokens": [ + -0.6714567542076111 + ], + "logprob_reward": -0.7964567542076111, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7964567542076111, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1127607524394989, + 0.040756210684776306, + 0.05496974289417267, + 0.2722814679145813, + 0.31024855375289917, + 0.044190578162670135, + 0.023638414219021797, + 0.07426060736179352, + 0.0170211773365736, + 0.049872446805238724 + ], + "credit_map": [ + 0.011961364187300205, + 0.016155479475855827, + 0.014169909060001373, + 0.016178252175450325, + 0.012308345176279545, + 0.01410179678350687, + 0.012909910641610622, + 0.014975698664784431, + 0.0034376727417111397, + 0.005370334722101688, + 0.004879381041973829, + 0.0064501045271754265, + 0.00505575817078352, + 0.004445304162800312, + 0.004308705218136311, + 0.006808949634432793, + 0.005298534408211708, + 0.004331962205469608, + 0.012582308612763882, + 0.004463887307792902, + 0.004689828027039766, + 0.0069318246096372604, + 0.004167893435806036, + 0.012503504753112793, + 0.013140321709215641, + 0.00487475423142314, + 0.11536199599504471, + 0.00818789191544056, + 0.0053830016404390335, + 0.010054069571197033, + 0.0041585820727050304, + 0.11112084239721298, + 0.017969200387597084, + 0.008846057578921318, + 0.1142028272151947, + 0.012858610600233078, + 0.010618090629577637, + 0.018269706517457962, + 0.008172519505023956, + 0.11931155622005463, + 0.003951504826545715, + 0.003665096592158079, + 0.009969396516680717, + 0.003977490589022636, + 0.003471061121672392, + 0.0065716225653886795, + 0.0035242107696831226, + 0.00906019750982523, + 0.0027066583279520273, + 0.002439618343487382, + 0.003665030701085925, + 0.0020394239109009504, + 0.0037977390456944704, + 0.003265470964834094, + 0.0023029414005577564, + 0.0034215333871543407, + 0.008412052877247334, + 0.006364833097904921, + 0.009456208907067776, + 0.007388934958726168, + 0.00880730152130127, + 0.018375784158706665, + 0.0071584791876375675, + 0.008297013118863106, + 0.0017593655502423644, + 0.0020452686585485935, + 0.003174442332237959, + 0.0018298779614269733, + 0.001474764314480126, + 0.002678354736417532, + 0.0012966104550287127, + 0.002762493444606662, + 0.005394601728767157, + 0.007982845418155193, + 0.005837123841047287, + 0.00723160058259964, + 0.005360871087759733, + 0.008225730620324612, + 0.005028521176427603, + 0.004811153281480074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.7964567542076111 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009810683317482471 + ], + "opposite_logprob_tokens": [ + -4.634810447692871 + ], + "logprob_reward": -0.009810683317482471, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009810683317482471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08061032742261887, + 0.027857491746544838, + 0.05908190831542015, + 0.03725428134202957, + 0.36492738127708435, + 0.03315626084804535, + 0.020503263920545578, + 0.020011937245726585, + 0.06540482491254807, + 0.29119229316711426 + ], + "credit_map": [ + 0.007559322752058506, + 0.009742222726345062, + 0.010878690518438816, + 0.013278028927743435, + 0.011193177662789822, + 0.009266712702810764, + 0.009239889681339264, + 0.009452280588448048, + 0.002145216567441821, + 0.0037420133594423532, + 0.002993086352944374, + 0.004759078845381737, + 0.00502408854663372, + 0.0028108947444707155, + 0.0031167957931756973, + 0.003266318002715707, + 0.005327288527041674, + 0.010055233724415302, + 0.009116294793784618, + 0.006308155134320259, + 0.006719725206494331, + 0.007362653035670519, + 0.004134231712669134, + 0.010058324784040451, + 0.0021429075859487057, + 0.0017620183061808348, + 0.011864932253956795, + 0.002215506276115775, + 0.003615706693381071, + 0.0023390529677271843, + 0.0017792576691135764, + 0.011534899473190308, + 0.020719103515148163, + 0.007158676162362099, + 0.14302802085876465, + 0.01565529592335224, + 0.008474286645650864, + 0.013953063637018204, + 0.006891694851219654, + 0.14904725551605225, + 0.0027277665212750435, + 0.006076877471059561, + 0.004570961929857731, + 0.003965239506214857, + 0.005030197557061911, + 0.003755049081519246, + 0.003333259606733918, + 0.0036969096399843693, + 0.0021640625782310963, + 0.0027700650971382856, + 0.002376314252614975, + 0.0032081829849630594, + 0.0028103236109018326, + 0.002369992434978485, + 0.0025963822845369577, + 0.0022079425398260355, + 0.002591164316982031, + 0.001668130629695952, + 0.002887289971113205, + 0.0018855941016227007, + 0.002293612342327833, + 0.004537283908575773, + 0.0016731752548366785, + 0.002475687535479665, + 0.0062485989183187485, + 0.004203039687126875, + 0.019846074283123016, + 0.00390985980629921, + 0.0033803291153162718, + 0.006499443668872118, + 0.002686063526198268, + 0.01863141730427742, + 0.026440564543008804, + 0.039401013404130936, + 0.03590250015258789, + 0.05592590570449829, + 0.04531370848417282, + 0.03587152808904648, + 0.027209123596549034, + 0.025127967819571495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.009810683317482471 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595545070245862 + ], + "opposite_logprob_tokens": [ + -5.130955219268799 + ], + "logprob_reward": -0.00595545070245862, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595545070245862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08389478921890259, + 0.07587083429098129, + 0.10500207543373108, + 0.3752848207950592, + 0.1365218609571457, + 0.12983864545822144, + 0.03008509799838066, + 0.02707516774535179, + 0.014378410764038563, + 0.022048281505703926 + ], + "credit_map": [ + 0.009238197468221188, + 0.013589365407824516, + 0.011160272173583508, + 0.011735295876860619, + 0.006740032695233822, + 0.011886322870850563, + 0.008716256357729435, + 0.010829037986695766, + 0.0048436992801725864, + 0.015083237551152706, + 0.008732408285140991, + 0.012759089469909668, + 0.011407297104597092, + 0.007611914537847042, + 0.007291803602129221, + 0.008141379803419113, + 0.00875984039157629, + 0.012481369078159332, + 0.024385398253798485, + 0.009860852733254433, + 0.006900656968355179, + 0.010659844614565372, + 0.006509486585855484, + 0.025444623082876205, + 0.018146930262446404, + 0.007043893914669752, + 0.15825797617435455, + 0.01096401084214449, + 0.008363577537238598, + 0.013925425708293915, + 0.0057135168462991714, + 0.15286950767040253, + 0.007754930295050144, + 0.002592266071587801, + 0.054265059530735016, + 0.005432616453617811, + 0.0025759460404515266, + 0.005377208814024925, + 0.0022822143509984016, + 0.05624161288142204, + 0.009962758049368858, + 0.013285225257277489, + 0.027874870225787163, + 0.010344319976866245, + 0.016615005210042, + 0.016885867342352867, + 0.010686809197068214, + 0.024183787405490875, + 0.0028667678125202656, + 0.002746202517300844, + 0.005463281646370888, + 0.003457492683082819, + 0.003977858927100897, + 0.0029274248518049717, + 0.0022382582537829876, + 0.00640780758112669, + 0.0031392821110785007, + 0.002501256763935089, + 0.003801859449595213, + 0.002379076788201928, + 0.0027370979078114033, + 0.006814377382397652, + 0.0022296058014035225, + 0.003472611540928483, + 0.001748691312968731, + 0.0014178972924128175, + 0.002202830044552684, + 0.0014977596001699567, + 0.0016588609432801604, + 0.0029002847149968147, + 0.0014801549259573221, + 0.0014719314640387893, + 0.00282709370367229, + 0.0022427865769714117, + 0.0031868976075202227, + 0.0031830407679080963, + 0.0017920525278896093, + 0.0035237965639680624, + 0.002647113287821412, + 0.002645498840138316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.00595545070245862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054962873458862305 + ], + "opposite_logprob_tokens": [ + -2.9299628734588623 + ], + "logprob_reward": -0.054962873458862305, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054962873458862305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16722598671913147, + 0.0293810423463583, + 0.06539922952651978, + 0.042601458728313446, + 0.40668660402297974, + 0.04752922058105469, + 0.03522201254963875, + 0.14165587723255157, + 0.019557463005185127, + 0.04474106431007385 + ], + "credit_map": [ + 0.018172601237893105, + 0.022098276764154434, + 0.026475343853235245, + 0.02552153915166855, + 0.01356984581798315, + 0.020899491384625435, + 0.013361779972910881, + 0.02712710201740265, + 0.002523000817745924, + 0.002734104171395302, + 0.0037224197294563055, + 0.0050271241925656796, + 0.0033942447043955326, + 0.0033306682016700506, + 0.004658002406358719, + 0.003991477657109499, + 0.005230800714343786, + 0.004254515282809734, + 0.018990397453308105, + 0.003734531346708536, + 0.003484150627627969, + 0.006613241974264383, + 0.0027401605620980263, + 0.020351434126496315, + 0.0027292191516608, + 0.0021744277328252792, + 0.012422650121152401, + 0.003176675410941243, + 0.00497068278491497, + 0.002644708612933755, + 0.0027440693229436874, + 0.011739022098481655, + 0.025621801614761353, + 0.010940884239971638, + 0.15050841867923737, + 0.02176278829574585, + 0.011779534630477428, + 0.019628793001174927, + 0.009992234408855438, + 0.15645214915275574, + 0.0037076203152537346, + 0.005945668555796146, + 0.010552887804806232, + 0.0036919910926371813, + 0.006556340958923101, + 0.0038736541755497456, + 0.0042924461886286736, + 0.00890861265361309, + 0.00363258202560246, + 0.004464122466742992, + 0.005655764602124691, + 0.003432132536545396, + 0.005643616896122694, + 0.003972469829022884, + 0.003012210363522172, + 0.005409114062786102, + 0.01794833317399025, + 0.012574941851198673, + 0.01954641006886959, + 0.01186204794794321, + 0.016425810754299164, + 0.0337950624525547, + 0.01191538106650114, + 0.01758788898587227, + 0.0020126397721469402, + 0.0029554071370512247, + 0.0038398040924221277, + 0.00199467153288424, + 0.001990187680348754, + 0.002558452542871237, + 0.001182092004455626, + 0.003024207428097725, + 0.004265739582479, + 0.0038464809767901897, + 0.006487350445240736, + 0.00913126952946186, + 0.004568479023873806, + 0.006857609376311302, + 0.004948224872350693, + 0.004635912831872702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.054962873458862305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048646263778209686 + ], + "opposite_logprob_tokens": [ + -3.0486462116241455 + ], + "logprob_reward": -0.048646263778209686, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048646263778209686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29439300298690796, + 0.04868413507938385, + 0.04456156864762306, + 0.029515380039811134, + 0.2138071060180664, + 0.030970681458711624, + 0.03792973607778549, + 0.19330917298793793, + 0.04533304274082184, + 0.06149614602327347 + ], + "credit_map": [ + 0.03711915388703346, + 0.05651712045073509, + 0.03394699841737747, + 0.041360240429639816, + 0.024596726521849632, + 0.03787984326481819, + 0.024010509252548218, + 0.038962412625551224, + 0.00630619190633297, + 0.005865258164703846, + 0.005805668421089649, + 0.006250990554690361, + 0.005318526178598404, + 0.007388025056570768, + 0.00468211667612195, + 0.007067354861646891, + 0.0054593211971223354, + 0.004815108608454466, + 0.007198726758360863, + 0.005043643061071634, + 0.003965959418565035, + 0.00610157148912549, + 0.004205956123769283, + 0.00777128292247653, + 0.0020652366802096367, + 0.0016848300583660603, + 0.008152603171765804, + 0.002453520195558667, + 0.00305150356143713, + 0.0026871589943766594, + 0.0015025794273242354, + 0.007917948067188263, + 0.012697079218924046, + 0.005791112780570984, + 0.07999889552593231, + 0.011253471486270428, + 0.006524648983031511, + 0.009032532572746277, + 0.004832561127841473, + 0.08367680758237839, + 0.0018772785551846027, + 0.0030345977284014225, + 0.008370148949325085, + 0.002959547331556678, + 0.0027178050950169563, + 0.0025274709332734346, + 0.001968974946066737, + 0.007514855358749628, + 0.003431156277656555, + 0.004453378729522228, + 0.005517715588212013, + 0.0028830666560679674, + 0.0056022354401648045, + 0.00475870119407773, + 0.004585053771734238, + 0.00669843191280961, + 0.020786102861166, + 0.015261419117450714, + 0.02629403956234455, + 0.019439756870269775, + 0.01992904394865036, + 0.048718664795160294, + 0.017961079254746437, + 0.024919066578149796, + 0.004977938253432512, + 0.004373581148684025, + 0.007293191272765398, + 0.00781370047479868, + 0.004735917318612337, + 0.006950950715690851, + 0.00404865900054574, + 0.0051391031593084335, + 0.007455566897988319, + 0.006228805985301733, + 0.01054325420409441, + 0.007094130851328373, + 0.006131602916866541, + 0.00892665982246399, + 0.0074830856174230576, + 0.007633041124790907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.048646263778209686 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, graceful crane shot descending from a wide angle to a medium shot, intimately depicting a young man and woman. They glide gracefully across a vast outdoor ice rink, bathed in the warm, golden glow of the afternoon sun. The man, with windswept black hair, wears a dark gray cardigan and dark pants; a gentle, adoring smile graces his face as he gazes upon his partner. The woman, her long blond hair flowing freely behind her, wears a cream turtleneck sweater and a flowing red skating skirt. Her cheeks are flushed by the cold, and her bright, warm smile mirrors his. Hand in hand, their movements perfectly synchronized, they carve a smooth, sweeping arc across the pristine ice, their skates stirring up a fine, shimmering mist. The background is a soft-focus winter wonderland, with snow-capped pine trees and the distant, blurred movements of the other skaters. The low-angle sunlight creates long, soft shadows and beautiful, cinematic lens flares, enriching the scene’s warm tones of gold, red, and dark gray, and contributing to an atmosphere of profound romance, tranquility, and joy.", + "seed": 18457, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_8696_seed_18457_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": -0.09037446342408657, + "vlm_logprob": -0.09037446342408657, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09037446342408657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26596683263778687 + ], + "opposite_logprob_tokens": [ + -1.6409668922424316 + ], + "logprob_reward": -0.26596683263778687, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.26596683263778687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09677060693502426, + 0.05452670156955719, + 0.04626982659101486, + 0.062335677444934845, + 0.042165376245975494, + 0.06779221445322037, + 0.02953924424946308, + 0.023684419691562653, + 0.06683462858200073, + 0.5100812911987305 + ], + "credit_map": [ + 0.007514612749218941, + 0.007636614143848419, + 0.02473435550928116, + 0.005538011901080608, + 0.012763654813170433, + 0.0089305080473423, + 0.0057368893176317215, + 0.023915970697999, + 0.003890143008902669, + 0.007753346581012011, + 0.008331049233675003, + 0.011635667644441128, + 0.0059089334681630135, + 0.006845941301435232, + 0.004603497684001923, + 0.0055581312626600266, + 0.004229838959872723, + 0.005548876244574785, + 0.007871619425714016, + 0.011676553636789322, + 0.004418519325554371, + 0.004379946738481522, + 0.0032712193205952644, + 0.004873258527368307, + 0.004710703622549772, + 0.0033104540780186653, + 0.01946995034813881, + 0.005918406415730715, + 0.00377950188703835, + 0.0037545096129179, + 0.0032742416951805353, + 0.018117913976311684, + 0.006256038323044777, + 0.00524301640689373, + 0.004493513610213995, + 0.005603518337011337, + 0.004450495354831219, + 0.006471436470746994, + 0.0036088216584175825, + 0.006038538180291653, + 0.00783624593168497, + 0.00860727671533823, + 0.00877784751355648, + 0.007829179987311363, + 0.01099757757037878, + 0.005794933065772057, + 0.007667441852390766, + 0.010281716473400593, + 0.0031504561193287373, + 0.004921189043670893, + 0.004346405155956745, + 0.004090346395969391, + 0.003683797549456358, + 0.002560448134317994, + 0.0034538174513727427, + 0.00333278626203537, + 0.0025509523693472147, + 0.003949456382542849, + 0.0024670183192938566, + 0.002744179219007492, + 0.0035024271346628666, + 0.002232842380180955, + 0.003095069667324424, + 0.0031424788758158684, + 0.0045786285772919655, + 0.013687826693058014, + 0.007723336573690176, + 0.00884326733648777, + 0.00904103834182024, + 0.006949907634407282, + 0.010258983820676804, + 0.005751646589487791, + 0.03188809007406235, + 0.03653936833143234, + 0.14444978535175323, + 0.09418369829654694, + 0.029597138985991478, + 0.13075388967990875, + 0.02189277671277523, + 0.02077658101916313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.26596683263778687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207640469074249 + ], + "opposite_logprob_tokens": [ + -2.8120763301849365 + ], + "logprob_reward": -0.06207640469074249, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207640469074249, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0992889255285263, + 0.08498220145702362, + 0.06434228271245956, + 0.08788532018661499, + 0.02827613428235054, + 0.03778509795665741, + 0.07692573219537735, + 0.06822427362203598, + 0.07251568138599396, + 0.37977442145347595 + ], + "credit_map": [ + 0.009761876426637173, + 0.008937153965234756, + 0.018444720655679703, + 0.0101859699934721, + 0.016022758558392525, + 0.012459557503461838, + 0.0077265603467822075, + 0.015750329941511154, + 0.006715297233313322, + 0.016495205461978912, + 0.011304880492389202, + 0.013203511945903301, + 0.013116074725985527, + 0.010410751216113567, + 0.006987893488258123, + 0.006748585030436516, + 0.0053390017710626125, + 0.008351588621735573, + 0.007719103712588549, + 0.008447765372693539, + 0.012144683860242367, + 0.005557066295295954, + 0.010472546331584454, + 0.006310527212917805, + 0.00589982233941555, + 0.006277522072196007, + 0.022619685158133507, + 0.010209945030510426, + 0.0070053935050964355, + 0.004566832445561886, + 0.009261331520974636, + 0.02204478345811367, + 0.002390235895290971, + 0.0037472716066986322, + 0.003736353712156415, + 0.005160942208021879, + 0.0037672496400773525, + 0.002108326181769371, + 0.003694019978865981, + 0.0036717350594699383, + 0.004861720837652683, + 0.0038507014978677034, + 0.0048327213153243065, + 0.0059007033705711365, + 0.004379915073513985, + 0.003117440966889262, + 0.005922000389546156, + 0.004919892176985741, + 0.010179495438933372, + 0.0068786186166107655, + 0.010838475078344345, + 0.007684675510972738, + 0.008661157451570034, + 0.013586347922682762, + 0.008385241962969303, + 0.01071171648800373, + 0.007400726433843374, + 0.012034580111503601, + 0.00763431191444397, + 0.0075197904370725155, + 0.009807459078729153, + 0.005935301072895527, + 0.008454820141196251, + 0.009437285363674164, + 0.008348057046532631, + 0.009019440971314907, + 0.00803531613200903, + 0.0055171349085867405, + 0.011456310749053955, + 0.012654019519686699, + 0.007194108795374632, + 0.010291295126080513, + 0.027470430359244347, + 0.02573840133845806, + 0.10271023213863373, + 0.06812271475791931, + 0.02859998308122158, + 0.09431669861078262, + 0.019100764766335487, + 0.013715206645429134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.06207640469074249 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743649020791054 + ], + "opposite_logprob_tokens": [ + -5.006743431091309 + ], + "logprob_reward": -0.006743649020791054, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743649020791054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10869067162275314, + 0.07299712300300598, + 0.04458050802350044, + 0.04557909443974495, + 0.07125253230333328, + 0.07309678196907043, + 0.07518432289361954, + 0.09195522218942642, + 0.0701211616396904, + 0.3465425968170166 + ], + "credit_map": [ + 0.008600389584898949, + 0.006690355017781258, + 0.0347667932510376, + 0.00601474242284894, + 0.007270150817930698, + 0.008510946296155453, + 0.005026242695748806, + 0.03181103989481926, + 0.005721323192119598, + 0.01402563601732254, + 0.009663170203566551, + 0.00965292938053608, + 0.009160038083791733, + 0.012473120354115963, + 0.005339056719094515, + 0.006961841601878405, + 0.0035991016775369644, + 0.006284434348344803, + 0.005370995961129665, + 0.006802842952311039, + 0.006223410367965698, + 0.005562234204262495, + 0.007058403920382261, + 0.0036790859885513783, + 0.00413435697555542, + 0.004350642673671246, + 0.010589543730020523, + 0.005518951918929815, + 0.003960221540182829, + 0.004061261657625437, + 0.0034170092549175024, + 0.009547107852995396, + 0.015958379954099655, + 0.010229391045868397, + 0.009006318636238575, + 0.007669265381991863, + 0.007467157673090696, + 0.004904779139906168, + 0.004502853844314814, + 0.011514383368194103, + 0.009574521332979202, + 0.008284248411655426, + 0.008283902890980244, + 0.008114427328109741, + 0.013647199608385563, + 0.0058218007907271385, + 0.007710474543273449, + 0.011660201475024223, + 0.012149624526500702, + 0.007389236707240343, + 0.010536924004554749, + 0.007699811831116676, + 0.007091347593814135, + 0.012368878349661827, + 0.006407331675291061, + 0.011541158892214298, + 0.00797361508011818, + 0.014415048994123936, + 0.010865566320717335, + 0.012149288319051266, + 0.013968732208013535, + 0.01000700518488884, + 0.009667227044701576, + 0.012908733449876308, + 0.009308337233960629, + 0.006345224101096392, + 0.008755338378250599, + 0.006231218110769987, + 0.009486333467066288, + 0.012442829087376595, + 0.006826928351074457, + 0.010724949650466442, + 0.022470694035291672, + 0.023755686357617378, + 0.10093153268098831, + 0.06583265215158463, + 0.01919366419315338, + 0.09292992204427719, + 0.011828194372355938, + 0.009600245393812656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.006743649020791054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108546286821365 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.038108546286821365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108546286821365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10070513933897018, + 0.08040402829647064, + 0.150517538189888, + 0.18587364256381989, + 0.044328201562166214, + 0.06373938173055649, + 0.1002417802810669, + 0.08220433443784714, + 0.039395030587911606, + 0.15259084105491638 + ], + "credit_map": [ + 0.009348931722342968, + 0.011532586999237537, + 0.020170504227280617, + 0.008143017999827862, + 0.011452881619334221, + 0.009867897257208824, + 0.007269623223692179, + 0.022919699549674988, + 0.005577836651355028, + 0.009168623015284538, + 0.012420592829585075, + 0.019398361444473267, + 0.009859930723905563, + 0.009524965658783913, + 0.006519495043903589, + 0.007934224791824818, + 0.016488613560795784, + 0.029642214998602867, + 0.02105814591050148, + 0.02734663151204586, + 0.016375556588172913, + 0.01542676705867052, + 0.010151260532438755, + 0.014028347097337246, + 0.013526014052331448, + 0.01347311120480299, + 0.04528877139091492, + 0.03157563507556915, + 0.013407839462161064, + 0.01508263498544693, + 0.015479284338653088, + 0.03804035112261772, + 0.007010928355157375, + 0.007560104597359896, + 0.004902991466224194, + 0.005733912345021963, + 0.0044814483262598515, + 0.004802639130502939, + 0.0027990052476525307, + 0.007037172093987465, + 0.009381457231938839, + 0.008171550929546356, + 0.008762320503592491, + 0.006381779443472624, + 0.007142952177673578, + 0.005697730462998152, + 0.006389666348695755, + 0.011811918579041958, + 0.01013554073870182, + 0.010139551013708115, + 0.0136902816593647, + 0.01114650908857584, + 0.01146039366722107, + 0.018675576895475388, + 0.008190320804715157, + 0.01680361106991768, + 0.006047840695828199, + 0.012555781751871109, + 0.008456316776573658, + 0.008340397849678993, + 0.012501501478254795, + 0.01099788211286068, + 0.012096594087779522, + 0.011208022944629192, + 0.003102691611275077, + 0.004339085426181555, + 0.0056042904034256935, + 0.005062222480773926, + 0.0053724260069429874, + 0.004849148914217949, + 0.004023327957838774, + 0.007041838485747576, + 0.010998671874403954, + 0.01125695277005434, + 0.04047781229019165, + 0.02683274634182453, + 0.009192842058837414, + 0.037832409143447876, + 0.0077126650139689445, + 0.00828673504292965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.038108546286821365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897688448429108 + ], + "opposite_logprob_tokens": [ + -2.57897686958313 + ], + "logprob_reward": -0.07897688448429108, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897688448429108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06787367165088654, + 0.042424459010362625, + 0.042018912732601166, + 0.06502766907215118, + 0.034360602498054504, + 0.09503291547298431, + 0.14189931750297546, + 0.09487949311733246, + 0.12576726078987122, + 0.29071563482284546 + ], + "credit_map": [ + 0.004097374156117439, + 0.00393860274925828, + 0.021385494619607925, + 0.0038676769472658634, + 0.006317092105746269, + 0.004606312606483698, + 0.0033246930688619614, + 0.02033642679452896, + 0.0029077285435050726, + 0.0057771955616772175, + 0.006886574439704418, + 0.010069583542644978, + 0.00413740798830986, + 0.004948476329445839, + 0.0039077261462807655, + 0.003789772279560566, + 0.003587129758670926, + 0.008428910747170448, + 0.005493967793881893, + 0.007554725278168917, + 0.0050798808224499226, + 0.00488681485876441, + 0.003157164203003049, + 0.0038303215987980366, + 0.005915180779993534, + 0.005373012274503708, + 0.01450658030807972, + 0.012508081272244453, + 0.0043859020806849, + 0.004811626859009266, + 0.004065588116645813, + 0.0134617043659091, + 0.007584217470139265, + 0.005536986980587244, + 0.004171990789473057, + 0.00365054770372808, + 0.0031991731375455856, + 0.002673287410289049, + 0.0029710817616432905, + 0.004573319107294083, + 0.012587739154696465, + 0.009190895594656467, + 0.014794111251831055, + 0.010900826193392277, + 0.00847913883626461, + 0.007149913348257542, + 0.009092122316360474, + 0.022838173434138298, + 0.013877108693122864, + 0.02113192528486252, + 0.017821578308939934, + 0.011409983970224857, + 0.020534291863441467, + 0.01839352585375309, + 0.019252588972449303, + 0.019478317350149155, + 0.01324745174497366, + 0.010284892283380032, + 0.011163179762661457, + 0.009383555501699448, + 0.01380737591534853, + 0.015228820033371449, + 0.008097240701317787, + 0.013666984625160694, + 0.013204746879637241, + 0.012878824025392532, + 0.016494816169142723, + 0.012658319436013699, + 0.018372105434536934, + 0.019880658015608788, + 0.010236633010208607, + 0.022041169926524162, + 0.01897910237312317, + 0.020413389429450035, + 0.08459971100091934, + 0.05513932555913925, + 0.015047896653413773, + 0.07807331532239914, + 0.010136446915566921, + 0.008326457813382149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.07897688448429108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene begins with a slow, graceful crane shot descending from a wide angle to a medium shot, intimately depicting a young man and woman. They glide gracefully across a vast outdoor ice rink, bathed in the warm, golden glow of the afternoon sun. The man, with windswept black hair, wears a dark gray cardigan and dark pants; a gentle, adoring smile graces his face as he gazes upon his partner. The woman, her long blond hair flowing freely behind her, wears a cream turtleneck sweater and a flowing red skating skirt. Her cheeks are flushed by the cold, and her bright, warm smile mirrors his. Hand in hand, their movements perfectly synchronized, they carve a smooth, sweeping arc across the pristine ice, their skates stirring up a fine, shimmering mist. The background is a soft-focus winter wonderland, with snow-capped pine trees and the distant, blurred movements of the other skaters. The low-angle sunlight creates long, soft shadows and beautiful, cinematic lens flares, enriching the scene’s warm tones of gold, red, and dark gray, and contributing to an atmosphere of profound romance, tranquility, and joy.", + "seed": 18458, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_5/index_8696_seed_18458_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": -0.12613806426525115, + "vlm_logprob": -0.12613806426525115, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.12613806426525115, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5643054842948914 + ], + "opposite_logprob_tokens": [ + -0.9393054842948914 + ], + "logprob_reward": -0.5643054842948914, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5643054842948914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15589168667793274, + 0.04804668202996254, + 0.18078938126564026, + 0.0747118815779686, + 0.046115659177303314, + 0.09706727415323257, + 0.13686619699001312, + 0.042777083814144135, + 0.18414542078971863, + 0.033588699996471405 + ], + "credit_map": [ + 0.019717151299118996, + 0.018992222845554352, + 0.021817844361066818, + 0.024110138416290283, + 0.019347310066223145, + 0.01868266984820366, + 0.01189151220023632, + 0.021332843229174614, + 0.006631738506257534, + 0.005634968169033527, + 0.006871296092867851, + 0.005875397473573685, + 0.005327695515006781, + 0.005379976239055395, + 0.00377152138389647, + 0.008554087951779366, + 0.02879849076271057, + 0.017155492678284645, + 0.03265710547566414, + 0.011257828213274479, + 0.014177500270307064, + 0.024146664887666702, + 0.012622048147022724, + 0.03997424989938736, + 0.005579384975135326, + 0.005709455348551273, + 0.017460988834500313, + 0.0031232943292707205, + 0.006296977866441011, + 0.006937824189662933, + 0.006703075487166643, + 0.022900883108377457, + 0.007800278719514608, + 0.004239728674292564, + 0.006364600732922554, + 0.00708672683686018, + 0.004013053607195616, + 0.004831389058381319, + 0.004047269467264414, + 0.007732614409178495, + 0.006733485963195562, + 0.007992583326995373, + 0.013055487535893917, + 0.01703539676964283, + 0.01986726187169552, + 0.009011488407850266, + 0.0073129176162183285, + 0.016058648005127907, + 0.018465183675289154, + 0.039982251822948456, + 0.015621410682797432, + 0.014972223900258541, + 0.01678330823779106, + 0.015593797899782658, + 0.007049906998872757, + 0.008398103527724743, + 0.00363176385872066, + 0.005941655486822128, + 0.005920378491282463, + 0.0088988421484828, + 0.004789587575942278, + 0.004475669469684362, + 0.00416780449450016, + 0.004951381124556065, + 0.02709975279867649, + 0.03511931374669075, + 0.024274460971355438, + 0.03460129722952843, + 0.016518326476216316, + 0.014251167885959148, + 0.01890164613723755, + 0.013379462994635105, + 0.004614938981831074, + 0.004163801204413176, + 0.0039291284047067165, + 0.004752707201987505, + 0.004116373602300882, + 0.0037084852810949087, + 0.0029304164927452803, + 0.0053728497587144375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.5643054842948914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823429882526398 + ], + "opposite_logprob_tokens": [ + -3.5298235416412354 + ], + "logprob_reward": -0.029823429882526398, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823429882526398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1422404795885086, + 0.08259012550115585, + 0.3687673509120941, + 0.09705010801553726, + 0.041745468974113464, + 0.10483791679143906, + 0.04942636936903, + 0.03914942592382431, + 0.042408403009176254, + 0.03178432956337929 + ], + "credit_map": [ + 0.01365402527153492, + 0.013591007329523563, + 0.01825914904475212, + 0.022786570712924004, + 0.021659480407834053, + 0.016384588554501534, + 0.014785542152822018, + 0.021120112389326096, + 0.007653587963432074, + 0.008765597827732563, + 0.012921618297696114, + 0.010872569866478443, + 0.00752559257671237, + 0.01205985527485609, + 0.005634370259940624, + 0.017156932502985, + 0.043508000671863556, + 0.03851981461048126, + 0.0745987743139267, + 0.03315553814172745, + 0.03825456649065018, + 0.051405634731054306, + 0.025061504915356636, + 0.06426352262496948, + 0.006705353036522865, + 0.009298402816057205, + 0.022704705595970154, + 0.006211276166141033, + 0.009183457121253014, + 0.00760707026347518, + 0.005611429922282696, + 0.029728412628173828, + 0.004008862655609846, + 0.00459718145430088, + 0.00594285735860467, + 0.006799034308642149, + 0.004670776426792145, + 0.0049809967167675495, + 0.0038724446203559637, + 0.006873312406241894, + 0.007993243634700775, + 0.009913809597492218, + 0.01703234575688839, + 0.01263067964464426, + 0.02157849259674549, + 0.00961229857057333, + 0.007619453594088554, + 0.01845759153366089, + 0.00406630989164114, + 0.0055755809880793095, + 0.0076116761192679405, + 0.005590988788753748, + 0.008975610136985779, + 0.005056507885456085, + 0.005141782108694315, + 0.00740791205316782, + 0.0045650433748960495, + 0.004726869985461235, + 0.005883597303181887, + 0.008084416389465332, + 0.005075679160654545, + 0.004421353340148926, + 0.003031097585335374, + 0.0033613683190196753, + 0.004891803488135338, + 0.008872314356267452, + 0.0054924022406339645, + 0.007205227855592966, + 0.0043976791203022, + 0.004314139019697905, + 0.003754209727048874, + 0.003480629064142704, + 0.00455161277204752, + 0.005190418567508459, + 0.004082533530890942, + 0.005106640513986349, + 0.004153026733547449, + 0.003468380542472005, + 0.0023116853553801775, + 0.0029200285207480192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.029823429882526398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23002515733242035, + 0.0830472782254219, + 0.22569572925567627, + 0.1093640923500061, + 0.04240449517965317, + 0.07397276908159256, + 0.08760448545217514, + 0.05302282050251961, + 0.06480895727872849, + 0.030054211616516113 + ], + "credit_map": [ + 0.03199378401041031, + 0.03156205266714096, + 0.02801552414894104, + 0.03358563035726547, + 0.03799917921423912, + 0.028902661055326462, + 0.01688009314239025, + 0.021086227148771286, + 0.008531919680535793, + 0.008830895647406578, + 0.010526399128139019, + 0.012348147109150887, + 0.014842712320387363, + 0.009890041314065456, + 0.006617409642785788, + 0.011459744535386562, + 0.0198116023093462, + 0.019772609695792198, + 0.03897954523563385, + 0.032953571528196335, + 0.03136621043086052, + 0.021294383332133293, + 0.019244195893406868, + 0.042273614555597305, + 0.008532265201210976, + 0.008523211814463139, + 0.026873445138335228, + 0.005668906960636377, + 0.009386453777551651, + 0.008084609173238277, + 0.01055049803107977, + 0.031744688749313354, + 0.0038835755549371243, + 0.005432596895843744, + 0.0062820990569889545, + 0.00508071668446064, + 0.004572301171720028, + 0.006782872602343559, + 0.004188884980976582, + 0.006181441713124514, + 0.007249512709677219, + 0.007390758953988552, + 0.011231201700866222, + 0.009444531053304672, + 0.010649441741406918, + 0.010068877600133419, + 0.0064825392328202724, + 0.0114558981731534, + 0.009022694081068039, + 0.012461568228900433, + 0.012275406159460545, + 0.014408581890165806, + 0.011968836188316345, + 0.008956541307270527, + 0.008195065893232822, + 0.01031578704714775, + 0.006466837599873543, + 0.005398509558290243, + 0.008366655558347702, + 0.010927734896540642, + 0.005386119242757559, + 0.005200724117457867, + 0.004753402434289455, + 0.006522834300994873, + 0.007629173342138529, + 0.007230872288346291, + 0.008780047297477722, + 0.01328691653907299, + 0.00734350923448801, + 0.005501426290720701, + 0.010621102526783943, + 0.0044159009121358395, + 0.003666214644908905, + 0.003931254148483276, + 0.0031268345192074776, + 0.004386803135275841, + 0.0051568313501775265, + 0.003947680350393057, + 0.0023139382246881723, + 0.0035246533807367086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640049956738949 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008640049956738949, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640049956738949, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1384904682636261, + 0.10819095373153687, + 0.09155390411615372, + 0.07573021948337555, + 0.10650079697370529, + 0.15403057634830475, + 0.07341741025447845, + 0.0823448970913887, + 0.1295788735151291, + 0.040161944925785065 + ], + "credit_map": [ + 0.02256012335419655, + 0.014517302624881268, + 0.016701221466064453, + 0.01685572788119316, + 0.017520977184176445, + 0.017334619536995888, + 0.013247696682810783, + 0.019752804189920425, + 0.013227243907749653, + 0.010310047306120396, + 0.015589871443808079, + 0.015099314041435719, + 0.01320524699985981, + 0.010899706743657589, + 0.00887775607407093, + 0.020981764420866966, + 0.01235591433942318, + 0.008740692399442196, + 0.015154460445046425, + 0.009798722341656685, + 0.008134793490171432, + 0.010569337755441666, + 0.008198165334761143, + 0.01860181614756584, + 0.006273207720369101, + 0.015320507809519768, + 0.010024909861385822, + 0.006052431184798479, + 0.010679560713469982, + 0.006274438463151455, + 0.009235135279595852, + 0.011870026588439941, + 0.00944132823497057, + 0.011641396209597588, + 0.017097000032663345, + 0.019415702670812607, + 0.009557326324284077, + 0.014418465085327625, + 0.009088197723031044, + 0.015841376036405563, + 0.009658992290496826, + 0.01161831896752119, + 0.026552695780992508, + 0.03153059631586075, + 0.022736741229891777, + 0.010229116305708885, + 0.011826558038592339, + 0.0298775602132082, + 0.007995869033038616, + 0.0130632808431983, + 0.013752428814768791, + 0.014801176264882088, + 0.005852028261870146, + 0.005341568496078253, + 0.004416620824486017, + 0.008194437250494957, + 0.008631827309727669, + 0.009555243887007236, + 0.010348492302000523, + 0.01448938250541687, + 0.010966329835355282, + 0.010661936365067959, + 0.009040276519954205, + 0.008651413023471832, + 0.013059550896286964, + 0.01759057305753231, + 0.015777720138430595, + 0.022996488958597183, + 0.01640019379556179, + 0.010240250267088413, + 0.022754771634936333, + 0.010759316384792328, + 0.004571794532239437, + 0.004998229444026947, + 0.0039364066906273365, + 0.004027220886200666, + 0.00742129422724247, + 0.003996691200882196, + 0.005175570026040077, + 0.006034737452864647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.008640049956738949 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" 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 camera start with a wide-angle shot and slowly descend to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the main subjects a young man with black hair and a woman with long blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman wearing a cream turtleneck sweater and a flowing red skating skirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the skaters carving a smooth, sweeping arc across a vast outdoor ice rink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft-focus winter wonderland with snow-capped pine trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327561192214489 + ], + "opposite_logprob_tokens": [ + -3.77327561378479 + ], + "logprob_reward": -0.02327561192214489, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327561192214489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2751869261264801, + 0.1570475548505783, + 0.10311143100261688, + 0.06410830467939377, + 0.06589943170547485, + 0.13947156071662903, + 0.09077528864145279, + 0.05953025072813034, + 0.0265507809817791, + 0.018318455666303635 + ], + "credit_map": [ + 0.05006212368607521, + 0.03748713433742523, + 0.029671242460608482, + 0.03344780206680298, + 0.03841209411621094, + 0.03927518054842949, + 0.019290369004011154, + 0.027540989220142365, + 0.019969619810581207, + 0.018857643008232117, + 0.022130126133561134, + 0.01904144696891308, + 0.016608530655503273, + 0.016843752935528755, + 0.009815705008804798, + 0.03378073498606682, + 0.016442548483610153, + 0.010411091148853302, + 0.015597967430949211, + 0.0056877704337239265, + 0.008272521197795868, + 0.024281803518533707, + 0.005754532292485237, + 0.016663188114762306, + 0.005263784900307655, + 0.004482100252062082, + 0.012803459540009499, + 0.0031841034069657326, + 0.005790029652416706, + 0.006720868870615959, + 0.006745894905179739, + 0.019118066877126694, + 0.004731920547783375, + 0.009394531138241291, + 0.010649616830050945, + 0.00714506721124053, + 0.006137556862086058, + 0.011778203770518303, + 0.006095194723457098, + 0.009967338293790817, + 0.01284593716263771, + 0.010826925747096539, + 0.01889835111796856, + 0.022509032860398293, + 0.020610593259334564, + 0.01783663220703602, + 0.00932188518345356, + 0.02662220038473606, + 0.011386116035282612, + 0.016808142885565758, + 0.016624070703983307, + 0.017248710617423058, + 0.006076662801206112, + 0.009897200390696526, + 0.0050351121462881565, + 0.007699280045926571, + 0.008479133248329163, + 0.007623243611305952, + 0.009082760661840439, + 0.013301391154527664, + 0.006434460170567036, + 0.00456601707264781, + 0.0040699089877307415, + 0.0059733339585363865, + 0.0038268661592155695, + 0.005389688536524773, + 0.003075060434639454, + 0.003961961250752211, + 0.0029908556025475264, + 0.002323521999642253, + 0.0026440825313329697, + 0.002338743768632412, + 0.0027387505397200584, + 0.0030239580664783716, + 0.0020939509849995375, + 0.0021353198681026697, + 0.002278206404298544, + 0.0018167456146329641, + 0.001557328854687512, + 0.00267419614829123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02327561192214489 + } + ], + "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/0000158/group_6/index_tensor([6311, 6311])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_tensor([6311, 6311])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..759658ec5eb45433f3bbe04c6d6c4e0f9c4f6473 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_tensor([6311, 6311])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly high-angled medium shot frames an ornate, dark mahogany desk, its polished surface gleaming in the focused, warm glow of a single green bank lamp in a dimly lit, wood-paneled study. The stillness of the scene is suddenly shattered as the desk begins to tremble with an almost imperceptible, low vibration, causing the fountain pen in its brass holder to rattle softly. This initial tremor escalates rapidly and violently into a convulsive earthquake, the entire heavy wooden frame rocking back and forth with unnatural force. Loose, previously neatly arranged sheets of handwritten parchment and modern typed documents are thrown upward and begin to flutter chaotically, scattering across the desk's surface and drifting out of frame to the floor. The unbroken shot tensely captures the escalating chaos, capturing the moment when the antique glass inkwell tilts and spills a pool of deep black ink, which spreads across the polished wood like a creeping stain, creating a stark, dramatic contrast that amplifies the scene's intensifying, baffling horror.", + "seed": 6539, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6539_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": -0.012188604287803172, + "vlm_logprob": -0.012188604287803172, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.012188604287803172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010596439242362976 + ], + "opposite_logprob_tokens": [ + -5.13559627532959 + ], + "logprob_reward": -0.010596439242362976, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.010596439242362976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1483837217092514, + 0.065774105489254, + 0.059521354734897614, + 0.06704167276620865, + 0.07763143628835678, + 0.06085963547229767, + 0.1039152443408966, + 0.11120857298374176, + 0.11014225333929062, + 0.19552206993103027 + ], + "credit_map": [ + 0.0307314395904541, + 0.019512034952640533, + 0.013116592541337013, + 0.022323351353406906, + 0.01944885030388832, + 0.018175344914197922, + 0.014212694950401783, + 0.01086342427879572, + 0.010942121036350727, + 0.009295938536524773, + 0.004778326489031315, + 0.006057690363377333, + 0.010080521926283836, + 0.009977663867175579, + 0.007239903789013624, + 0.007401940878480673, + 0.009862762875854969, + 0.009731346741318703, + 0.004298764746636152, + 0.004472578410059214, + 0.011665016412734985, + 0.008153361268341541, + 0.004970581270754337, + 0.006366944871842861, + 0.011135094799101353, + 0.021033406257629395, + 0.004035342019051313, + 0.003944908734411001, + 0.003595991525799036, + 0.005858176853507757, + 0.0031387172639369965, + 0.014300032518804073, + 0.012487753294408321, + 0.006887787953019142, + 0.009187977761030197, + 0.007918012328445911, + 0.005969188641756773, + 0.011810208670794964, + 0.0086758341640234, + 0.014694679528474808, + 0.007149083539843559, + 0.004505758639425039, + 0.0077941156923770905, + 0.00910007394850254, + 0.005940080154687166, + 0.010990087874233723, + 0.007635222747921944, + 0.00774521054700017, + 0.009336504153907299, + 0.010417389683425426, + 0.017236851155757904, + 0.013965833932161331, + 0.010138297453522682, + 0.012349342927336693, + 0.016400109976530075, + 0.014070912264287472, + 0.020718080922961235, + 0.010594720020890236, + 0.015419204719364643, + 0.011042483150959015, + 0.008485069498419762, + 0.008354540914297104, + 0.013198702596127987, + 0.023395776748657227, + 0.014708983711898327, + 0.0182714331895113, + 0.02326906844973564, + 0.017376352101564407, + 0.011057974770665169, + 0.005538885947316885, + 0.011677932925522327, + 0.008241619914770126, + 0.027660390362143517, + 0.027425037696957588, + 0.03281111270189285, + 0.029347464442253113, + 0.022841086611151695, + 0.014868594706058502, + 0.02070743218064308, + 0.019860941916704178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.010596439242362976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012579617090523243 + ], + "opposite_logprob_tokens": [ + -4.387579441070557 + ], + "logprob_reward": -0.012579617090523243, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012579617090523243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09398995339870453, + 0.113250732421875, + 0.12091047316789627, + 0.12629610300064087, + 0.1846143901348114, + 0.04715276509523392, + 0.04160510376095772, + 0.13841556012630463, + 0.05348750948905945, + 0.08027739077806473 + ], + "credit_map": [ + 0.017337825149297714, + 0.012822059914469719, + 0.007053903304040432, + 0.007951394654810429, + 0.015882819890975952, + 0.012608298100531101, + 0.009439830668270588, + 0.010893815197050571, + 0.0256344024091959, + 0.019311774522066116, + 0.005633960012346506, + 0.006145354826003313, + 0.011211990378797054, + 0.01723518595099449, + 0.01318130549043417, + 0.01489674299955368, + 0.021265270188450813, + 0.019092721864581108, + 0.0058784144930541515, + 0.007030939683318138, + 0.027956169098615646, + 0.018107324838638306, + 0.007217378355562687, + 0.014362243935465813, + 0.01730855368077755, + 0.04073521867394447, + 0.00489458441734314, + 0.004552204627543688, + 0.012794438749551773, + 0.01092134602367878, + 0.005125753115862608, + 0.029963994398713112, + 0.03518517687916756, + 0.012808240950107574, + 0.01043407991528511, + 0.01370299980044365, + 0.01972545124590397, + 0.022040486335754395, + 0.03103821910917759, + 0.03967972472310066, + 0.006242708768695593, + 0.003367628436535597, + 0.007876183837652206, + 0.006746845319867134, + 0.004284042399376631, + 0.006703737657517195, + 0.005448014475405216, + 0.006483599077910185, + 0.00551783200353384, + 0.004326578229665756, + 0.006143373437225819, + 0.006258446257561445, + 0.004064237233251333, + 0.003089164150878787, + 0.006877074483782053, + 0.005328397732228041, + 0.03708811476826668, + 0.011325539089739323, + 0.008566062897443771, + 0.013670223765075207, + 0.007513352204114199, + 0.007035659626126289, + 0.013161868788301945, + 0.04005473852157593, + 0.006134577561169863, + 0.008758321404457092, + 0.010513133369386196, + 0.008733074180781841, + 0.003930483013391495, + 0.004762593191117048, + 0.005165778566151857, + 0.005489544942975044, + 0.01105941366404295, + 0.011592098511755466, + 0.009591218084096909, + 0.014794392511248589, + 0.010239437222480774, + 0.00637042848393321, + 0.009050337597727776, + 0.0075800600461661816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.012579617090523243 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191400915384293 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.018191400915384293, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191400915384293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10653012245893478, + 0.05585130676627159, + 0.10120649635791779, + 0.04037678986787796, + 0.1575680822134018, + 0.11086096614599228, + 0.07439979165792465, + 0.09152957797050476, + 0.17985112965106964, + 0.08182574808597565 + ], + "credit_map": [ + 0.03217105567455292, + 0.019287914037704468, + 0.007114246021956205, + 0.008373002521693707, + 0.010015234351158142, + 0.011926385574042797, + 0.006447548978030682, + 0.011194722726941109, + 0.008651819080114365, + 0.008646580390632153, + 0.0037217268254607916, + 0.004002534784376621, + 0.005500491708517075, + 0.011504370719194412, + 0.0047339582815766335, + 0.009089816361665726, + 0.014019954949617386, + 0.016956143081188202, + 0.009792715311050415, + 0.008383422158658504, + 0.009790698066353798, + 0.00906455609947443, + 0.012581877410411835, + 0.02061712183058262, + 0.008084378205239773, + 0.006507198791950941, + 0.003928788471966982, + 0.0028639601077884436, + 0.004588947165757418, + 0.004910836927592754, + 0.00337041518650949, + 0.006122262217104435, + 0.030303528532385826, + 0.013248836621642113, + 0.016048340126872063, + 0.013948986306786537, + 0.016992930322885513, + 0.020747214555740356, + 0.017533866688609123, + 0.028744373470544815, + 0.012638336047530174, + 0.006552783772349358, + 0.0173901729285717, + 0.02231091447174549, + 0.007633716333657503, + 0.01093376986682415, + 0.018398761749267578, + 0.015002509579062462, + 0.005844018887728453, + 0.0046936385333538055, + 0.013071810826659203, + 0.013290398754179478, + 0.0071196602657437325, + 0.007726692128926516, + 0.011260593309998512, + 0.011392972432076931, + 0.019045500084757805, + 0.008985940366983414, + 0.01236006524413824, + 0.005606520920991898, + 0.005557496566325426, + 0.007169388700276613, + 0.012817640788853168, + 0.019987015053629875, + 0.018687108531594276, + 0.028042618185281754, + 0.04499661177396774, + 0.03342246264219284, + 0.011594912968575954, + 0.010928425006568432, + 0.016023822128772736, + 0.016155147925019264, + 0.008276480250060558, + 0.012408341281116009, + 0.009015997871756554, + 0.016284774988889694, + 0.012113026343286037, + 0.005516916513442993, + 0.010129881091415882, + 0.008080324158072472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.018191400915384293 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.06389863044023514, + 0.04623414948582649, + 0.06915217638015747, + 0.10913177579641342, + 0.13749650120735168, + 0.12124746292829514, + 0.03706469386816025, + 0.17820894718170166, + 0.09543957561254501, + 0.14212609827518463 + ], + "credit_map": [ + 0.010710559785366058, + 0.008807731792330742, + 0.007130297366529703, + 0.010421784594655037, + 0.006865167990326881, + 0.007112112827599049, + 0.007384233642369509, + 0.005466741044074297, + 0.008580578491091728, + 0.006545275449752808, + 0.004150568973273039, + 0.006666242145001888, + 0.005410556681454182, + 0.0038938368670642376, + 0.0055702910758554935, + 0.0054167951457202435, + 0.011749405413866043, + 0.011619052849709988, + 0.005001118406653404, + 0.004216516390442848, + 0.005848629400134087, + 0.005349595099687576, + 0.006710129790008068, + 0.018657727167010307, + 0.016596894711256027, + 0.0205690935254097, + 0.010053730569779873, + 0.010409333743155003, + 0.00864294171333313, + 0.01291428692638874, + 0.008258816786110401, + 0.021686669439077377, + 0.023580245673656464, + 0.013218692503869534, + 0.014318304136395454, + 0.013217846862971783, + 0.013527335599064827, + 0.018339021131396294, + 0.017730632796883583, + 0.023564422503113747, + 0.012948783114552498, + 0.006949160248041153, + 0.025624219328165054, + 0.02243807725608349, + 0.007649844046682119, + 0.012499865144491196, + 0.01368474680930376, + 0.019452763721346855, + 0.005026970058679581, + 0.0026766699738800526, + 0.005060703493654728, + 0.005223297979682684, + 0.003615014487877488, + 0.0038606123998761177, + 0.006441186647862196, + 0.005160236265510321, + 0.052193012088537216, + 0.013464899733662605, + 0.02023930475115776, + 0.010476484894752502, + 0.0075492518953979015, + 0.009743757545948029, + 0.01107726339250803, + 0.05346496030688286, + 0.014212644658982754, + 0.016157187521457672, + 0.01826229877769947, + 0.010495525784790516, + 0.007980523630976677, + 0.006013845093548298, + 0.010917390696704388, + 0.011400147341191769, + 0.01395732257515192, + 0.024783890694379807, + 0.01505372952669859, + 0.02502835914492607, + 0.028827371075749397, + 0.009968203492462635, + 0.014349274337291718, + 0.010157939977943897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097876638174057 + ], + "opposite_logprob_tokens": [ + -4.634787559509277 + ], + "logprob_reward": -0.0097876638174057, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097876638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06774978339672089, + 0.0778975635766983, + 0.05518071725964546, + 0.04439052194356918, + 0.11032678931951523, + 0.13211466372013092, + 0.12218921631574631, + 0.25513768196105957, + 0.060321640223264694, + 0.07469142228364944 + ], + "credit_map": [ + 0.01129384245723486, + 0.008875373750925064, + 0.00777721032500267, + 0.011851447634398937, + 0.00822334736585617, + 0.005893465597182512, + 0.0074303424917161465, + 0.0064047579653561115, + 0.022224964573979378, + 0.014195743016898632, + 0.004069200716912746, + 0.004770913626998663, + 0.00605063745751977, + 0.00775873102247715, + 0.007381870411336422, + 0.01144550833851099, + 0.008590255863964558, + 0.007637673057615757, + 0.006591134238988161, + 0.0056762536987662315, + 0.00682086544111371, + 0.007382259704172611, + 0.006213315762579441, + 0.006268959958106279, + 0.007620488293468952, + 0.005324048455804586, + 0.004487604834139347, + 0.00456121563911438, + 0.003985923249274492, + 0.0054786293767392635, + 0.004343304317444563, + 0.008589310571551323, + 0.020654115825891495, + 0.012043456546962261, + 0.012984596192836761, + 0.011295605450868607, + 0.007535281591117382, + 0.011985068209469318, + 0.010066927410662174, + 0.023761741816997528, + 0.014369551092386246, + 0.007740713190287352, + 0.023220622912049294, + 0.030932988971471786, + 0.007866477593779564, + 0.013448831625282764, + 0.01595556177198887, + 0.018579905852675438, + 0.016070760786533356, + 0.011031311005353928, + 0.01735573075711727, + 0.019514895975589752, + 0.013520575128495693, + 0.013806370086967945, + 0.014420118182897568, + 0.016469450667500496, + 0.06022019684314728, + 0.021545767784118652, + 0.025258442386984825, + 0.030148467049002647, + 0.01439219992607832, + 0.014430839568376541, + 0.024993976578116417, + 0.06414778530597687, + 0.009864444844424725, + 0.009580982849001884, + 0.012246502563357353, + 0.009759459644556046, + 0.004118815530091524, + 0.0028644839767366648, + 0.005814519710838795, + 0.0060724299401044846, + 0.008281193673610687, + 0.012320132926106453, + 0.00843559019267559, + 0.010141994804143906, + 0.01591397263109684, + 0.004939493257552385, + 0.00699443556368351, + 0.0076646045781672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.0097876638174057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames an ornate, dark mahogany desk, its polished surface gleaming in the focused, warm glow of a single green bank lamp in a dimly lit, wood-paneled study. The stillness of the scene is suddenly shattered as the desk begins to tremble with an almost imperceptible, low vibration, causing the fountain pen in its brass holder to rattle softly. This initial tremor escalates rapidly and violently into a convulsive earthquake, the entire heavy wooden frame rocking back and forth with unnatural force. Loose, previously neatly arranged sheets of handwritten parchment and modern typed documents are thrown upward and begin to flutter chaotically, scattering across the desk's surface and drifting out of frame to the floor. The unbroken shot tensely captures the escalating chaos, capturing the moment when the antique glass inkwell tilts and spills a pool of deep black ink, which spreads across the polished wood like a creeping stain, creating a stark, dramatic contrast that amplifies the scene's intensifying, baffling horror.", + "seed": 6540, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6540_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": -0.030055138282477856, + "vlm_logprob": -0.030055138282477856, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.030055138282477856, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013903080485761166 + ], + "opposite_logprob_tokens": [ + -4.7639031410217285 + ], + "logprob_reward": -0.013903080485761166, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013903080485761166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1689983755350113, + 0.04157773032784462, + 0.039247285574674606, + 0.11384660750627518, + 0.09507086873054504, + 0.12275081127882004, + 0.11388035118579865, + 0.14414507150650024, + 0.04883154109120369, + 0.11165130883455276 + ], + "credit_map": [ + 0.02794121392071247, + 0.01438995636999607, + 0.022850826382637024, + 0.023122794926166534, + 0.014031064696609974, + 0.019495265558362007, + 0.01544799841940403, + 0.03171925246715546, + 0.005527875851839781, + 0.004945501685142517, + 0.006804315838962793, + 0.004883730784058571, + 0.004573781043291092, + 0.003420608351007104, + 0.005803384818136692, + 0.00561853451654315, + 0.010030103847384453, + 0.003720537992194295, + 0.0037721346598118544, + 0.0038032501470297575, + 0.0037312658969312906, + 0.003369938815012574, + 0.004904701374471188, + 0.005915354005992413, + 0.0176588986068964, + 0.018496304750442505, + 0.012911955825984478, + 0.014028779231011868, + 0.016456691548228264, + 0.008820325136184692, + 0.013440984301269054, + 0.012032668106257915, + 0.015816600993275642, + 0.015364918857812881, + 0.005153259262442589, + 0.005690043326467276, + 0.009411723352968693, + 0.028051599860191345, + 0.00743807153776288, + 0.008144654333591461, + 0.008540153503417969, + 0.02208740822970867, + 0.029448620975017548, + 0.011743667535483837, + 0.009130747988820076, + 0.009970681741833687, + 0.008444325998425484, + 0.023385202512145042, + 0.01603325642645359, + 0.016636930406093597, + 0.016972554847598076, + 0.015133016742765903, + 0.023211421445012093, + 0.010817223228514194, + 0.005821750499308109, + 0.00925420317798853, + 0.020141372457146645, + 0.03254736587405205, + 0.021840931847691536, + 0.01184008177369833, + 0.021383678540587425, + 0.012751009315252304, + 0.013147308491170406, + 0.010493313893675804, + 0.0066565778106451035, + 0.009129582904279232, + 0.0050852978602051735, + 0.006349479779601097, + 0.005935774650424719, + 0.004420448560267687, + 0.006276319269090891, + 0.004978061653673649, + 0.01482511218637228, + 0.03267112374305725, + 0.02339817024767399, + 0.009364217519760132, + 0.008441530168056488, + 0.009779807180166245, + 0.006856657564640045, + 0.006314692087471485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.013903080485761166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816511854529381 + ], + "opposite_logprob_tokens": [ + -3.2881650924682617 + ], + "logprob_reward": -0.03816511854529381, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816511854529381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09780943393707275, + 0.0680224671959877, + 0.07275143265724182, + 0.13705182075500488, + 0.08622314035892487, + 0.03940347582101822, + 0.1173064336180687, + 0.19722311198711395, + 0.059252120554447174, + 0.12495657801628113 + ], + "credit_map": [ + 0.013625911436975002, + 0.014479691162705421, + 0.010726808570325375, + 0.014843631535768509, + 0.008833999745547771, + 0.008111514151096344, + 0.009416796267032623, + 0.017771083861589432, + 0.01731320656836033, + 0.010991550981998444, + 0.00837782584130764, + 0.004975592251867056, + 0.00614094128832221, + 0.004094576463103294, + 0.006139541510492563, + 0.009989232756197453, + 0.024272773414850235, + 0.006736520677804947, + 0.006249197293072939, + 0.005167789291590452, + 0.006976556498557329, + 0.005473259370774031, + 0.0073405723087489605, + 0.010534761473536491, + 0.019804412499070168, + 0.024341128766536713, + 0.014158890582621098, + 0.01274914387613535, + 0.01558451633900404, + 0.009350971318781376, + 0.02480260469019413, + 0.01626015082001686, + 0.0138113833963871, + 0.013914460316300392, + 0.0039936755783855915, + 0.004940450191497803, + 0.008001190610229969, + 0.02707553468644619, + 0.006735527887940407, + 0.007750911172479391, + 0.0048527843318879604, + 0.00403588218614459, + 0.006053732708096504, + 0.004281213507056236, + 0.004519792273640633, + 0.005241082515567541, + 0.005105722229927778, + 0.005313267465680838, + 0.018749982118606567, + 0.018631011247634888, + 0.013324780389666557, + 0.01754705421626568, + 0.021232040598988533, + 0.01276955846697092, + 0.007760492153465748, + 0.007291513495147228, + 0.028072375804185867, + 0.044926516711711884, + 0.02671956457197666, + 0.01994592510163784, + 0.02875799871981144, + 0.020665351301431656, + 0.015929535031318665, + 0.012205848470330238, + 0.007843189872801304, + 0.010554037056863308, + 0.007276125252246857, + 0.007797846570611, + 0.009984028525650501, + 0.006343233399093151, + 0.005364610813558102, + 0.0040890490636229515, + 0.02405175007879734, + 0.03480058163404465, + 0.01692081242799759, + 0.011032731272280216, + 0.011643213219940662, + 0.011751837097108364, + 0.0071863168850541115, + 0.007569336798042059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.03816511854529381 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811795637011528 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811795637011528, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811795637011528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11841654777526855, + 0.04018855094909668, + 0.02656829170882702, + 0.13495315611362457, + 0.09797725826501846, + 0.1455344706773758, + 0.07558604329824448, + 0.12694893777370453, + 0.1339583694934845, + 0.09986835718154907 + ], + "credit_map": [ + 0.021748075261712074, + 0.010141587816178799, + 0.023195384070277214, + 0.018010247498750687, + 0.008275824598968029, + 0.009120462462306023, + 0.00909053161740303, + 0.0188344307243824, + 0.004982693120837212, + 0.004613794852048159, + 0.005987585987895727, + 0.00547836534678936, + 0.0038924533873796463, + 0.0030114378314465284, + 0.007962759584188461, + 0.004259458743035793, + 0.005216652527451515, + 0.002695114351809025, + 0.002667604712769389, + 0.0035863176453858614, + 0.0035329004749655724, + 0.0017857470083981752, + 0.00290270010009408, + 0.00418125418946147, + 0.018028954043984413, + 0.024500492960214615, + 0.015135983936488628, + 0.012667286209762096, + 0.014396250247955322, + 0.009320316836237907, + 0.023792563006281853, + 0.017111310735344887, + 0.016328178346157074, + 0.013831853866577148, + 0.00507492246106267, + 0.005693517159670591, + 0.012630373239517212, + 0.025813058018684387, + 0.008083267137408257, + 0.010522082448005676, + 0.015195883810520172, + 0.02344079688191414, + 0.03906446322798729, + 0.012836327776312828, + 0.011370046064257622, + 0.013467743992805481, + 0.0093988087028265, + 0.02076040767133236, + 0.00860287994146347, + 0.00947105884552002, + 0.012802292592823505, + 0.009207742288708687, + 0.013328462839126587, + 0.008188499137759209, + 0.0065753161907196045, + 0.007409791927784681, + 0.02027059905230999, + 0.02193477749824524, + 0.020778408274054527, + 0.016290292143821716, + 0.015297862701117992, + 0.011758111417293549, + 0.01249844953417778, + 0.00812042597681284, + 0.017755217850208282, + 0.015218062326312065, + 0.014194796793162823, + 0.01734866388142109, + 0.019599560648202896, + 0.013255609199404716, + 0.017000310122966766, + 0.01958615705370903, + 0.014309094287455082, + 0.01626945473253727, + 0.016167446970939636, + 0.012909938581287861, + 0.008637730032205582, + 0.011592802591621876, + 0.01129931304603815, + 0.00868257787078619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.03811795637011528 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369880884885788 + ], + "opposite_logprob_tokens": [ + -3.65136981010437 + ], + "logprob_reward": -0.026369880884885788, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369880884885788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0953124389052391, + 0.028456464409828186, + 0.019590985029935837, + 0.24449455738067627, + 0.0658923014998436, + 0.08244393765926361, + 0.08594394475221634, + 0.18841023743152618, + 0.10744591802358627, + 0.0820092186331749 + ], + "credit_map": [ + 0.017277603968977928, + 0.00644589401781559, + 0.01975548267364502, + 0.017277704551815987, + 0.006187812425196171, + 0.006805181968957186, + 0.008647216483950615, + 0.01291554607450962, + 0.00278116250410676, + 0.0032645699102431536, + 0.003415726125240326, + 0.0043534100987017155, + 0.003875207621604204, + 0.002589674200862646, + 0.004868398420512676, + 0.0033083143644034863, + 0.0048649194650352, + 0.0018063475145027041, + 0.002035589888691902, + 0.0024613169953227043, + 0.002163961762562394, + 0.0012472120579332113, + 0.0021569873206317425, + 0.0028546501416713, + 0.033722229301929474, + 0.04227764531970024, + 0.02852543815970421, + 0.024714509025216103, + 0.025230905041098595, + 0.015628838911652565, + 0.04490303620696068, + 0.029491961002349854, + 0.009736333973705769, + 0.00784615520387888, + 0.004383775405585766, + 0.005129092372953892, + 0.009186838753521442, + 0.016513332724571228, + 0.005968178156763315, + 0.007128594443202019, + 0.008798536844551563, + 0.009468285366892815, + 0.020998777821660042, + 0.007540725637227297, + 0.008279846049845219, + 0.007921637035906315, + 0.007206704933196306, + 0.012229429557919502, + 0.012757428921759129, + 0.008802462369203568, + 0.012618575245141983, + 0.014712885022163391, + 0.010514112189412117, + 0.009879597462713718, + 0.009322810918092728, + 0.007336067501455545, + 0.033120084553956985, + 0.029436443001031876, + 0.0413784496486187, + 0.024508047848939896, + 0.018608853220939636, + 0.012993743643164635, + 0.015486192889511585, + 0.012878432869911194, + 0.013751816004514694, + 0.010176953859627247, + 0.012842003256082535, + 0.013622039929032326, + 0.013414256274700165, + 0.009808002039790154, + 0.016080515459179878, + 0.017750324681401253, + 0.0065471879206597805, + 0.023186052218079567, + 0.021970674395561218, + 0.007722408510744572, + 0.005665776785463095, + 0.0065875486470758915, + 0.006498402450233698, + 0.0038311618845909834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.026369880884885788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371965512633324 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371965512633324, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371965512633324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05970029532909393, + 0.018916422501206398, + 0.04599333927035332, + 0.0835636779665947, + 0.23165373504161835, + 0.10497540235519409, + 0.08306492865085602, + 0.23744960129261017, + 0.06610515713691711, + 0.06857741624116898 + ], + "credit_map": [ + 0.012864259071648121, + 0.0047149015590548515, + 0.015312504954636097, + 0.005955833476036787, + 0.00421633617952466, + 0.004335280507802963, + 0.00417100777849555, + 0.008130175061523914, + 0.003248751861974597, + 0.0016279644332826138, + 0.0031723775900900364, + 0.003085693810135126, + 0.0017065772553905845, + 0.0011795309837907553, + 0.002374972216784954, + 0.002520556328818202, + 0.013961748220026493, + 0.003788364352658391, + 0.005039445124566555, + 0.004213619511574507, + 0.0037345746532082558, + 0.004721084143966436, + 0.00417628837749362, + 0.00635821558535099, + 0.013997159898281097, + 0.01918569952249527, + 0.008528712205588818, + 0.0070770615711808205, + 0.010998493060469627, + 0.005002859979867935, + 0.01243498269468546, + 0.006338716018944979, + 0.03708485886454582, + 0.03701961040496826, + 0.009675120934844017, + 0.013863260857760906, + 0.02449815161526203, + 0.07650131732225418, + 0.01542041264474392, + 0.017591018229722977, + 0.008699672296643257, + 0.015368234366178513, + 0.02078878879547119, + 0.010510378517210484, + 0.008269325830042362, + 0.00946307647973299, + 0.007765370886772871, + 0.024110564962029457, + 0.009302501566708088, + 0.013033038936555386, + 0.010236121714115143, + 0.013259819708764553, + 0.012638657353818417, + 0.007296787109225988, + 0.009765717200934887, + 0.0075322906486690044, + 0.044760555028915405, + 0.035288769751787186, + 0.054812561720609665, + 0.02526111528277397, + 0.02542613446712494, + 0.017603956162929535, + 0.02273443527519703, + 0.011562076397240162, + 0.008735520765185356, + 0.00955128576606512, + 0.0074267303571105, + 0.0085551543161273, + 0.010378342121839523, + 0.006680458318442106, + 0.008160808123648167, + 0.006616865284740925, + 0.009894540533423424, + 0.0172081608325243, + 0.011685408651828766, + 0.004800622351467609, + 0.007736656814813614, + 0.006937003694474697, + 0.004439779557287693, + 0.005875247064977884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.03371965512633324 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames an ornate, dark mahogany desk, its polished surface gleaming in the focused, warm glow of a single green bank lamp in a dimly lit, wood-paneled study. The stillness of the scene is suddenly shattered as the desk begins to tremble with an almost imperceptible, low vibration, causing the fountain pen in its brass holder to rattle softly. This initial tremor escalates rapidly and violently into a convulsive earthquake, the entire heavy wooden frame rocking back and forth with unnatural force. Loose, previously neatly arranged sheets of handwritten parchment and modern typed documents are thrown upward and begin to flutter chaotically, scattering across the desk's surface and drifting out of frame to the floor. The unbroken shot tensely captures the escalating chaos, capturing the moment when the antique glass inkwell tilts and spills a pool of deep black ink, which spreads across the polished wood like a creeping stain, creating a stark, dramatic contrast that amplifies the scene's intensifying, baffling horror.", + "seed": 6541, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6541_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": -0.007991287671029568, + "vlm_logprob": -0.007991287671029568, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.007991287671029568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007288177963346243 + ], + "opposite_logprob_tokens": [ + -5.757287979125977 + ], + "logprob_reward": -0.007288177963346243, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007288177963346243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0562191940844059, + 0.023674121126532555, + 0.06213947385549545, + 0.2728530168533325, + 0.07603013515472412, + 0.11292925477027893, + 0.0782705768942833, + 0.15693244338035583, + 0.10649140179157257, + 0.05446041747927666 + ], + "credit_map": [ + 0.0071905492804944515, + 0.006292333360761404, + 0.00625077448785305, + 0.006034723483026028, + 0.006711160764098167, + 0.006320292130112648, + 0.006762875244021416, + 0.010656480677425861, + 0.0024914327077567577, + 0.0029511740431189537, + 0.0019513326697051525, + 0.0021527716889977455, + 0.003901829943060875, + 0.003778257640078664, + 0.00295652961358428, + 0.003490790259093046, + 0.009012248367071152, + 0.0075349463149905205, + 0.009565912187099457, + 0.005534113850444555, + 0.0070341806858778, + 0.007678785361349583, + 0.0077278101816773415, + 0.008051472716033459, + 0.013127326034009457, + 0.008602611720561981, + 0.09663932025432587, + 0.018382590264081955, + 0.007448089774698019, + 0.011339695192873478, + 0.014502367936074734, + 0.10281100124120712, + 0.010867154225707054, + 0.009323909878730774, + 0.0075533087365329266, + 0.010787628591060638, + 0.00947485864162445, + 0.011740872636437416, + 0.00863848440349102, + 0.007643912918865681, + 0.016082774847745895, + 0.01672867313027382, + 0.0072094714269042015, + 0.018443822860717773, + 0.010236725211143494, + 0.013837119564414024, + 0.017764130607247353, + 0.012626520358026028, + 0.006952082738280296, + 0.005930161569267511, + 0.006923136301338673, + 0.015529190190136433, + 0.0057865954004228115, + 0.00877412036061287, + 0.010400049388408661, + 0.017975233495235443, + 0.021929465234279633, + 0.012015534564852715, + 0.017206531018018723, + 0.027615081518888474, + 0.012113386765122414, + 0.011850097216665745, + 0.010532118380069733, + 0.04367021843791008, + 0.008258123882114887, + 0.005969387013465166, + 0.018291890621185303, + 0.012352182529866695, + 0.019189810380339622, + 0.00949705671519041, + 0.008524066768586636, + 0.024408863857388496, + 0.010191191919147968, + 0.0045484136790037155, + 0.008207052014768124, + 0.00747779430821538, + 0.00635620579123497, + 0.004849298857152462, + 0.004977284464985132, + 0.007853172719478607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.007288177963346243 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008662977255880833 + ], + "opposite_logprob_tokens": [ + -4.758663177490234 + ], + "logprob_reward": -0.008662977255880833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008662977255880833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12513110041618347, + 0.06930523365736008, + 0.06256671249866486, + 0.09400533884763718, + 0.09544488042593002, + 0.129652202129364, + 0.08300462365150452, + 0.20446854829788208, + 0.08901882916688919, + 0.047402530908584595 + ], + "credit_map": [ + 0.01633095182478428, + 0.01510768011212349, + 0.01365533098578453, + 0.017985373735427856, + 0.015340355224907398, + 0.01561793778091669, + 0.015010741539299488, + 0.016082724556326866, + 0.009050814434885979, + 0.010890408419072628, + 0.006828237324953079, + 0.009206960909068584, + 0.008604386821389198, + 0.008634019643068314, + 0.009113789536058903, + 0.006976618897169828, + 0.013342797756195068, + 0.008857439272105694, + 0.007541929837316275, + 0.006244795396924019, + 0.005793015472590923, + 0.006642613559961319, + 0.006303955335170031, + 0.0078401705250144, + 0.00912302453070879, + 0.007531268056482077, + 0.01950833387672901, + 0.013527040369808674, + 0.006117840763181448, + 0.007732785772532225, + 0.0076347882859408855, + 0.022830255329608917, + 0.01623343490064144, + 0.009597325697541237, + 0.011721717193722725, + 0.015504470095038414, + 0.007563212886452675, + 0.014051951467990875, + 0.010638231411576271, + 0.010134532116353512, + 0.02010221965610981, + 0.02369708940386772, + 0.0076501318253576756, + 0.020692400634288788, + 0.013185136020183563, + 0.016945848241448402, + 0.016174638643860817, + 0.011204741895198822, + 0.007460910361260176, + 0.006209374871104956, + 0.013974280096590519, + 0.0105001674965024, + 0.008993104100227356, + 0.008693750947713852, + 0.009312214329838753, + 0.01786082237958908, + 0.027039343491196632, + 0.015146381221711636, + 0.02112208493053913, + 0.037112005054950714, + 0.02022385224699974, + 0.0161263570189476, + 0.01642918772995472, + 0.05126933753490448, + 0.005263010039925575, + 0.0038582677952945232, + 0.0202653631567955, + 0.008550034835934639, + 0.0075656333938241005, + 0.00716470368206501, + 0.005815526936203241, + 0.030536292120814323, + 0.004441118333488703, + 0.0029624965973198414, + 0.007236073259264231, + 0.005769938696175814, + 0.005199142266064882, + 0.0062714540399611, + 0.004544814117252827, + 0.010977497324347496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.008662977255880833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542534619569778 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542534619569778, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542534619569778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09765942394733429, + 0.04178904742002487, + 0.08148553222417831, + 0.19702711701393127, + 0.18998904526233673, + 0.15020504593849182, + 0.09209122508764267, + 0.07038915157318115, + 0.03944570943713188, + 0.03991871699690819 + ], + "credit_map": [ + 0.008863340131938457, + 0.012130388990044594, + 0.014235280454158783, + 0.012158684432506561, + 0.011853935196995735, + 0.013225354254245758, + 0.012337930500507355, + 0.012854501605033875, + 0.0042267050594091415, + 0.00542394770309329, + 0.0030568200163543224, + 0.004453260917216539, + 0.006499198265373707, + 0.007490408141165972, + 0.005149020813405514, + 0.005489683710038662, + 0.011074434034526348, + 0.008946288377046585, + 0.012279357761144638, + 0.0077987476252019405, + 0.008461766876280308, + 0.012452268041670322, + 0.008841251023113728, + 0.011631406843662262, + 0.01421439554542303, + 0.01195429265499115, + 0.048894576728343964, + 0.020713485777378082, + 0.015200861729681492, + 0.01782339997589588, + 0.014954084530472755, + 0.053272005170583725, + 0.024594206362962723, + 0.01839270070195198, + 0.018060116097331047, + 0.03306688740849495, + 0.011507540009915829, + 0.02649741992354393, + 0.037023600190877914, + 0.020846564322710037, + 0.022519882768392563, + 0.03745841607451439, + 0.00891716219484806, + 0.022554714232683182, + 0.014125281944870949, + 0.013468075543642044, + 0.01632327027618885, + 0.014838236384093761, + 0.00833286065608263, + 0.010067082941532135, + 0.012368172407150269, + 0.015469937585294247, + 0.012324352748692036, + 0.007940574549138546, + 0.013445505872368813, + 0.012142728082835674, + 0.004453618079423904, + 0.00700867036357522, + 0.0060332100838422775, + 0.010242024436593056, + 0.013962995260953903, + 0.009247678332030773, + 0.006264808587729931, + 0.013176143169403076, + 0.004411607049405575, + 0.0027888664044439793, + 0.0058385287411510944, + 0.0047898972406983376, + 0.0059322635643184185, + 0.0043001179583370686, + 0.004419426433742046, + 0.006964998319745064, + 0.003779100487008691, + 0.0052895573899149895, + 0.006386531516909599, + 0.005082452204078436, + 0.005211489740759134, + 0.004250246565788984, + 0.002913485514000058, + 0.007005850784480572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.012542534619569778 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640758693218231 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640758693218231, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640758693218231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039897941052913666, + 0.02230250835418701, + 0.029347173869609833, + 0.4958031475543976, + 0.0921221673488617, + 0.07161299139261246, + 0.08936623483896255, + 0.06833204627037048, + 0.06659281998872757, + 0.024623000994324684 + ], + "credit_map": [ + 0.004766217898577452, + 0.0046401480212807655, + 0.007501361891627312, + 0.005391642451286316, + 0.00451704254373908, + 0.004675733391195536, + 0.004306076560169458, + 0.004099713172763586, + 0.0028828736394643784, + 0.003013454843312502, + 0.0020263036713004112, + 0.0026644906029105186, + 0.0024189630057662725, + 0.0027612901758402586, + 0.0035024448297917843, + 0.0030326847918331623, + 0.0030798898078501225, + 0.003280408214777708, + 0.003349310252815485, + 0.0028847018256783485, + 0.003487310605123639, + 0.005829492583870888, + 0.0032672330271452665, + 0.0041688233613967896, + 0.02312820963561535, + 0.014932459220290184, + 0.17716942727565765, + 0.02623189426958561, + 0.016742244362831116, + 0.02447490766644478, + 0.02220407873392105, + 0.1909199357032776, + 0.011892001144587994, + 0.01074110809713602, + 0.007971025072038174, + 0.016133122146129608, + 0.005767879541963339, + 0.013033377937972546, + 0.017899498343467712, + 0.008684147149324417, + 0.009308983571827412, + 0.014983925968408585, + 0.005848858505487442, + 0.009347558952867985, + 0.006415260955691338, + 0.008765223436057568, + 0.0092524578794837, + 0.007690714672207832, + 0.007532869931310415, + 0.007188037969172001, + 0.011202537454664707, + 0.016094893217086792, + 0.006752696819603443, + 0.006424602121114731, + 0.01887732557952404, + 0.015293272212147713, + 0.00940659362822771, + 0.006762288510799408, + 0.006475385278463364, + 0.01027035154402256, + 0.007271185051649809, + 0.007852300070226192, + 0.006701318081468344, + 0.013592615723609924, + 0.008725348860025406, + 0.004216181114315987, + 0.012838977389037609, + 0.008403018116950989, + 0.0048475004732608795, + 0.005754011683166027, + 0.00588702317327261, + 0.01592075452208519, + 0.0027625313960015774, + 0.0030736939515918493, + 0.003538112621754408, + 0.003339158371090889, + 0.003420150140300393, + 0.0027680175844579935, + 0.0016811387613415718, + 0.004040198866277933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.008640758693218231 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028219898231327534 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028219898231327534, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028219898231327534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07416543364524841, + 0.024676945060491562, + 0.04440343379974365, + 0.2196662724018097, + 0.11888925731182098, + 0.16963797807693481, + 0.08440151810646057, + 0.11258750408887863, + 0.08675412088632584, + 0.06481754779815674 + ], + "credit_map": [ + 0.008587093092501163, + 0.005583162885159254, + 0.013141963630914688, + 0.012226300314068794, + 0.007636371534317732, + 0.008837507106363773, + 0.009010144509375095, + 0.00914289616048336, + 0.0027926452457904816, + 0.0024633274879306555, + 0.001435935264453292, + 0.002473290078341961, + 0.003702739952132106, + 0.004920785315334797, + 0.003948255907744169, + 0.0029399667400866747, + 0.005147450137883425, + 0.004424720071256161, + 0.0049040955491364, + 0.0033588584046810865, + 0.005809194408357143, + 0.010031871497631073, + 0.005783684551715851, + 0.004943555220961571, + 0.020840227603912354, + 0.016731735318899155, + 0.03760802373290062, + 0.046391263604164124, + 0.015242010354995728, + 0.022708287462592125, + 0.023506131023168564, + 0.03663858771324158, + 0.020784370601177216, + 0.0093533368781209, + 0.012972448021173477, + 0.027713125571608543, + 0.008371635340154171, + 0.01650385372340679, + 0.012040049768984318, + 0.011150435544550419, + 0.02911318838596344, + 0.049925561994314194, + 0.014916813932359219, + 0.021748565137386322, + 0.012797084636986256, + 0.01489283237606287, + 0.014102905057370663, + 0.012141036801040173, + 0.006369826849550009, + 0.008196782320737839, + 0.008396296761929989, + 0.016321539878845215, + 0.012083957903087139, + 0.007479249034076929, + 0.010425113141536713, + 0.01512875221669674, + 0.01241299882531166, + 0.012774533592164516, + 0.010379648767411709, + 0.0132293701171875, + 0.01554152462631464, + 0.013443948701024055, + 0.009608815424144268, + 0.02519666589796543, + 0.008724275045096874, + 0.006622113287448883, + 0.014646000228822231, + 0.010970791801810265, + 0.011846603825688362, + 0.008617467246949673, + 0.007433258928358555, + 0.017893612384796143, + 0.010693546384572983, + 0.007866374216973782, + 0.008087554946541786, + 0.007075832225382328, + 0.009904779493808746, + 0.007620010059326887, + 0.00486812274903059, + 0.008701328188180923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.0028219898231327534 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly high-angled medium shot frames an ornate, dark mahogany desk, its polished surface gleaming in the focused, warm glow of a single green bank lamp in a dimly lit, wood-paneled study. The stillness of the scene is suddenly shattered as the desk begins to tremble with an almost imperceptible, low vibration, causing the fountain pen in its brass holder to rattle softly. This initial tremor escalates rapidly and violently into a convulsive earthquake, the entire heavy wooden frame rocking back and forth with unnatural force. Loose, previously neatly arranged sheets of handwritten parchment and modern typed documents are thrown upward and begin to flutter chaotically, scattering across the desk's surface and drifting out of frame to the floor. The unbroken shot tensely captures the escalating chaos, capturing the moment when the antique glass inkwell tilts and spills a pool of deep black ink, which spreads across the polished wood like a creeping stain, creating a stark, dramatic contrast that amplifies the scene's intensifying, baffling horror.", + "seed": 6542, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_6/index_6311_seed_6542_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": -0.016255848482251166, + "vlm_logprob": -0.016255848482251166, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.016255848482251166, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012883991934359074 + ], + "opposite_logprob_tokens": [ + -4.887884140014648 + ], + "logprob_reward": -0.012883991934359074, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012883991934359074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2238902598619461, + 0.04177601635456085, + 0.2865902781486511, + 0.02393006905913353, + 0.032854560762643814, + 0.08046729117631912, + 0.09908978641033173, + 0.060040008276700974, + 0.10040576756000519, + 0.05095599219202995 + ], + "credit_map": [ + 0.030406907200813293, + 0.018962901085615158, + 0.024510741233825684, + 0.020742906257510185, + 0.016042308881878853, + 0.04148668795824051, + 0.028485678136348724, + 0.043252140283584595, + 0.006471005734056234, + 0.0033675478771328926, + 0.00528609799221158, + 0.004130018409341574, + 0.002354574156925082, + 0.005611908156424761, + 0.006988125387579203, + 0.007566740270704031, + 0.04051750898361206, + 0.015023710206151009, + 0.021431978791952133, + 0.02959286794066429, + 0.019513826817274094, + 0.01442757062613964, + 0.06485185772180557, + 0.08123096078634262, + 0.0025669464375823736, + 0.002882150234654546, + 0.0034620678052306175, + 0.0031210132874548435, + 0.003273384878411889, + 0.0026256064884364605, + 0.0037481668405234814, + 0.002250732621178031, + 0.002142382552847266, + 0.002139113610610366, + 0.00653546629473567, + 0.0054837544448673725, + 0.0020289444364607334, + 0.0053683919832110405, + 0.0039027882739901543, + 0.0052537210285663605, + 0.009263659827411175, + 0.008347712457180023, + 0.008922787383198738, + 0.01234037522226572, + 0.013172810897231102, + 0.012916797772049904, + 0.00826400425285101, + 0.007239141967147589, + 0.011468288488686085, + 0.013934594579041004, + 0.008236231282353401, + 0.00793435052037239, + 0.021363193169236183, + 0.016762053593993187, + 0.009496642276644707, + 0.009894434362649918, + 0.008683916181325912, + 0.007434960454702377, + 0.006727166008204222, + 0.005353366024792194, + 0.006110580638051033, + 0.010114872828125954, + 0.007464212365448475, + 0.008150934241712093, + 0.008294657804071903, + 0.011610862798988819, + 0.007763685658574104, + 0.013318692333996296, + 0.013731597922742367, + 0.01811080053448677, + 0.015717362985014915, + 0.01185811311006546, + 0.003052409505471587, + 0.006415616720914841, + 0.0062535107135772705, + 0.004678433761000633, + 0.005714030470699072, + 0.010791974142193794, + 0.007724136579781771, + 0.0063258809968829155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.012883991934359074 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233310479670763 + ], + "opposite_logprob_tokens": [ + -3.7733311653137207 + ], + "logprob_reward": -0.0233310479670763, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233310479670763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19053836166858673, + 0.04439666494727135, + 0.2855718433856964, + 0.05739035829901695, + 0.08064348995685577, + 0.06253824383020401, + 0.10731938481330872, + 0.06705065071582794, + 0.04166221618652344, + 0.0628887340426445 + ], + "credit_map": [ + 0.028698522597551346, + 0.015794990584254265, + 0.02001997083425522, + 0.02232365868985653, + 0.016021817922592163, + 0.028717266395688057, + 0.022096823900938034, + 0.036865316331386566, + 0.005962473340332508, + 0.003326371544972062, + 0.00447802385315299, + 0.004010837525129318, + 0.0035984315909445286, + 0.008080205880105495, + 0.008299790322780609, + 0.006640539038926363, + 0.03867363929748535, + 0.020320778712630272, + 0.032380975782871246, + 0.04906109720468521, + 0.02609683945775032, + 0.023286592215299606, + 0.05840691551566124, + 0.03734506294131279, + 0.005192760843783617, + 0.007392204832285643, + 0.010242661461234093, + 0.008230302482843399, + 0.006734644528478384, + 0.006357220467180014, + 0.006724013481289148, + 0.006516559980809689, + 0.006069930270314217, + 0.008492249064147472, + 0.01143118180334568, + 0.008300895802676678, + 0.015768978744745255, + 0.010692107491195202, + 0.009250286035239697, + 0.010637872852385044, + 0.003703007474541664, + 0.006729422602802515, + 0.007551985327154398, + 0.010511172004044056, + 0.010803873650729656, + 0.00998311210423708, + 0.005994903389364481, + 0.007260775193572044, + 0.009794001467525959, + 0.01573227345943451, + 0.010477946139872074, + 0.012483594007790089, + 0.021580761298537254, + 0.014972562901675701, + 0.012606944888830185, + 0.009671309031546116, + 0.007158477324992418, + 0.005926568992435932, + 0.004744496662169695, + 0.008582071401178837, + 0.007397200912237167, + 0.00809183344244957, + 0.015858717262744904, + 0.009291286580264568, + 0.002612633164972067, + 0.004252428654581308, + 0.003304192330688238, + 0.0033745630644261837, + 0.004823412746191025, + 0.010244022123515606, + 0.005202220752835274, + 0.007848748005926609, + 0.0027714483439922333, + 0.004621490836143494, + 0.004039942752569914, + 0.004195619374513626, + 0.007011872716248035, + 0.016292650252580643, + 0.005322093144059181, + 0.018633626401424408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0233310479670763 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.13349692523479462, + 0.06138797849416733, + 0.09789600968360901, + 0.09897851198911667, + 0.05520656332373619, + 0.07823310047388077, + 0.11596936732530594, + 0.08391138166189194, + 0.13192768394947052, + 0.14299243688583374 + ], + "credit_map": [ + 0.012078087776899338, + 0.011412830092012882, + 0.010679027996957302, + 0.011819828301668167, + 0.010315639898180962, + 0.025500645861029625, + 0.017938990145921707, + 0.033751875162124634, + 0.008223440498113632, + 0.004915887489914894, + 0.00665898434817791, + 0.006762134376913309, + 0.004994911141693592, + 0.009980881586670876, + 0.010585171170532703, + 0.009266570210456848, + 0.011960739269852638, + 0.007292598485946655, + 0.008065934292972088, + 0.008991657756268978, + 0.008775887079536915, + 0.0060292696580290794, + 0.02067284844815731, + 0.026107072830200195, + 0.014904621988534927, + 0.0075555951334536076, + 0.017929114401340485, + 0.014434801414608955, + 0.014785226434469223, + 0.009472294710576534, + 0.008799456059932709, + 0.011097400449216366, + 0.004833534359931946, + 0.006481618620455265, + 0.007701622787863016, + 0.0048942482098937035, + 0.011429604142904282, + 0.0077657331712543964, + 0.005901545751839876, + 0.0061986553482711315, + 0.005996546242386103, + 0.009999602101743221, + 0.00980470608919859, + 0.013429404236376286, + 0.00968848541378975, + 0.012607639655470848, + 0.008159901015460491, + 0.008546818979084492, + 0.010571617633104324, + 0.013688781298696995, + 0.013524971902370453, + 0.009988963603973389, + 0.025744033977389336, + 0.017043817788362503, + 0.015063349157571793, + 0.01034383475780487, + 0.01065649464726448, + 0.012442662380635738, + 0.00916940439492464, + 0.007317740470170975, + 0.010982241481542587, + 0.011966566555202007, + 0.010798320174217224, + 0.010577947832643986, + 0.010444567538797855, + 0.012602453120052814, + 0.010183011181652546, + 0.020872756838798523, + 0.01831710711121559, + 0.02200540341436863, + 0.019488440826535225, + 0.01801394857466221, + 0.007443741429597139, + 0.017869021743535995, + 0.014335337094962597, + 0.013880993239581585, + 0.018938113003969193, + 0.028864413499832153, + 0.018466532230377197, + 0.023194288834929466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.029801098629832268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122843474149704, + 0.048556141555309296, + 0.14083248376846313, + 0.09983839094638824, + 0.05737698823213577, + 0.0653371810913086, + 0.08422181755304337, + 0.12266183644533157, + 0.1434200406074524, + 0.12547074258327484 + ], + "credit_map": [ + 0.009197032079100609, + 0.009917248971760273, + 0.010965771973133087, + 0.009843235835433006, + 0.010185657069087029, + 0.023123331367969513, + 0.013767538592219353, + 0.025284530594944954, + 0.006174308247864246, + 0.003294809954240918, + 0.00582900270819664, + 0.006471824366599321, + 0.002829024102538824, + 0.007172773592174053, + 0.00749941123649478, + 0.009284988977015018, + 0.029130443930625916, + 0.007465872913599014, + 0.012035707011818886, + 0.02593209780752659, + 0.012021424248814583, + 0.009116781875491142, + 0.01633729785680771, + 0.028792856261134148, + 0.012827673926949501, + 0.006834811065346003, + 0.015587796457111835, + 0.012489824555814266, + 0.010047470219433308, + 0.01197078824043274, + 0.011853177100419998, + 0.018226854503154755, + 0.005238506477326155, + 0.006445170845836401, + 0.007180423010140657, + 0.006006432231515646, + 0.011201580055058002, + 0.007446494419127703, + 0.007574894931167364, + 0.006283487658947706, + 0.005658511538058519, + 0.005848695058375597, + 0.008023540489375591, + 0.014518453739583492, + 0.006593142636120319, + 0.009345277212560177, + 0.007958774454891682, + 0.0073907822370529175, + 0.009446927346289158, + 0.010514532215893269, + 0.007905111648142338, + 0.009867378510534763, + 0.016999153420329094, + 0.014014995656907558, + 0.008220667019486427, + 0.007253055926412344, + 0.013258067891001701, + 0.017143819481134415, + 0.010755090974271297, + 0.016183579340577126, + 0.01495163980871439, + 0.013544153422117233, + 0.020713454112410545, + 0.016112029552459717, + 0.010698061436414719, + 0.016748445108532906, + 0.015146705321967602, + 0.01557887252420187, + 0.019170647487044334, + 0.02451530285179615, + 0.017124557867646217, + 0.02443745732307434, + 0.006070671137422323, + 0.014014833606779575, + 0.012743365950882435, + 0.012363884598016739, + 0.015541477128863335, + 0.025123117491602898, + 0.014906530268490314, + 0.024706872180104256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.007632025051862001 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 dimly lit, wood-paneled study featuring a dark mahogany desk with a polished surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a single green bank lamp providing warm, focused light on the desk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the desk begin to tremble with a low vibration that escalates into a violent, convulsive earthquake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do handwritten parchment and typed documents scatter chaotically across the desk and drift out of frame to the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does an antique glass inkwell tilt and spill a pool of deep black ink that spreads across the polished wood?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0336783230304718, + 0.04176519438624382, + 0.07802516222000122, + 0.040585070848464966, + 0.0777793675661087, + 0.09815693646669388, + 0.15342815220355988, + 0.1303577721118927, + 0.13072530925273895, + 0.21549874544143677 + ], + "credit_map": [ + 0.006926281377673149, + 0.0027762267272919416, + 0.0026898307260125875, + 0.004353977739810944, + 0.0027058047708123922, + 0.004579741507768631, + 0.00488684605807066, + 0.004759614355862141, + 0.004041648004204035, + 0.002089502988383174, + 0.003259401535615325, + 0.0030639306642115116, + 0.0019657728262245655, + 0.009404486045241356, + 0.009976792149245739, + 0.00796365737915039, + 0.011754565872251987, + 0.004796708468347788, + 0.006476872600615025, + 0.0071257478557527065, + 0.0068682399578392506, + 0.0049789538607001305, + 0.018398180603981018, + 0.017625894397497177, + 0.004825209733098745, + 0.003203693078830838, + 0.004760535899549723, + 0.005564416758716106, + 0.005098256282508373, + 0.0052316379733383656, + 0.00414041243493557, + 0.007760909851640463, + 0.006566130090504885, + 0.008366905152797699, + 0.00849673431366682, + 0.008195661008358002, + 0.016291670501232147, + 0.010259370319545269, + 0.010311238467693329, + 0.009291657246649265, + 0.009865319356322289, + 0.009881522506475449, + 0.015127657912671566, + 0.02004540152847767, + 0.010900028981268406, + 0.013258487917482853, + 0.010811964049935341, + 0.008266553282737732, + 0.01810387335717678, + 0.019192487001419067, + 0.016007589176297188, + 0.022315479815006256, + 0.026244381442666054, + 0.016488349065184593, + 0.018584338948130608, + 0.016491655260324478, + 0.012878435663878918, + 0.012988738715648651, + 0.013661038130521774, + 0.01221306063234806, + 0.01820913329720497, + 0.030272163450717926, + 0.014681159518659115, + 0.015454036183655262, + 0.008516733534634113, + 0.01313608605414629, + 0.010994256474077702, + 0.009555981494486332, + 0.017129553481936455, + 0.03353920951485634, + 0.015272569842636585, + 0.022580912336707115, + 0.009954679757356644, + 0.022466961294412613, + 0.020737264305353165, + 0.01857195608317852, + 0.02284053899347782, + 0.04635089635848999, + 0.031105872243642807, + 0.04347056522965431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.007631078828126192 + } + ], + "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/0000158/group_7/index_tensor([9167, 9167])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_tensor([9167, 9167])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..077e55a922aae082ba80fb0a3e639e2a56c8f53c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_tensor([9167, 9167])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot captures a futuristic chain-driven snowmobile slamming on the brakes, its heavy, exposed chain drive and cleated tracks violently churning deep into the pristine snow. Designed with a sleek, aerodynamic body composed of matte gunmetal gray composite panels and sharp, angular lines, the vehicle unleashes a vast, dazzling plume of crystalline powder that instantly engulfs it, against a breathtaking backdrop of awe-inspiring, jagged ice peaks piercing the pale, frosty sky. As the machine hums to life and falls into a profound silence, the white cloud of snow slowly dissipates, revealing its silent form. Then, in a smooth and deliberate doll-like retreat, the camera slowly pulls away from the lone vehicle. The frame gradually widens, revealing the immense scale of the surrounding landscape: a vast, silent, and completely empty expanse of untouched snow stretching to the distant mountains. The long, blue shadows cast by the low-hanging sun underscore the profound, quiet solitude of the vast winter wilderness.", + "seed": 20353, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20353_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": -0.031378220720216635, + "vlm_logprob": -0.031378220720216635, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.031378220720216635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022517476230859756 + ], + "opposite_logprob_tokens": [ + -3.897517442703247 + ], + "logprob_reward": -0.022517476230859756, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022517476230859756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16403795778751373, + 0.15171436965465546, + 0.08120892196893692, + 0.09943006932735443, + 0.06760595738887787, + 0.2584310472011566, + 0.05362200364470482, + 0.046401817351579666, + 0.033231671899557114, + 0.0443161204457283 + ], + "credit_map": [ + 0.022155893966555595, + 0.022832032293081284, + 0.018002374097704887, + 0.013000705279409885, + 0.024897566065192223, + 0.023910708725452423, + 0.015425626188516617, + 0.02381305955350399, + 0.020248234272003174, + 0.015124361030757427, + 0.022495027631521225, + 0.01588350161910057, + 0.031100470572710037, + 0.018395287916064262, + 0.014131990261375904, + 0.014335502870380878, + 0.014379014261066914, + 0.011499564163386822, + 0.009649447165429592, + 0.0086923036724329, + 0.009348521940410137, + 0.010812005028128624, + 0.009315011091530323, + 0.007513063494116068, + 0.015370069071650505, + 0.014929920434951782, + 0.015559425577521324, + 0.010784882120788097, + 0.013241235166788101, + 0.00983586348593235, + 0.010674409568309784, + 0.009034276008605957, + 0.01192332711070776, + 0.014767887070775032, + 0.011353655718266964, + 0.005723622627556324, + 0.005991995334625244, + 0.005511595867574215, + 0.0060749780386686325, + 0.006258898880332708, + 0.03126462921500206, + 0.0295228473842144, + 0.03498329222202301, + 0.03110845573246479, + 0.026835549622774124, + 0.0451185405254364, + 0.03351995721459389, + 0.02607780136168003, + 0.01030905731022358, + 0.008742444217205048, + 0.005690484307706356, + 0.005552446004003286, + 0.00522185955196619, + 0.007199331186711788, + 0.005706943105906248, + 0.005199445877224207, + 0.0075538442470133305, + 0.004770039580762386, + 0.005802607163786888, + 0.0051580993458628654, + 0.0052514574490487576, + 0.0066312151029706, + 0.004306531976908445, + 0.006928029470145702, + 0.004301007837057114, + 0.006370168644934893, + 0.0018361101392656565, + 0.003582064062356949, + 0.006708870641887188, + 0.00431649899110198, + 0.002959911711513996, + 0.003157042432576418, + 0.00577579578384757, + 0.008555525913834572, + 0.005364520009607077, + 0.005361178424209356, + 0.008064765483140945, + 0.003022567369043827, + 0.003300135489553213, + 0.004871638026088476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.022517476230859756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292766034603119 + ], + "opposite_logprob_tokens": [ + -2.2379276752471924 + ], + "logprob_reward": -0.11292766034603119, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292766034603119, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21221154928207397, + 0.10244131088256836, + 0.1030084639787674, + 0.05765817314386368, + 0.0855095237493515, + 0.10760800540447235, + 0.09062549471855164, + 0.11831913888454437, + 0.10052965581417084, + 0.02208872139453888 + ], + "credit_map": [ + 0.02455073408782482, + 0.022790031507611275, + 0.026220593601465225, + 0.02624359354376793, + 0.02716987207531929, + 0.030944228172302246, + 0.029953299090266228, + 0.02433917485177517, + 0.016240248456597328, + 0.00967737752944231, + 0.013966193422675133, + 0.009843749925494194, + 0.019053934141993523, + 0.013566305860877037, + 0.01022468600422144, + 0.0098688043653965, + 0.017726171761751175, + 0.012098368257284164, + 0.013676861301064491, + 0.008265676908195019, + 0.016452347859740257, + 0.012389075011014938, + 0.013852380216121674, + 0.008547566831111908, + 0.008881391026079655, + 0.009136701934039593, + 0.0086190365254879, + 0.005912865046411753, + 0.008115449920296669, + 0.0064479452557861805, + 0.00582039775326848, + 0.004724378231912851, + 0.014737569727003574, + 0.01795540377497673, + 0.014636586420238018, + 0.008314182981848717, + 0.008480696007609367, + 0.006942214909940958, + 0.0075485906563699245, + 0.00689427787438035, + 0.015277093276381493, + 0.013445790857076645, + 0.015240192413330078, + 0.013545825146138668, + 0.012954508885741234, + 0.010298972018063068, + 0.016002580523490906, + 0.01084302831441164, + 0.012791968882083893, + 0.011441160924732685, + 0.014697263017296791, + 0.01743289828300476, + 0.0065662856213748455, + 0.00701325386762619, + 0.00955845508724451, + 0.011124198324978352, + 0.026936160400509834, + 0.016260117292404175, + 0.007144571747630835, + 0.013035308569669724, + 0.01689089462161064, + 0.012127241119742393, + 0.013915632851421833, + 0.01200919970870018, + 0.021734576672315598, + 0.013352183625102043, + 0.007704108487814665, + 0.010491864755749702, + 0.01187633816152811, + 0.011810535565018654, + 0.011809080839157104, + 0.011750965379178524, + 0.0035280997399240732, + 0.004412657115608454, + 0.0021768249571323395, + 0.002546684816479683, + 0.003606697777286172, + 0.002064659260213375, + 0.0019419473828747869, + 0.0018111476674675941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.11292766034603119 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15340425074100494, + 0.24732472002506256, + 0.06325490772724152, + 0.062156904488801956, + 0.1204088106751442, + 0.15339119732379913, + 0.0636153370141983, + 0.05213777348399162, + 0.06352625042200089, + 0.020779866725206375 + ], + "credit_map": [ + 0.02034355141222477, + 0.014207490719854832, + 0.020949626341462135, + 0.017089055851101875, + 0.014226430095732212, + 0.024279184639453888, + 0.016746193170547485, + 0.025562714785337448, + 0.044114962220191956, + 0.040870826691389084, + 0.036083027720451355, + 0.019265295937657356, + 0.04238603636622429, + 0.03304501622915268, + 0.013915786519646645, + 0.01764376275241375, + 0.009807426482439041, + 0.00548779871314764, + 0.009029914624989033, + 0.007573574315756559, + 0.007807145360857248, + 0.005873069167137146, + 0.01008040551096201, + 0.007595572154968977, + 0.008804279379546642, + 0.008588302880525589, + 0.00886252149939537, + 0.00594840943813324, + 0.006769412662833929, + 0.0059576136991381645, + 0.006720355246216059, + 0.010506011545658112, + 0.020569995045661926, + 0.028186874464154243, + 0.020627066493034363, + 0.013083686120808125, + 0.01099498849362135, + 0.008128421381115913, + 0.011109601706266403, + 0.007708172779530287, + 0.01548488438129425, + 0.013597387820482254, + 0.02404961735010147, + 0.021760260686278343, + 0.015509691089391708, + 0.024061860516667366, + 0.022738609462976456, + 0.01618887484073639, + 0.010336135514080524, + 0.008681071922183037, + 0.008125972002744675, + 0.009547596797347069, + 0.004973811097443104, + 0.0058183507062494755, + 0.009568396955728531, + 0.006564000621438026, + 0.008329910226166248, + 0.006537904962897301, + 0.005658377427607775, + 0.0074117532931268215, + 0.006801638286560774, + 0.0050415596924722195, + 0.007523221429437399, + 0.004833405837416649, + 0.014645928516983986, + 0.00890995841473341, + 0.005829919595271349, + 0.0076111117377877235, + 0.008802835829555988, + 0.0061139799654483795, + 0.0060964287258684635, + 0.005516094155609608, + 0.003492305288091302, + 0.004445657134056091, + 0.0017440383089706302, + 0.00280487397685647, + 0.003752463962882757, + 0.0017818502383306623, + 0.0015488103963434696, + 0.001209866488352418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021934278775006533 + ], + "opposite_logprob_tokens": [ + -6.127193450927734 + ], + "logprob_reward": -0.0021934278775006533, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021934278775006533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14106905460357666, + 0.2320566028356552, + 0.07405905425548553, + 0.03632494807243347, + 0.059479016810655594, + 0.11215473711490631, + 0.08563973754644394, + 0.09232565760612488, + 0.1110692247748375, + 0.05582194775342941 + ], + "credit_map": [ + 0.02153605781495571, + 0.015068769454956055, + 0.0167742520570755, + 0.01656343974173069, + 0.01447097398340702, + 0.022150667384266853, + 0.017499538138508797, + 0.017005356028676033, + 0.0367053784430027, + 0.03506595641374588, + 0.03592440113425255, + 0.016749601811170578, + 0.03895646706223488, + 0.033491816371679306, + 0.015321028418838978, + 0.01984195038676262, + 0.0194898322224617, + 0.008930585347115993, + 0.013095333240926266, + 0.007371037267148495, + 0.006549037992954254, + 0.008111665025353432, + 0.004218160640448332, + 0.006293403450399637, + 0.005730521399527788, + 0.0053673675283789635, + 0.0050890082493424416, + 0.0032901098020374775, + 0.004450169391930103, + 0.00321975233964622, + 0.005891758482903242, + 0.0032862606458365917, + 0.012582470662891865, + 0.011522860266268253, + 0.010035473853349686, + 0.005225159227848053, + 0.005224634427577257, + 0.004224547650665045, + 0.005348604172468185, + 0.0053152707405388355, + 0.019894592463970184, + 0.017492592334747314, + 0.017443660646677017, + 0.01357630081474781, + 0.009126233868300915, + 0.012112674303352833, + 0.01137674693018198, + 0.011131936684250832, + 0.02296191267669201, + 0.013430227525532246, + 0.01031099259853363, + 0.00781444925814867, + 0.00498224189504981, + 0.010285108350217342, + 0.005821307189762592, + 0.010033500380814075, + 0.019014105200767517, + 0.011273208074271679, + 0.008593355305492878, + 0.01271190494298935, + 0.01163287740200758, + 0.008639921434223652, + 0.010330898687243462, + 0.01012938842177391, + 0.032424118369817734, + 0.015456502325832844, + 0.012945708818733692, + 0.010188574902713299, + 0.014715561643242836, + 0.009491986595094204, + 0.008881854824721813, + 0.006964921951293945, + 0.010836920700967312, + 0.011357608251273632, + 0.0038365349173545837, + 0.0048399982042610645, + 0.012209214270114899, + 0.005304395221173763, + 0.004049011040478945, + 0.0033882667776197195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.0021934278775006533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003187221009284258 + ], + "opposite_logprob_tokens": [ + -5.75318717956543 + ], + "logprob_reward": -0.003187221009284258, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003187221009284258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07589337974786758, + 0.061198506504297256, + 0.059595197439193726, + 0.05110037326812744, + 0.14735275506973267, + 0.09029177576303482, + 0.1388428509235382, + 0.13581489026546478, + 0.13958074152469635, + 0.10032955557107925 + ], + "credit_map": [ + 0.010597490705549717, + 0.009952150285243988, + 0.007661818992346525, + 0.0092008700594306, + 0.005785494111478329, + 0.011983825825154781, + 0.010449448600411415, + 0.010262280702590942, + 0.00923959817737341, + 0.008027300238609314, + 0.008756515569984913, + 0.006078633014112711, + 0.00962262973189354, + 0.00916154496371746, + 0.005323306191712618, + 0.0049889711663126945, + 0.01120334304869175, + 0.007321030832827091, + 0.008073370903730392, + 0.006210511550307274, + 0.0066389781422913074, + 0.007043306715786457, + 0.006384697742760181, + 0.006719956640154123, + 0.007314436603337526, + 0.005754563491791487, + 0.008061438798904419, + 0.006936999503523111, + 0.005739391315728426, + 0.005431099329143763, + 0.0058512454852461815, + 0.006011195480823517, + 0.024741007015109062, + 0.02274518646299839, + 0.021550722420215607, + 0.017144503071904182, + 0.015266535803675652, + 0.013691367581486702, + 0.01869013160467148, + 0.0135232824832201, + 0.010032804682850838, + 0.011234006844460964, + 0.012459236197173595, + 0.011127549223601818, + 0.009955311194062233, + 0.01301830168813467, + 0.013753763400018215, + 0.008710798807442188, + 0.020443584769964218, + 0.017139192670583725, + 0.016836578026413918, + 0.022143570706248283, + 0.010314182378351688, + 0.01301693357527256, + 0.020255139097571373, + 0.018693672493100166, + 0.020281312987208366, + 0.015510744415223598, + 0.012281622737646103, + 0.022672101855278015, + 0.012258435599505901, + 0.021734317764639854, + 0.014719093218445778, + 0.016357261687517166, + 0.04287001118063927, + 0.022990161553025246, + 0.011239719577133656, + 0.011310317553579807, + 0.02140725962817669, + 0.012082831934094429, + 0.008935526944696903, + 0.008744914084672928, + 0.012177315540611744, + 0.017233233898878098, + 0.007574041839689016, + 0.00801711343228817, + 0.018525851890444756, + 0.009308098815381527, + 0.012165604159235954, + 0.015328281559050083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.003187221009284258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot captures a futuristic chain-driven snowmobile slamming on the brakes, its heavy, exposed chain drive and cleated tracks violently churning deep into the pristine snow. Designed with a sleek, aerodynamic body composed of matte gunmetal gray composite panels and sharp, angular lines, the vehicle unleashes a vast, dazzling plume of crystalline powder that instantly engulfs it, against a breathtaking backdrop of awe-inspiring, jagged ice peaks piercing the pale, frosty sky. As the machine hums to life and falls into a profound silence, the white cloud of snow slowly dissipates, revealing its silent form. Then, in a smooth and deliberate doll-like retreat, the camera slowly pulls away from the lone vehicle. The frame gradually widens, revealing the immense scale of the surrounding landscape: a vast, silent, and completely empty expanse of untouched snow stretching to the distant mountains. The long, blue shadows cast by the low-hanging sun underscore the profound, quiet solitude of the vast winter wilderness.", + "seed": 20354, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20354_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": -0.03368455190211535, + "vlm_logprob": -0.03368455190211535, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.03368455190211535, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.045854970812797546 + ], + "opposite_logprob_tokens": [ + -3.1708550453186035 + ], + "logprob_reward": -0.045854970812797546, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.045854970812797546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07284418493509293, + 0.030082877725362778, + 0.17028971016407013, + 0.02644907310605049, + 0.051915932446718216, + 0.057856280356645584, + 0.49772506952285767, + 0.05108799785375595, + 0.028141407296061516, + 0.013607487082481384 + ], + "credit_map": [ + 0.016992228105664253, + 0.00672693969681859, + 0.017134318128228188, + 0.009221365675330162, + 0.003921443596482277, + 0.004920843057334423, + 0.006558711640536785, + 0.0073683420196175575, + 0.004711040761321783, + 0.0033123346511274576, + 0.00503678061068058, + 0.0035550228785723448, + 0.002792872954159975, + 0.0031418511644005775, + 0.0044130850583314896, + 0.0031198901124298573, + 0.022570554167032242, + 0.02489805966615677, + 0.016277823597192764, + 0.022498557344079018, + 0.019503969699144363, + 0.02150871977210045, + 0.02316967211663723, + 0.019862348213791847, + 0.004252323880791664, + 0.005362777039408684, + 0.003036476904526353, + 0.003474240191280842, + 0.0032485807314515114, + 0.0033805205021053553, + 0.001993426587432623, + 0.0017007263377308846, + 0.008628742769360542, + 0.006112601142376661, + 0.00299027725122869, + 0.0035030285362154245, + 0.008774987421929836, + 0.012130496092140675, + 0.003523262683302164, + 0.0062525393441319466, + 0.010395271703600883, + 0.00539717823266983, + 0.0053022513166069984, + 0.0037085111252963543, + 0.011466935276985168, + 0.007443437352776527, + 0.00662598479539156, + 0.0075167091563344, + 0.03632286190986633, + 0.20563584566116333, + 0.14269620180130005, + 0.01664714515209198, + 0.0164398942142725, + 0.058043573051691055, + 0.009356354363262653, + 0.012583204545080662, + 0.01208513043820858, + 0.006144002079963684, + 0.0077897896990180016, + 0.004573923535645008, + 0.004967503249645233, + 0.006689996458590031, + 0.004490376450121403, + 0.0043472712859511375, + 0.004325381480157375, + 0.00411817105486989, + 0.0021974309347569942, + 0.002597079612314701, + 0.0038720646407455206, + 0.003951309714466333, + 0.002496834145858884, + 0.004583137575536966, + 0.00157150998711586, + 0.0018472931114956737, + 0.0013929944252595305, + 0.001543222344480455, + 0.0018136611906811595, + 0.00231731659732759, + 0.0014548486797139049, + 0.0016666401643306017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.045854970812797546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003124788403511 + ], + "opposite_logprob_tokens": [ + -2.3503124713897705 + ], + "logprob_reward": -0.1003124788403511, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003124788403511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3521542251110077, + 0.031936757266521454, + 0.0545806922018528, + 0.04121916741132736, + 0.14056739211082458, + 0.06745073944330215, + 0.12789416313171387, + 0.09676321595907211, + 0.055533237755298615, + 0.03190041705965996 + ], + "credit_map": [ + 0.06630895286798477, + 0.027025070041418076, + 0.029035532847046852, + 0.07585246860980988, + 0.024680878967046738, + 0.03566811978816986, + 0.0454409085214138, + 0.04814225807785988, + 0.005200696177780628, + 0.0033996051643043756, + 0.00707713607698679, + 0.0040373122319579124, + 0.004224697593599558, + 0.0027690441347658634, + 0.002905048429965973, + 0.0023232181556522846, + 0.00774773582816124, + 0.008331921882927418, + 0.005575857125222683, + 0.005802799016237259, + 0.006391587201505899, + 0.007386669982224703, + 0.006057108752429485, + 0.007287010550498962, + 0.004959624260663986, + 0.009931834414601326, + 0.005413200240582228, + 0.004011199809610844, + 0.005332170054316521, + 0.004450669512152672, + 0.0035984963178634644, + 0.003521972568705678, + 0.021130353212356567, + 0.013211815617978573, + 0.00882866233587265, + 0.010017751716077328, + 0.02370777539908886, + 0.034750472754240036, + 0.010123816318809986, + 0.018796749413013458, + 0.011830577626824379, + 0.011674267239868641, + 0.010935338214039803, + 0.005335433408617973, + 0.006904589477926493, + 0.008078823797404766, + 0.00606576120480895, + 0.006625944748520851, + 0.01585964486002922, + 0.03329446539282799, + 0.03049083612859249, + 0.011746145784854889, + 0.008549264632165432, + 0.012322522699832916, + 0.006530204322189093, + 0.009101083502173424, + 0.01931874081492424, + 0.009995689615607262, + 0.010782359167933464, + 0.010584711097180843, + 0.013997416011989117, + 0.013274402357637882, + 0.008397410623729229, + 0.010412488132715225, + 0.010729951784014702, + 0.007049491163343191, + 0.004734329879283905, + 0.0049671027809381485, + 0.009452075697481632, + 0.006448920350521803, + 0.006731655914336443, + 0.0054197125136852264, + 0.004066759720444679, + 0.005009282846003771, + 0.0041314163245260715, + 0.0031849804800003767, + 0.005312558729201555, + 0.004639171529561281, + 0.0019924920052289963, + 0.003563753329217434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.1003124788403511 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16401256620883942, + 0.033096443861722946, + 0.05543261021375656, + 0.04071677103638649, + 0.03211481124162674, + 0.08637670427560806, + 0.2999263107776642, + 0.16621296107769012, + 0.07049281150102615, + 0.051618024706840515 + ], + "credit_map": [ + 0.03726951405405998, + 0.01789456605911255, + 0.013475796207785606, + 0.047422271221876144, + 0.01002824679017067, + 0.01212201826274395, + 0.015279211103916168, + 0.010520934127271175, + 0.006371070630848408, + 0.0030449479818344116, + 0.01166135910898447, + 0.0035634469240903854, + 0.0023176956456154585, + 0.0018402757123112679, + 0.00250420393422246, + 0.0017934476491063833, + 0.007675043772906065, + 0.008695836178958416, + 0.0056258877739310265, + 0.0065077608451247215, + 0.006443917751312256, + 0.007010678295046091, + 0.00689161941409111, + 0.006581868510693312, + 0.00605414342135191, + 0.0076724267564713955, + 0.004960829392075539, + 0.0048661669716238976, + 0.005235191434621811, + 0.0060756574384868145, + 0.0032238864805549383, + 0.002628467744216323, + 0.0046148826368153095, + 0.005287621170282364, + 0.002535238629207015, + 0.002884522546082735, + 0.006038369610905647, + 0.0034556149039417505, + 0.0028921528719365597, + 0.004406408406794071, + 0.019054867327213287, + 0.009201274253427982, + 0.010721362195909023, + 0.005547418259084225, + 0.009059568867087364, + 0.019836636260151863, + 0.006611982826143503, + 0.0063435896299779415, + 0.035872578620910645, + 0.11543154716491699, + 0.05911838635802269, + 0.01580585539340973, + 0.01690618507564068, + 0.035497117787599564, + 0.010080654174089432, + 0.011213993653655052, + 0.04103752225637436, + 0.020759649574756622, + 0.02400743030011654, + 0.016423551365733147, + 0.012518866918981075, + 0.02288724295794964, + 0.011582530103623867, + 0.01699616014957428, + 0.010913258418440819, + 0.008904154412448406, + 0.005524535197764635, + 0.007273836061358452, + 0.011002280749380589, + 0.011624232865869999, + 0.007334958761930466, + 0.007915557362139225, + 0.006344178691506386, + 0.0066613201051950455, + 0.005891277454793453, + 0.006240697577595711, + 0.008354363031685352, + 0.005896870978176594, + 0.004625383298844099, + 0.007603929843753576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.014193889684975147 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815570682287216 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815570682287216, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815570682287216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06514715403318405, + 0.04327172413468361, + 0.1063336580991745, + 0.028526779264211655, + 0.04646456241607666, + 0.06091366335749626, + 0.3170633316040039, + 0.11229004710912704, + 0.14589078724384308, + 0.07409828156232834 + ], + "credit_map": [ + 0.012015494517982006, + 0.004930136725306511, + 0.009137087501585484, + 0.017379114404320717, + 0.004119119606912136, + 0.007485877722501755, + 0.005607462488114834, + 0.004472862463444471, + 0.006784813012927771, + 0.005175357218831778, + 0.007469356060028076, + 0.0047922078520059586, + 0.004426843486726284, + 0.0037527058739215136, + 0.007013619411736727, + 0.0038568233139812946, + 0.014141718856990337, + 0.01579446531832218, + 0.010762772522866726, + 0.013840936124324799, + 0.012441804632544518, + 0.013438727706670761, + 0.013612642884254456, + 0.012300580739974976, + 0.0043719131499528885, + 0.005130427423864603, + 0.0029771102126687765, + 0.004441515542566776, + 0.004400940146297216, + 0.0032931407913565636, + 0.0019010669784620404, + 0.0020106639713048935, + 0.007233388721942902, + 0.006431673187762499, + 0.002736602211371064, + 0.0031303949654102325, + 0.008450135588645935, + 0.010195793583989143, + 0.003259491641074419, + 0.005027082283049822, + 0.013132686726748943, + 0.00654976349323988, + 0.009858881123363972, + 0.003644546028226614, + 0.007020489778369665, + 0.010551492683589458, + 0.004471109248697758, + 0.005684694275259972, + 0.06331207603216171, + 0.12460343539714813, + 0.03498796746134758, + 0.015899114310741425, + 0.016924286261200905, + 0.0396549291908741, + 0.009937729686498642, + 0.01174379326403141, + 0.025056634098291397, + 0.012955822050571442, + 0.014504441991448402, + 0.012628784403204918, + 0.010992618277668953, + 0.014074550941586494, + 0.009280419908463955, + 0.012796775437891483, + 0.024898644536733627, + 0.023888828232884407, + 0.012569280341267586, + 0.014456206932663918, + 0.011640786193311214, + 0.01913931593298912, + 0.0072973365895450115, + 0.03200039267539978, + 0.009074856527149677, + 0.01134986337274313, + 0.012203322723507881, + 0.01119617372751236, + 0.008666764944791794, + 0.008651850745081902, + 0.005577579606324434, + 0.007377861067652702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.002815570682287216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052458494901657104 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.0052458494901657104, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052458494901657104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09802012890577316, + 0.029718665406107903, + 0.06197112426161766, + 0.045658353716135025, + 0.0695880874991417, + 0.042435262352228165, + 0.19456632435321808, + 0.06726549565792084, + 0.28491824865341187, + 0.10585830360651016 + ], + "credit_map": [ + 0.01927298679947853, + 0.007989668287336826, + 0.013942391611635685, + 0.012912602163851261, + 0.006594009231775999, + 0.012694655917584896, + 0.013881221413612366, + 0.010732591152191162, + 0.004450841806828976, + 0.004063276574015617, + 0.007074026390910149, + 0.0033568604849278927, + 0.0019978347700089216, + 0.002168356440961361, + 0.002814286621287465, + 0.0037931823171675205, + 0.008338365703821182, + 0.010274243541061878, + 0.005800290498882532, + 0.00958620011806488, + 0.00719629367813468, + 0.007161032408475876, + 0.00773088913410902, + 0.005883809644728899, + 0.006074691656976938, + 0.009032671339809895, + 0.004926625639200211, + 0.009048346430063248, + 0.005004870239645243, + 0.004285767674446106, + 0.004221209324896336, + 0.0030641742050647736, + 0.011463680304586887, + 0.007916524074971676, + 0.005228723865002394, + 0.005500191356986761, + 0.012504938058555126, + 0.014771013520658016, + 0.00508916936814785, + 0.007113845087587833, + 0.007261201273649931, + 0.004138449672609568, + 0.005154303275048733, + 0.0036213737912476063, + 0.007116714026778936, + 0.006145429331809282, + 0.004487600177526474, + 0.004510190803557634, + 0.014157081954181194, + 0.07301606982946396, + 0.055065155029296875, + 0.008343435823917389, + 0.008684328757226467, + 0.022317947819828987, + 0.005478700622916222, + 0.007503618020564318, + 0.014167692512273788, + 0.007817274890840054, + 0.005508693866431713, + 0.010552506893873215, + 0.008798498660326004, + 0.0077846297062933445, + 0.0072916001081466675, + 0.00534459576010704, + 0.06292513012886047, + 0.03341873362660408, + 0.02613130584359169, + 0.029596691951155663, + 0.02792038768529892, + 0.0565413199365139, + 0.020871616899967194, + 0.02751307189464569, + 0.009231853298842907, + 0.011900809593498707, + 0.016794411465525627, + 0.013029458932578564, + 0.017761260271072388, + 0.011701040901243687, + 0.013825353235006332, + 0.011614114046096802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0052458494901657104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot captures a futuristic chain-driven snowmobile slamming on the brakes, its heavy, exposed chain drive and cleated tracks violently churning deep into the pristine snow. Designed with a sleek, aerodynamic body composed of matte gunmetal gray composite panels and sharp, angular lines, the vehicle unleashes a vast, dazzling plume of crystalline powder that instantly engulfs it, against a breathtaking backdrop of awe-inspiring, jagged ice peaks piercing the pale, frosty sky. As the machine hums to life and falls into a profound silence, the white cloud of snow slowly dissipates, revealing its silent form. Then, in a smooth and deliberate doll-like retreat, the camera slowly pulls away from the lone vehicle. The frame gradually widens, revealing the immense scale of the surrounding landscape: a vast, silent, and completely empty expanse of untouched snow stretching to the distant mountains. The long, blue shadows cast by the low-hanging sun underscore the profound, quiet solitude of the vast winter wilderness.", + "seed": 20355, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20355_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": -0.044644896779209374, + "vlm_logprob": -0.044644896779209374, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.044644896779209374, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05644238367676735 + ], + "opposite_logprob_tokens": [ + -2.9314424991607666 + ], + "logprob_reward": -0.05644238367676735, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05644238367676735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1509397327899933, + 0.04009781777858734, + 0.1780814230442047, + 0.0738319531083107, + 0.10099409520626068, + 0.11876646429300308, + 0.05033251643180847, + 0.03860629349946976, + 0.056155573576688766, + 0.1921941190958023 + ], + "credit_map": [ + 0.009889723733067513, + 0.024673212319612503, + 0.02595990151166916, + 0.021995846182107925, + 0.027196252718567848, + 0.01423827838152647, + 0.012653321027755737, + 0.014333189465105534, + 0.004960211925208569, + 0.0059972722083330154, + 0.0029427523259073496, + 0.004008056595921516, + 0.007853937335312366, + 0.004800021182745695, + 0.004318966995924711, + 0.005216601770371199, + 0.05716085433959961, + 0.00970444455742836, + 0.010586773045361042, + 0.006228714715689421, + 0.03576204180717468, + 0.01202555000782013, + 0.00775185925886035, + 0.038861192762851715, + 0.009668980725109577, + 0.012802783399820328, + 0.009379669092595577, + 0.007736754138022661, + 0.008531621657311916, + 0.011246412992477417, + 0.0057711293920874596, + 0.0086946040391922, + 0.012385930866003036, + 0.00902328360825777, + 0.020961610600352287, + 0.011934749782085419, + 0.013063441030681133, + 0.017652183771133423, + 0.009002522565424442, + 0.00697036599740386, + 0.013884189538657665, + 0.01370984222739935, + 0.017019737511873245, + 0.01353661622852087, + 0.021686991676688194, + 0.018664531409740448, + 0.009041283279657364, + 0.01122326124459505, + 0.005889035761356354, + 0.00682874396443367, + 0.004424109123647213, + 0.006045854650437832, + 0.010979574173688889, + 0.007312450557947159, + 0.004248900804668665, + 0.00460384925827384, + 0.0058813258074223995, + 0.005040801130235195, + 0.004693858325481415, + 0.0036566469352692366, + 0.006320693530142307, + 0.004519648849964142, + 0.0038607867900282145, + 0.004632532596588135, + 0.00454788189381361, + 0.007087240926921368, + 0.0070262486115098, + 0.00896591879427433, + 0.0050599356181919575, + 0.0077566285617649555, + 0.006709875073283911, + 0.009001845493912697, + 0.02347942627966404, + 0.009734674356877804, + 0.04990820959210396, + 0.009758190251886845, + 0.009225703775882721, + 0.0728212371468544, + 0.007192647084593773, + 0.010074022226035595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.05644238367676735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129358634352684 + ], + "opposite_logprob_tokens": [ + -2.237935781478882 + ], + "logprob_reward": -0.1129358634352684, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129358634352684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06456763297319412, + 0.07207192480564117, + 0.2873261570930481, + 0.29939335584640503, + 0.04332059621810913, + 0.04395386204123497, + 0.030037228018045425, + 0.02046205662190914, + 0.04162516072392464, + 0.09724202007055283 + ], + "credit_map": [ + 0.00557509483769536, + 0.009992324747145176, + 0.011003797873854637, + 0.006854370702058077, + 0.007737953681498766, + 0.009007652290165424, + 0.006310639902949333, + 0.008085798472166061, + 0.010293797589838505, + 0.009111284278333187, + 0.006978364195674658, + 0.005652395077049732, + 0.010804563760757446, + 0.0070039695128798485, + 0.008480419404804707, + 0.013747131451964378, + 0.0867004469037056, + 0.018127912655472755, + 0.019649367779493332, + 0.010364165529608727, + 0.029716936871409416, + 0.027120674028992653, + 0.012726851738989353, + 0.08291978389024734, + 0.03676760196685791, + 0.04337036982178688, + 0.03959698602557182, + 0.03500097244977951, + 0.042261186987161636, + 0.0286969393491745, + 0.026491530239582062, + 0.047207798808813095, + 0.005686609540134668, + 0.004027956165373325, + 0.010915092192590237, + 0.004641946405172348, + 0.005503420252352953, + 0.005508980248123407, + 0.004496221896260977, + 0.002540369750931859, + 0.006063474342226982, + 0.008380277082324028, + 0.004843639675527811, + 0.0064067840576171875, + 0.0071328431367874146, + 0.004506842233240604, + 0.0038356243167072535, + 0.0027843753341585398, + 0.003858352778479457, + 0.004500216338783503, + 0.003200496081262827, + 0.003647956531494856, + 0.004421298857778311, + 0.004371543414890766, + 0.002372132148593664, + 0.0036652309354394674, + 0.0041727633215487, + 0.001954869134351611, + 0.001564705977216363, + 0.0016779908910393715, + 0.002372132847085595, + 0.002666509710252285, + 0.0021900306455790997, + 0.003863054793328047, + 0.0040266807191073895, + 0.00692617567256093, + 0.004590120166540146, + 0.005863912403583527, + 0.005128567107021809, + 0.005100511945784092, + 0.005144656635820866, + 0.0048445360735058784, + 0.01546680647879839, + 0.006265430245548487, + 0.02016185224056244, + 0.004690479952841997, + 0.005505499895662069, + 0.03716994449496269, + 0.003408279735594988, + 0.004573729820549488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.1129358634352684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369579464197159 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.03369579464197159, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369579464197159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2756970524787903, + 0.0891651138663292, + 0.212485209107399, + 0.1025494858622551, + 0.043216194957494736, + 0.061234187334775925, + 0.10404760390520096, + 0.03770215064287186, + 0.03407025337219238, + 0.039832714945077896 + ], + "credit_map": [ + 0.023602863773703575, + 0.056496843695640564, + 0.04845946282148361, + 0.028419647365808487, + 0.046029314398765564, + 0.023211980238556862, + 0.026820732280611992, + 0.02265624888241291, + 0.013082658872008324, + 0.008688719943165779, + 0.007278235163539648, + 0.006578627973794937, + 0.01193994004279375, + 0.010261432267725468, + 0.012596125714480877, + 0.018739385530352592, + 0.0271291583776474, + 0.0148321483284235, + 0.013052466325461864, + 0.016262618824839592, + 0.04856984317302704, + 0.021957002580165863, + 0.009087124839425087, + 0.061594851315021515, + 0.0140693923458457, + 0.01655837520956993, + 0.013291390612721443, + 0.009928715415298939, + 0.012631385587155819, + 0.01560442615300417, + 0.009317544288933277, + 0.011148268356919289, + 0.006280642468482256, + 0.004199805203825235, + 0.010410171002149582, + 0.00342025188729167, + 0.005250907968729734, + 0.0071487766690552235, + 0.003461375366896391, + 0.0030442695133388042, + 0.00941760279238224, + 0.009126460179686546, + 0.008100789040327072, + 0.005422181915491819, + 0.009465083479881287, + 0.009541724808514118, + 0.006074387580156326, + 0.0040859654545784, + 0.012137897312641144, + 0.01957767829298973, + 0.009921560995280743, + 0.012179110199213028, + 0.019806791096925735, + 0.013956367038190365, + 0.008109807968139648, + 0.008358395658433437, + 0.005595523398369551, + 0.004175232723355293, + 0.0050093065947294235, + 0.005610685329884291, + 0.004083428997546434, + 0.00419031735509634, + 0.003711473662406206, + 0.005326184909790754, + 0.004840502981096506, + 0.0034908659290522337, + 0.004149659536778927, + 0.006563943345099688, + 0.003248575609177351, + 0.004713691305369139, + 0.004849640186876059, + 0.002213379368185997, + 0.007900776341557503, + 0.004317157436162233, + 0.0033094054087996483, + 0.002317565493285656, + 0.004183149430900812, + 0.012838144786655903, + 0.0026116499211639166, + 0.002354868920519948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.03369579464197159 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004090989474207163 + ], + "opposite_logprob_tokens": [ + -5.504090785980225 + ], + "logprob_reward": -0.004090989474207163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004090989474207163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21196158230304718, + 0.07377767562866211, + 0.3339587450027466, + 0.09719850867986679, + 0.06375553458929062, + 0.05750009045004845, + 0.054099924862384796, + 0.033476706594228745, + 0.030034974217414856, + 0.04423626884818077 + ], + "credit_map": [ + 0.022811930626630783, + 0.033496856689453125, + 0.04148856922984123, + 0.021715596318244934, + 0.031723443418741226, + 0.017955461516976357, + 0.017561709508299828, + 0.025208009406924248, + 0.008815095759928226, + 0.01014722604304552, + 0.005555432755500078, + 0.006176969036459923, + 0.01619494892656803, + 0.01063748449087143, + 0.008334314450621605, + 0.007916200906038284, + 0.10752025246620178, + 0.01719732955098152, + 0.016311341896653175, + 0.010021684691309929, + 0.05423438176512718, + 0.03237208351492882, + 0.014429997652769089, + 0.08187167346477509, + 0.014530088752508163, + 0.010409734211862087, + 0.013956940732896328, + 0.00842177402228117, + 0.010452915914356709, + 0.015437676571309566, + 0.012384424917399883, + 0.01160494890064001, + 0.007306620944291353, + 0.007670840714126825, + 0.012574572116136551, + 0.006541695911437273, + 0.007672020699828863, + 0.011943366378545761, + 0.005632875487208366, + 0.0044135404750704765, + 0.006988960783928633, + 0.006757482886314392, + 0.007302116136997938, + 0.0075681391172111034, + 0.011700094677507877, + 0.009174113161861897, + 0.004254307132214308, + 0.0037548765540122986, + 0.00664979824796319, + 0.007797159720212221, + 0.005097715184092522, + 0.006326581817120314, + 0.0098942955955863, + 0.008435675874352455, + 0.004562458489090204, + 0.005336237605661154, + 0.0060416474007070065, + 0.0030615406576544046, + 0.004389462061226368, + 0.0024307607673108578, + 0.003531266003847122, + 0.004161616321653128, + 0.002606174908578396, + 0.007254236843436956, + 0.003556097624823451, + 0.0035334692802280188, + 0.0029365113005042076, + 0.006184616591781378, + 0.0024965328630059958, + 0.0041289133951067924, + 0.003572125919163227, + 0.0036267037503421307, + 0.005530830938369036, + 0.004923912696540356, + 0.013554339297115803, + 0.004050736781209707, + 0.005687325727194548, + 0.004806398879736662, + 0.0033323545940220356, + 0.002350366674363613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.004090989474207163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059452667832375 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059452667832375, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059452667832375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11484633386135101, + 0.10372090339660645, + 0.4587023854255676, + 0.04194308817386627, + 0.025829430669546127, + 0.022495107725262642, + 0.03799223527312279, + 0.014735150150954723, + 0.01359557919204235, + 0.16613976657390594 + ], + "credit_map": [ + 0.010978383012115955, + 0.022986220195889473, + 0.018104907125234604, + 0.011618068441748619, + 0.01978728361427784, + 0.010262998752295971, + 0.012377886101603508, + 0.008730585686862469, + 0.012344716116786003, + 0.01208556815981865, + 0.007742098066955805, + 0.007672597654163837, + 0.017287353053689003, + 0.009912046603858471, + 0.014735027216374874, + 0.021941497921943665, + 0.12284939736127853, + 0.03395146504044533, + 0.03177977725863457, + 0.020268438383936882, + 0.02746650017797947, + 0.04279927909374237, + 0.02271796204149723, + 0.15686960518360138, + 0.0044860742054879665, + 0.0049026720225811005, + 0.0054796538315713406, + 0.004908744711428881, + 0.006792737171053886, + 0.006576888728886843, + 0.004364090505987406, + 0.004432228859513998, + 0.004437311086803675, + 0.0032746719662100077, + 0.007125303149223328, + 0.001886736718006432, + 0.0024837153032422066, + 0.0034626484848558903, + 0.001617297064512968, + 0.0015417482936754823, + 0.00307800667360425, + 0.0031021512113511562, + 0.003342469222843647, + 0.0029961219988763332, + 0.0022758389823138714, + 0.0036687098909169436, + 0.002201272640377283, + 0.0018305388512089849, + 0.004195150453597307, + 0.006169023457914591, + 0.003165923058986664, + 0.004597648046910763, + 0.007450401782989502, + 0.006302393041551113, + 0.003187219612300396, + 0.00292447442188859, + 0.0018474312964826822, + 0.001884808181785047, + 0.0015972924884408712, + 0.0015026091132313013, + 0.002558268839493394, + 0.0019045135704800487, + 0.0015075147384777665, + 0.0019327132031321526, + 0.001343346550129354, + 0.001553931157104671, + 0.0014590518549084663, + 0.002220196183770895, + 0.001545302220620215, + 0.0017885179258883, + 0.0020805932581424713, + 0.0016046396922320127, + 0.022445153445005417, + 0.008120066486299038, + 0.040685590356588364, + 0.008063795045018196, + 0.006929995957762003, + 0.06548944860696793, + 0.005612073000520468, + 0.008793661370873451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.016059452667832375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot captures a futuristic chain-driven snowmobile slamming on the brakes, its heavy, exposed chain drive and cleated tracks violently churning deep into the pristine snow. Designed with a sleek, aerodynamic body composed of matte gunmetal gray composite panels and sharp, angular lines, the vehicle unleashes a vast, dazzling plume of crystalline powder that instantly engulfs it, against a breathtaking backdrop of awe-inspiring, jagged ice peaks piercing the pale, frosty sky. As the machine hums to life and falls into a profound silence, the white cloud of snow slowly dissipates, revealing its silent form. Then, in a smooth and deliberate doll-like retreat, the camera slowly pulls away from the lone vehicle. The frame gradually widens, revealing the immense scale of the surrounding landscape: a vast, silent, and completely empty expanse of untouched snow stretching to the distant mountains. The long, blue shadows cast by the low-hanging sun underscore the profound, quiet solitude of the vast winter wilderness.", + "seed": 20356, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_7/index_9167_seed_20356_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": -0.05542593197897076, + "vlm_logprob": -0.05542593197897076, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.05542593197897076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08203741908073425 + ], + "opposite_logprob_tokens": [ + -2.5820374488830566 + ], + "logprob_reward": -0.08203741908073425, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08203741908073425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06630509346723557, + 0.10175035148859024, + 0.0290688369423151, + 0.09969570487737656, + 0.10493773967027664, + 0.074781633913517, + 0.2835025191307068, + 0.17466233670711517, + 0.025006311014294624, + 0.040289461612701416 + ], + "credit_map": [ + 0.010087690316140652, + 0.01547300722450018, + 0.009035284630954266, + 0.0092460410669446, + 0.005273611284792423, + 0.007903173565864563, + 0.006034217309206724, + 0.003252072026953101, + 0.011206080205738544, + 0.008363503031432629, + 0.007382557261735201, + 0.007029090076684952, + 0.00918227806687355, + 0.023299584165215492, + 0.027222948148846626, + 0.008064310997724533, + 0.004843621980398893, + 0.00441362988203764, + 0.002524556126445532, + 0.002438253490254283, + 0.0038069207221269608, + 0.003958744928240776, + 0.004212051630020142, + 0.0028710553888231516, + 0.014675725251436234, + 0.017326518893241882, + 0.009781813248991966, + 0.009406168013811111, + 0.015514352358877659, + 0.012417043559253216, + 0.010892340913414955, + 0.009681749157607555, + 0.012244892306625843, + 0.0147400526329875, + 0.015323948115110397, + 0.011905521154403687, + 0.021442167460918427, + 0.011186265386641026, + 0.008888252079486847, + 0.009206640534102917, + 0.00955005269497633, + 0.008040416054427624, + 0.008527164347469807, + 0.014980894513428211, + 0.00994474720209837, + 0.007955794222652912, + 0.006463761441409588, + 0.009318803437054157, + 0.028364073485136032, + 0.056510865688323975, + 0.039846278727054596, + 0.038945138454437256, + 0.038055505603551865, + 0.0243045836687088, + 0.0347360335290432, + 0.02274002693593502, + 0.012326636351644993, + 0.03396987542510033, + 0.050977177917957306, + 0.020660877227783203, + 0.016125937923789024, + 0.010830766521394253, + 0.012434100732207298, + 0.017336968332529068, + 0.0024744842667132616, + 0.003915755078196526, + 0.0032457509078085423, + 0.0033354994375258684, + 0.0029408263508230448, + 0.0025710086338222027, + 0.0030644224025309086, + 0.003458563471212983, + 0.0033662416972219944, + 0.004859830252826214, + 0.00601767236366868, + 0.005088078789412975, + 0.0049240244552493095, + 0.005568892695009708, + 0.004993201699107885, + 0.005471518263220787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.08203741908073425 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987661063671112 + ], + "opposite_logprob_tokens": [ + -1.8048765659332275 + ], + "logprob_reward": -0.17987661063671112, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987661063671112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03599793463945389, + 0.05193770304322243, + 0.033008817583322525, + 0.035122454166412354, + 0.07604645937681198, + 0.1539391130208969, + 0.3666020929813385, + 0.16669194400310516, + 0.05258854106068611, + 0.028064919635653496 + ], + "credit_map": [ + 0.006407302804291248, + 0.004222603049129248, + 0.004976453725248575, + 0.004850446712225676, + 0.004495871253311634, + 0.0043515102006495, + 0.004068640060722828, + 0.0026251068338751793, + 0.006284487433731556, + 0.004288753494620323, + 0.003888497594743967, + 0.0034416934940963984, + 0.006485631223767996, + 0.011983941309154034, + 0.011545807123184204, + 0.00401888694614172, + 0.0059960284270346165, + 0.0034862428437918425, + 0.0025354649405926466, + 0.002165713347494602, + 0.006835072301328182, + 0.0035637582186609507, + 0.005226725712418556, + 0.003199812490493059, + 0.0048997001722455025, + 0.004842870403081179, + 0.003670717356726527, + 0.003787731984630227, + 0.00639858515933156, + 0.004620586987584829, + 0.0033725237008184195, + 0.003529738634824753, + 0.008931080810725689, + 0.01223691739141941, + 0.007516669575124979, + 0.009346785955131054, + 0.014372269622981548, + 0.009534654207527637, + 0.005061400588601828, + 0.009046687744557858, + 0.01981460116803646, + 0.015389502979815006, + 0.0181717686355114, + 0.03514854982495308, + 0.01698991283774376, + 0.012075409293174744, + 0.015034130774438381, + 0.021315239369869232, + 0.0368378572165966, + 0.06788191199302673, + 0.05142119526863098, + 0.0495934933423996, + 0.047869857400655746, + 0.03160037472844124, + 0.04822296276688576, + 0.03317444398999214, + 0.016163697466254234, + 0.020295271649956703, + 0.04096204787492752, + 0.02477261610329151, + 0.01719210296869278, + 0.013541700318455696, + 0.013923305086791515, + 0.019841192290186882, + 0.005658193025738001, + 0.007400476839393377, + 0.005201136227697134, + 0.00589113449677825, + 0.006369466427713633, + 0.005819087848067284, + 0.006406601518392563, + 0.009842442348599434, + 0.0024538380093872547, + 0.003528516972437501, + 0.00415012938901782, + 0.0027612601406872272, + 0.004203302785754204, + 0.0034881513565778732, + 0.003607703372836113, + 0.003872017376124859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.17987661063671112 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463874451816082 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.00463874451816082, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463874451816082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08695943653583527, + 0.05479539930820465, + 0.02024204656481743, + 0.029522281140089035, + 0.035106509923934937, + 0.15089061856269836, + 0.4335261881351471, + 0.09375754743814468, + 0.033205945044755936, + 0.061994075775146484 + ], + "credit_map": [ + 0.010308497585356236, + 0.0165626872330904, + 0.010710161179304123, + 0.013004874810576439, + 0.01169579103589058, + 0.009795712307095528, + 0.009670096449553967, + 0.00521161500364542, + 0.006640170235186815, + 0.004377918783575296, + 0.004023849498480558, + 0.0035169583279639482, + 0.006394858937710524, + 0.011968923732638359, + 0.014096925966441631, + 0.003775792196393013, + 0.003309298772364855, + 0.0030256016179919243, + 0.0018178033642470837, + 0.0023686420172452927, + 0.002991556655615568, + 0.0022056035231798887, + 0.0027182206977158785, + 0.0018053187523037195, + 0.004087470006197691, + 0.0037672342732548714, + 0.0035404020454734564, + 0.0033392608165740967, + 0.005723702721297741, + 0.0031166488770395517, + 0.0030000563710927963, + 0.002947504399344325, + 0.0044630407355725765, + 0.005169505253434181, + 0.004512820392847061, + 0.00471873302012682, + 0.003785255830734968, + 0.004599710460752249, + 0.002901506843045354, + 0.004955940879881382, + 0.020340697839856148, + 0.015557356178760529, + 0.018177058547735214, + 0.04273616522550583, + 0.015695316717028618, + 0.011425310745835304, + 0.013835521414875984, + 0.01312319003045559, + 0.04428599774837494, + 0.0840413048863411, + 0.06439491361379623, + 0.05744029954075813, + 0.05508913844823837, + 0.0364808589220047, + 0.05541028454899788, + 0.036383382976055145, + 0.006477203220129013, + 0.014528742991387844, + 0.01855376549065113, + 0.014289422892034054, + 0.011508915573358536, + 0.007753979880362749, + 0.009776340797543526, + 0.010869174264371395, + 0.003307166276499629, + 0.00467534177005291, + 0.0051447441801428795, + 0.004624930676072836, + 0.0035074972547590733, + 0.003195724915713072, + 0.004338630009442568, + 0.004411912988871336, + 0.004753425717353821, + 0.009726984426379204, + 0.006895388010889292, + 0.007158633787184954, + 0.007578992284834385, + 0.009054289199411869, + 0.007527503184974194, + 0.009298859164118767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.00463874451816082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944311618804932 + ], + "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": [ + 0.058127544820308685, + 0.040346454828977585, + 0.01625622622668743, + 0.03412049636244774, + 0.1118108406662941, + 0.07515673339366913, + 0.13584789633750916, + 0.07327192276716232, + 0.026806052774190903, + 0.4282557964324951 + ], + "credit_map": [ + 0.00853890459984541, + 0.006066615227609873, + 0.00688848877325654, + 0.007241915911436081, + 0.007377864792943001, + 0.008638309314846992, + 0.006928283721208572, + 0.0064471689984202385, + 0.0053231376223266125, + 0.0034285690635442734, + 0.003258933313190937, + 0.0032520482782274485, + 0.004299555439502001, + 0.008600900880992413, + 0.009497650898993015, + 0.0026856649201363325, + 0.0020916075445711613, + 0.002060920000076294, + 0.0017805886454880238, + 0.001577600254677236, + 0.0023121593985706568, + 0.002113881055265665, + 0.002285658149048686, + 0.002033812692388892, + 0.003990551922470331, + 0.004525710828602314, + 0.004126387648284435, + 0.0038544617127627134, + 0.006874739192426205, + 0.003683380549773574, + 0.0036836008075624704, + 0.003381666960194707, + 0.016115227714180946, + 0.009060101583600044, + 0.020876623690128326, + 0.010289600118994713, + 0.01395961083471775, + 0.016826171427965164, + 0.008262010291218758, + 0.01642150804400444, + 0.009261353872716427, + 0.009443950839340687, + 0.009282054379582405, + 0.02183811366558075, + 0.005600865930318832, + 0.006947070825845003, + 0.005182165186852217, + 0.007601163350045681, + 0.012704627588391304, + 0.028570478782057762, + 0.015617292374372482, + 0.013015826232731342, + 0.026598874479532242, + 0.01352359913289547, + 0.013400659896433353, + 0.012416556477546692, + 0.006713151466101408, + 0.009432117454707623, + 0.010910242795944214, + 0.00939276535063982, + 0.008522124029695988, + 0.007893241010606289, + 0.007052444852888584, + 0.013355839066207409, + 0.0027752513997256756, + 0.004940425511449575, + 0.003633513581007719, + 0.004269869532436132, + 0.002455918351188302, + 0.0021554604172706604, + 0.0031309325713664293, + 0.0034446853678673506, + 0.034130748361349106, + 0.0681460052728653, + 0.045124802738428116, + 0.05108391121029854, + 0.051795970648527145, + 0.06409335136413574, + 0.04992779344320297, + 0.06395325809717178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.005944311618804932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a futuristic snowmobile with a sleek, aerodynamic body composed of matte gunmetal gray composite panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the snowmobile slam on its brakes, causing its heavy exposed chain drive and cleated tracks to violently churn deep into the snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the vehicle unleash a vast, dazzling plume of crystalline powder that instantly engulfs it before slowly dissipating?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast expanse of untouched snow stretching to distant, jagged ice peaks under a pale, frosty sky with long blue shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a smooth, deliberate pull-away (zoom out) from the vehicle to reveal the immense scale of the empty landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004632574040442705 + ], + "opposite_logprob_tokens": [ + -5.379632472991943 + ], + "logprob_reward": -0.004632574040442705, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004632574040442705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04772128164768219, + 0.04873376339673996, + 0.0223232414573431, + 0.03184771537780762, + 0.06813240051269531, + 0.11537495255470276, + 0.2074888050556183, + 0.2946690022945404, + 0.10972903668880463, + 0.05397980287671089 + ], + "credit_map": [ + 0.006858309730887413, + 0.010227012448012829, + 0.006818283349275589, + 0.006379659287631512, + 0.005360767245292664, + 0.006639924366027117, + 0.00343813537620008, + 0.00199919193983078, + 0.0061845602467656136, + 0.005053987260907888, + 0.004220136906951666, + 0.003679354675114155, + 0.005417168606072664, + 0.01013143640011549, + 0.009901361539959908, + 0.004145764280110598, + 0.003692889353260398, + 0.002959893085062504, + 0.00222263322211802, + 0.002864647191017866, + 0.003361618844792247, + 0.0021566739305853844, + 0.0026860570069402456, + 0.0023788304533809423, + 0.003028790932148695, + 0.00391841446980834, + 0.0026430292055010796, + 0.0035998441744595766, + 0.007126178126782179, + 0.003862411715090275, + 0.004059405066072941, + 0.0036096456460654736, + 0.009275395423173904, + 0.005701432470232248, + 0.009042349644005299, + 0.008824114687740803, + 0.005888304673135281, + 0.010280718095600605, + 0.00596258882433176, + 0.013157499954104424, + 0.017382703721523285, + 0.014691425487399101, + 0.014520369470119476, + 0.025239229202270508, + 0.008249306119978428, + 0.011935308575630188, + 0.009968852624297142, + 0.013387768529355526, + 0.0262131504714489, + 0.03253821283578873, + 0.028383834287524223, + 0.02784854732453823, + 0.03157127648591995, + 0.017686886712908745, + 0.02145099639892578, + 0.021795909851789474, + 0.02472125180065632, + 0.046928804367780685, + 0.032862983644008636, + 0.03737668693065643, + 0.03439047932624817, + 0.035188574343919754, + 0.026911607012152672, + 0.05628861486911774, + 0.00968818087130785, + 0.01986047811806202, + 0.013420871458947659, + 0.015008198097348213, + 0.010803389362990856, + 0.010284660384058952, + 0.011651837266981602, + 0.01901143789291382, + 0.0036021247506141663, + 0.007839107885956764, + 0.006979402620345354, + 0.006704041734337807, + 0.007345563732087612, + 0.008356297388672829, + 0.005384528078138828, + 0.007768741808831692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.004632574040442705 + } + ], + "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/0000158/group_8/index_tensor([5750, 5750])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_tensor([5750, 5750])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45f62f23913afeef102c6c17e837162083bcab77 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_tensor([5750, 5750])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures an intimate moment of passionate conversation between two friends in a dimly lit, cozy bistro. A woman in her twenties, with wavy auburn hair loosely pulled back and wearing a soft cream-colored sweater, leans forward, a gentle smile on her face as she speaks. Across from her, a man of similar age, wearing a dark green button-down shirt with the sleeves casually rolled up to his forearms, listens intently, his head slightly tilted, a look of genuine engagement. Between them, on a small, round, dark wooden table, lie two elegant, large glasses half-filled with deep red wine. As the woman shares her story, her right hand gracefully and expressively gestures near the bowl of her glass, while the man's hand rests near the base of his own, his fingers leisurely tracing the polished wood. The entire scene is bathed in the warm golden light of a low-hanging lamp, casting soft shadows on their faces and causing the wine to gleam, creating a tranquil and welcoming atmosphere of easy friendship and shared confidence.", + "seed": 15303, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_5750_seed_15303_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": -0.015950530767440796, + "vlm_logprob": -0.015950530767440796, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.015950530767440796, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03396120294928551 + ], + "opposite_logprob_tokens": [ + -4.908961296081543 + ], + "logprob_reward": -0.03396120294928551, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.03396120294928551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23328913748264313, + 0.08604467660188675, + 0.05924958735704422, + 0.05689576268196106, + 0.15184727311134338, + 0.04153311625123024, + 0.08861374109983444, + 0.10269936919212341, + 0.08144409954547882, + 0.09838318824768066 + ], + "credit_map": [ + 0.031047215685248375, + 0.042018745094537735, + 0.017649270594120026, + 0.028966614976525307, + 0.03358181193470955, + 0.02306864969432354, + 0.024595066905021667, + 0.032361775636672974, + 0.012325005605816841, + 0.013825901784002781, + 0.009719669818878174, + 0.01085043977946043, + 0.011797499842941761, + 0.008116818964481354, + 0.00798422284424305, + 0.01142512820661068, + 0.009850694797933102, + 0.007032311987131834, + 0.007314028684049845, + 0.007585471961647272, + 0.006388770882040262, + 0.00604369817301631, + 0.00574783980846405, + 0.009286772459745407, + 0.008696895092725754, + 0.005920769181102514, + 0.006920702289789915, + 0.007710836827754974, + 0.009829098358750343, + 0.005411256104707718, + 0.0055613466538488865, + 0.0068448614329099655, + 0.020497450605034828, + 0.025539813563227654, + 0.015357744880020618, + 0.02469533681869507, + 0.014790567569434643, + 0.01394121628254652, + 0.014354038052260876, + 0.022671109065413475, + 0.005072661675512791, + 0.0049640703946352005, + 0.0044289883226156235, + 0.006276790983974934, + 0.0069144656881690025, + 0.0036247826647013426, + 0.0039055272936820984, + 0.006345831789076328, + 0.011327107436954975, + 0.017278457060456276, + 0.007082326803356409, + 0.012399796396493912, + 0.011129958555102348, + 0.009883667342364788, + 0.008306954056024551, + 0.011205481365323067, + 0.011721008457243443, + 0.01066422276198864, + 0.01294285710901022, + 0.015783891081809998, + 0.01090304832905531, + 0.009343680925667286, + 0.011567611247301102, + 0.01977306418120861, + 0.009620588272809982, + 0.017735442146658897, + 0.006873170379549265, + 0.0122091518715024, + 0.009839516133069992, + 0.009288453496992588, + 0.006966065615415573, + 0.008911717683076859, + 0.01230782549828291, + 0.01784815825521946, + 0.009248068556189537, + 0.016092101112008095, + 0.009621040895581245, + 0.009395888075232506, + 0.010263968259096146, + 0.013606148771941662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.03396120294928551 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032195420935750008 + ], + "opposite_logprob_tokens": [ + -5.7532196044921875 + ], + "logprob_reward": -0.0032195420935750008, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032195420935750008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25787419080734253, + 0.1550108641386032, + 0.10511062294244766, + 0.055634040385484695, + 0.07436393201351166, + 0.09537193179130554, + 0.06432610750198364, + 0.08774124830961227, + 0.044164687395095825, + 0.06040237098932266 + ], + "credit_map": [ + 0.03425118699669838, + 0.03522675484418869, + 0.030113311484456062, + 0.029894426465034485, + 0.03240026533603668, + 0.024812836199998856, + 0.02362416312098503, + 0.047551229596138, + 0.019487403333187103, + 0.02230328507721424, + 0.02001802623271942, + 0.018289120867848396, + 0.02090022899210453, + 0.013316097669303417, + 0.017259377986192703, + 0.023437323048710823, + 0.011966992169618607, + 0.02138926461338997, + 0.009456995874643326, + 0.019842013716697693, + 0.012705034576356411, + 0.007343614939600229, + 0.00968025904148817, + 0.012726454064249992, + 0.0060770465061068535, + 0.007621448952704668, + 0.00618561590090394, + 0.007279589306563139, + 0.0077872988767921925, + 0.004532476421445608, + 0.007329440210014582, + 0.008821122348308563, + 0.007913535460829735, + 0.00850363727658987, + 0.007281731814146042, + 0.010298463515937328, + 0.010924928821623325, + 0.007785428315401077, + 0.009096103720366955, + 0.0125601040199399, + 0.01078706607222557, + 0.01697460003197193, + 0.0069866180419921875, + 0.016415299847722054, + 0.011133436113595963, + 0.007649014238268137, + 0.010544555261731148, + 0.014881337992846966, + 0.00685837771743536, + 0.013740596361458302, + 0.005931262858211994, + 0.007415229920297861, + 0.010740661062300205, + 0.006167456042021513, + 0.007057200651615858, + 0.00641532801091671, + 0.010137590579688549, + 0.01229459885507822, + 0.010811676271259785, + 0.01147601380944252, + 0.009823163039982319, + 0.01009612251073122, + 0.008964711800217628, + 0.014137370511889458, + 0.006069117225706577, + 0.005623115226626396, + 0.004773041233420372, + 0.005397648550570011, + 0.0047125075943768024, + 0.006145116873085499, + 0.006279491353780031, + 0.00516465213149786, + 0.006305774208158255, + 0.010001365095376968, + 0.005099335219711065, + 0.009212602861225605, + 0.00782836228609085, + 0.004884619731456041, + 0.007343691773712635, + 0.009726621210575104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0032195420935750008 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41756024956703186, + 0.1024235188961029, + 0.06928684562444687, + 0.06900444626808167, + 0.07952404022216797, + 0.042557328939437866, + 0.0825413390994072, + 0.048281703144311905, + 0.05051862448453903, + 0.03830181062221527 + ], + "credit_map": [ + 0.04866486415266991, + 0.06852055341005325, + 0.04530346021056175, + 0.07109362632036209, + 0.05307170748710632, + 0.03885414078831673, + 0.0411837212741375, + 0.05086821690201759, + 0.011905970051884651, + 0.01254703663289547, + 0.012537744827568531, + 0.016946233808994293, + 0.01329529844224453, + 0.011900067329406738, + 0.010078770108520985, + 0.01321240235120058, + 0.00889764167368412, + 0.010600254870951176, + 0.008080780506134033, + 0.010319595225155354, + 0.00832651648670435, + 0.008281338959932327, + 0.006249569822102785, + 0.00853115413337946, + 0.008237604983150959, + 0.01064314879477024, + 0.0079279113560915, + 0.00858631543815136, + 0.008712072856724262, + 0.007313132751733065, + 0.00881703570485115, + 0.008767230436205864, + 0.009489071555435658, + 0.01042015291750431, + 0.007302065379917622, + 0.01568533480167389, + 0.00918659195303917, + 0.006340745836496353, + 0.010408001951873302, + 0.010692084208130836, + 0.0053754388354718685, + 0.00497733848169446, + 0.004959339741617441, + 0.006920674350112677, + 0.006708945147693157, + 0.004511155653744936, + 0.004574101883918047, + 0.004530339036136866, + 0.00923328846693039, + 0.016219399869441986, + 0.011272105388343334, + 0.012061022222042084, + 0.009497132152318954, + 0.007058922667056322, + 0.009032761678099632, + 0.008166715502738953, + 0.005583911202847958, + 0.00910823792219162, + 0.0045663099735975266, + 0.006168128922581673, + 0.005863317754119635, + 0.006333646830171347, + 0.00433847913518548, + 0.006319671403616667, + 0.0067795803770422935, + 0.007443973328918219, + 0.005281898658722639, + 0.006947559770196676, + 0.004681944847106934, + 0.006933405064046383, + 0.00662080617621541, + 0.005829463712871075, + 0.00522516667842865, + 0.005976331885904074, + 0.00411659199744463, + 0.005748181138187647, + 0.0042116460390388966, + 0.0038197976537048817, + 0.004545648116618395, + 0.004658451769500971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.005258656572550535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369821608066559 + ], + "opposite_logprob_tokens": [ + -3.408698320388794 + ], + "logprob_reward": -0.03369821608066559, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369821608066559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2325231432914734, + 0.08685950934886932, + 0.08901501446962357, + 0.06296569108963013, + 0.05362818390130997, + 0.06490550935268402, + 0.22056682407855988, + 0.02650163508951664, + 0.06776848435401917, + 0.09526605904102325 + ], + "credit_map": [ + 0.026556575670838356, + 0.0385717935860157, + 0.02105857990682125, + 0.038096703588962555, + 0.026899276301264763, + 0.014479054138064384, + 0.027243517339229584, + 0.0396176315844059, + 0.011432847008109093, + 0.01423447486013174, + 0.009864618070423603, + 0.014454074203968048, + 0.011893250048160553, + 0.00576804531738162, + 0.008327205665409565, + 0.010884989984333515, + 0.011237791739404202, + 0.015548013150691986, + 0.008807442151010036, + 0.015598601661622524, + 0.010374720208346844, + 0.0056519703939557076, + 0.009719889611005783, + 0.01207659114152193, + 0.006947523448616266, + 0.009246857836842537, + 0.005787602160125971, + 0.01139706838876009, + 0.006782205775380135, + 0.004765246529132128, + 0.00939964409917593, + 0.008639547042548656, + 0.006247036624699831, + 0.009822053834795952, + 0.005116166081279516, + 0.005116728134453297, + 0.008904152549803257, + 0.004609966184943914, + 0.005462811794131994, + 0.00834926962852478, + 0.007715353276580572, + 0.008596424013376236, + 0.006485247053205967, + 0.009969722479581833, + 0.007659680675715208, + 0.0057235900312662125, + 0.008074798621237278, + 0.010680696927011013, + 0.022949116304516792, + 0.056716494262218475, + 0.019457539543509483, + 0.03113231621682644, + 0.030604807659983635, + 0.017175788059830666, + 0.018687525764107704, + 0.02384323626756668, + 0.003315480425953865, + 0.003776111640036106, + 0.0035828740801662207, + 0.0027872847858816385, + 0.004030017182230949, + 0.002992873778566718, + 0.0022844949271529913, + 0.003732498502358794, + 0.009332524612545967, + 0.009054092690348625, + 0.008192742243409157, + 0.010231777094304562, + 0.006358894519507885, + 0.007549759931862354, + 0.00884347502142191, + 0.008205217309296131, + 0.011664331890642643, + 0.0134646687656641, + 0.009505758993327618, + 0.018991796299815178, + 0.010030992329120636, + 0.0070801363326609135, + 0.009855168871581554, + 0.014673207886517048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.03369821608066559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615036141127348 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.003615036141127348, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615036141127348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4323406517505646, + 0.10330101102590561, + 0.049711912870407104, + 0.051295798271894455, + 0.02936846949160099, + 0.03877304866909981, + 0.0608859658241272, + 0.13028621673583984, + 0.05751702934503555, + 0.04651994630694389 + ], + "credit_map": [ + 0.05719747394323349, + 0.08832896500825882, + 0.03557489067316055, + 0.06785047799348831, + 0.05154065042734146, + 0.034733209758996964, + 0.04172506928443909, + 0.055389877408742905, + 0.01409878209233284, + 0.023531829938292503, + 0.010852532461285591, + 0.014084559865295887, + 0.012049445882439613, + 0.008448109030723572, + 0.00907893106341362, + 0.011156809516251087, + 0.007596885319799185, + 0.010517262853682041, + 0.005196927580982447, + 0.006406860891729593, + 0.005710524041205645, + 0.003538677003234625, + 0.004877159371972084, + 0.005867613013833761, + 0.005790338385850191, + 0.008883196860551834, + 0.004381115082651377, + 0.010462423786520958, + 0.004477663431316614, + 0.0040925052016973495, + 0.006807662080973387, + 0.006400887854397297, + 0.004375445656478405, + 0.0034621567465364933, + 0.0028683661948889494, + 0.003772624535486102, + 0.004126079846173525, + 0.0028369594365358353, + 0.003134290687739849, + 0.004792544525116682, + 0.0051040733233094215, + 0.005331106949597597, + 0.0034038876183331013, + 0.00614320719614625, + 0.003458478255197406, + 0.0046569532714784145, + 0.004597974009811878, + 0.006077370140701532, + 0.007227245252579451, + 0.008426792919635773, + 0.00771315349265933, + 0.009670446626842022, + 0.008032461628317833, + 0.005074212793260813, + 0.007199892774224281, + 0.007541752886027098, + 0.0169540848582983, + 0.015028032474219799, + 0.019283412024378777, + 0.017118556424975395, + 0.012108792550861835, + 0.012377798557281494, + 0.013893121853470802, + 0.023522406816482544, + 0.007982717826962471, + 0.009255712851881981, + 0.005847332067787647, + 0.00871595460921526, + 0.006672470830380917, + 0.005832132883369923, + 0.005282827652990818, + 0.007927875965833664, + 0.0064407773315906525, + 0.006159983575344086, + 0.004489078186452389, + 0.010251609608530998, + 0.005034903064370155, + 0.004359355196356773, + 0.004489963874220848, + 0.005294273607432842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.003615036141127348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate moment of passionate conversation between two friends in a dimly lit, cozy bistro. A woman in her twenties, with wavy auburn hair loosely pulled back and wearing a soft cream-colored sweater, leans forward, a gentle smile on her face as she speaks. Across from her, a man of similar age, wearing a dark green button-down shirt with the sleeves casually rolled up to his forearms, listens intently, his head slightly tilted, a look of genuine engagement. Between them, on a small, round, dark wooden table, lie two elegant, large glasses half-filled with deep red wine. As the woman shares her story, her right hand gracefully and expressively gestures near the bowl of her glass, while the man's hand rests near the base of his own, his fingers leisurely tracing the polished wood. The entire scene is bathed in the warm golden light of a low-hanging lamp, casting soft shadows on their faces and causing the wine to gleam, creating a tranquil and welcoming atmosphere of easy friendship and shared confidence.", + "seed": 15304, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_5750_seed_15304_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": -0.02337231154087931, + "vlm_logprob": -0.02337231154087931, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.02337231154087931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07776597142219543 + ], + "opposite_logprob_tokens": [ + -4.202765941619873 + ], + "logprob_reward": -0.07776597142219543, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.07776597142219543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1177917942404747, + 0.07088517397642136, + 0.04679539054632187, + 0.0591658353805542, + 0.08608417212963104, + 0.058650970458984375, + 0.09953726083040237, + 0.24673911929130554, + 0.1087842658162117, + 0.10556599497795105 + ], + "credit_map": [ + 0.014789948239922523, + 0.018430931493639946, + 0.008284610696136951, + 0.013411984778940678, + 0.02053900994360447, + 0.014480069279670715, + 0.01003460306674242, + 0.01782064326107502, + 0.006181742995977402, + 0.007813215255737305, + 0.0045896898955106735, + 0.0074885571375489235, + 0.009396273642778397, + 0.00980285368859768, + 0.00954864826053381, + 0.016064193099737167, + 0.006516346242278814, + 0.005728121846914291, + 0.0035443997476249933, + 0.0043589272536337376, + 0.007314714137464762, + 0.0066480012610554695, + 0.004051536321640015, + 0.008633342571556568, + 0.0077678365632891655, + 0.006173884961754084, + 0.0029626397881656885, + 0.00770434457808733, + 0.010953951627016068, + 0.008136340416967869, + 0.006469288840889931, + 0.008997546508908272, + 0.009531359188258648, + 0.01100105233490467, + 0.007195372134447098, + 0.007760428357869387, + 0.016912519931793213, + 0.010728506371378899, + 0.010420284233987331, + 0.012534650973975658, + 0.006776975002139807, + 0.0068005952052772045, + 0.0036351410672068596, + 0.005919731222093105, + 0.009011635556817055, + 0.007438474800437689, + 0.01305883377790451, + 0.0060095773078501225, + 0.014207049272954464, + 0.007142334245145321, + 0.006787184160202742, + 0.005982388276606798, + 0.010754716582596302, + 0.01454073004424572, + 0.011271554976701736, + 0.028851307928562164, + 0.06549446284770966, + 0.011454087682068348, + 0.010439054109156132, + 0.01430966891348362, + 0.021448662504553795, + 0.07149985432624817, + 0.01909765973687172, + 0.03299567848443985, + 0.013102413155138493, + 0.013379231095314026, + 0.0073620653711259365, + 0.017256615683436394, + 0.014627350494265556, + 0.015068387612700462, + 0.015669064596295357, + 0.012319140136241913, + 0.015613446943461895, + 0.011928974650800228, + 0.009438645094633102, + 0.016154497861862183, + 0.010442566126585007, + 0.0158498864620924, + 0.015894737094640732, + 0.0102432481944561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.07776597142219543 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015397133538499475 + ], + "opposite_logprob_tokens": [ + -6.501539707183838 + ], + "logprob_reward": -0.0015397133538499475, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015397133538499475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16698117554187775, + 0.06762317568063736, + 0.06524639576673508, + 0.06499647349119186, + 0.07410680502653122, + 0.02966420352458954, + 0.07073132693767548, + 0.15461763739585876, + 0.15870681405067444, + 0.14732599258422852 + ], + "credit_map": [ + 0.019120080396533012, + 0.03332404047250748, + 0.010831277817487717, + 0.021086759865283966, + 0.025398114696145058, + 0.020818842574954033, + 0.016914641484618187, + 0.019487427547574043, + 0.005633956752717495, + 0.01157638244330883, + 0.005170187912881374, + 0.007621058728545904, + 0.010481756180524826, + 0.010211138054728508, + 0.00682453066110611, + 0.010104168206453323, + 0.006330291274935007, + 0.007017373573035002, + 0.004213779699057341, + 0.007891710847616196, + 0.012105388566851616, + 0.009600027464330196, + 0.00624191015958786, + 0.011845918372273445, + 0.008021438494324684, + 0.012346925213932991, + 0.004515601322054863, + 0.0071164462715387344, + 0.009729409590363503, + 0.00886994507163763, + 0.007509699556976557, + 0.006887007039040327, + 0.007553665898740292, + 0.007588658947497606, + 0.00459249084815383, + 0.005730732344090939, + 0.008427652530372143, + 0.010629041120409966, + 0.010210592299699783, + 0.019373971968889236, + 0.0035949761513620615, + 0.004256193060427904, + 0.002211860613897443, + 0.004072185140103102, + 0.004345559515058994, + 0.004364409018307924, + 0.0037499438039958477, + 0.0030690778512507677, + 0.008107257075607777, + 0.0073440964333713055, + 0.005501368548721075, + 0.007404097821563482, + 0.008234472945332527, + 0.009482896886765957, + 0.008900349959731102, + 0.015756793320178986, + 0.029513970017433167, + 0.01627204194664955, + 0.007710705511271954, + 0.010473656468093395, + 0.021695049479603767, + 0.03315439447760582, + 0.01566821150481701, + 0.020129604265093803, + 0.023242350667715073, + 0.016888123005628586, + 0.014911973848938942, + 0.01879502460360527, + 0.01870456337928772, + 0.024319089949131012, + 0.020515482872724533, + 0.021330203860998154, + 0.02218806743621826, + 0.019961154088377953, + 0.015414765104651451, + 0.02542116679251194, + 0.012910185381770134, + 0.018174804747104645, + 0.016988012939691544, + 0.01626782864332199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0015397133538499475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260909907519817 + ], + "opposite_logprob_tokens": [ + -5.255260944366455 + ], + "logprob_reward": -0.005260909907519817, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260909907519817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18282946944236755, + 0.10199669003486633, + 0.06359320133924484, + 0.09338835626840591, + 0.09716702252626419, + 0.0770951509475708, + 0.046859659254550934, + 0.0674930140376091, + 0.13204099237918854, + 0.13753639161586761 + ], + "credit_map": [ + 0.015493347309529781, + 0.02972368150949478, + 0.009917407296597958, + 0.018184758722782135, + 0.038815829902887344, + 0.024138949811458588, + 0.01796383410692215, + 0.02859167382121086, + 0.010177279822528362, + 0.0122876837849617, + 0.005945251323282719, + 0.008071644231677055, + 0.012248634360730648, + 0.013972468674182892, + 0.01589464396238327, + 0.02339908853173256, + 0.006376135628670454, + 0.006551013793796301, + 0.0034700362011790276, + 0.006540970876812935, + 0.011211362667381763, + 0.007928367704153061, + 0.0081007806584239, + 0.013414531014859676, + 0.012538844719529152, + 0.013835472986102104, + 0.005400781519711018, + 0.008966528810560703, + 0.015992121770977974, + 0.010774875991046429, + 0.010398440062999725, + 0.015481295995414257, + 0.014798318967223167, + 0.009506795555353165, + 0.0060486868023872375, + 0.00912624504417181, + 0.010227651335299015, + 0.01179369818419218, + 0.01225766446441412, + 0.0234079547226429, + 0.008262732066214085, + 0.006200353614985943, + 0.0030110152438282967, + 0.008023940026760101, + 0.011590927839279175, + 0.009378502145409584, + 0.018544290214776993, + 0.012083386071026325, + 0.004924757406115532, + 0.005853887181729078, + 0.0035384674556553364, + 0.0036988614592701197, + 0.00947553664445877, + 0.006239202339202166, + 0.004928271286189556, + 0.008200673386454582, + 0.007669163402169943, + 0.006624295841902494, + 0.005958680063486099, + 0.007923540659248829, + 0.008453389629721642, + 0.0118467528373003, + 0.008615391328930855, + 0.010401799343526363, + 0.018441664054989815, + 0.012612220831215382, + 0.012317148968577385, + 0.01965910941362381, + 0.011776674538850784, + 0.020429925993084908, + 0.016742274165153503, + 0.020061980932950974, + 0.020210998132824898, + 0.021915212273597717, + 0.020134976133704185, + 0.022732272744178772, + 0.011714211665093899, + 0.017027758061885834, + 0.011807763949036598, + 0.011993193067610264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.005260909907519817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348518207669258 + ], + "opposite_logprob_tokens": [ + -3.65134859085083 + ], + "logprob_reward": -0.026348518207669258, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348518207669258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11044149845838547, + 0.09816139191389084, + 0.034232404083013535, + 0.06885311007499695, + 0.040603235363960266, + 0.04584404081106186, + 0.1090303286910057, + 0.14799320697784424, + 0.2725198268890381, + 0.07232100516557693 + ], + "credit_map": [ + 0.00976085290312767, + 0.01249441783875227, + 0.00940930750221014, + 0.012115893885493279, + 0.017001168802380562, + 0.015049835667014122, + 0.01173231191933155, + 0.022877713665366173, + 0.007573141250759363, + 0.012274082750082016, + 0.005894103553146124, + 0.009764094837009907, + 0.011995231732726097, + 0.01438626367598772, + 0.014896093867719173, + 0.021378377452492714, + 0.0037918349262326956, + 0.003713908139616251, + 0.0017204551259055734, + 0.00361637189052999, + 0.005298068281263113, + 0.004686341620981693, + 0.0031242931727319956, + 0.008281130343675613, + 0.007452196441590786, + 0.007810004521161318, + 0.0034603490494191647, + 0.007246588822454214, + 0.01026931218802929, + 0.009266852401196957, + 0.007242159452289343, + 0.0161056499928236, + 0.004351869225502014, + 0.005930535029619932, + 0.0034128199331462383, + 0.004342255648225546, + 0.0073298681527376175, + 0.004726600833237171, + 0.0059424759820103645, + 0.0045668138191103935, + 0.004802068695425987, + 0.0034803631715476513, + 0.002189814345911145, + 0.003676085267215967, + 0.007501143030822277, + 0.006639810744673014, + 0.009903332218527794, + 0.007651423569768667, + 0.014100170694291592, + 0.012995296157896519, + 0.0071034119464457035, + 0.009733994491398335, + 0.01230844296514988, + 0.015778206288814545, + 0.015061152167618275, + 0.021949658170342445, + 0.023991087451577187, + 0.012591456063091755, + 0.008673061616718769, + 0.014219822362065315, + 0.01617085002362728, + 0.03138561174273491, + 0.018194647505879402, + 0.02276666648685932, + 0.03523975610733032, + 0.02411748841404915, + 0.026622869074344635, + 0.04715179651975632, + 0.02714330330491066, + 0.03707358241081238, + 0.041801925748586655, + 0.03336910158395767, + 0.0101504260674119, + 0.008596763014793396, + 0.006183084566146135, + 0.009531278163194656, + 0.007025704253464937, + 0.014532141387462616, + 0.006854325067251921, + 0.009447285905480385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.026348518207669258 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946444813162088 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946444813162088, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946444813162088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12974557280540466, + 0.05943085253238678, + 0.152065247297287, + 0.13866442441940308, + 0.17909151315689087, + 0.06829890608787537, + 0.042271196842193604, + 0.05366043373942375, + 0.06889656186103821, + 0.10787531733512878 + ], + "credit_map": [ + 0.01435022335499525, + 0.015634583309292793, + 0.011441548354923725, + 0.017435278743505478, + 0.021481599658727646, + 0.01790572516620159, + 0.013900372199714184, + 0.017596229910850525, + 0.006067332345992327, + 0.008648400194942951, + 0.00469892006367445, + 0.006883442401885986, + 0.006543820723891258, + 0.009059671312570572, + 0.006890646647661924, + 0.010638615116477013, + 0.025248613208532333, + 0.013054939918220043, + 0.010471564717590809, + 0.014587968587875366, + 0.02409295365214348, + 0.022821830585598946, + 0.016364866867661476, + 0.025422507897019386, + 0.015934549272060394, + 0.01991768553853035, + 0.008149324916303158, + 0.016719020903110504, + 0.022595226764678955, + 0.01996760256588459, + 0.01593358814716339, + 0.019447408616542816, + 0.01932080276310444, + 0.014272304251790047, + 0.010302146896719933, + 0.013737876899540424, + 0.02000044658780098, + 0.024877628311514854, + 0.030197247862815857, + 0.04638305678963661, + 0.0059408931992948055, + 0.005200983490794897, + 0.004942770581692457, + 0.006747054401785135, + 0.01129449624568224, + 0.00904683955013752, + 0.016540272161364555, + 0.00858558900654316, + 0.005260878708213568, + 0.00452817790210247, + 0.0035451874136924744, + 0.004029261413961649, + 0.0047785150818526745, + 0.0057866838760674, + 0.006190852727741003, + 0.008151634596288204, + 0.006882743909955025, + 0.004751427564769983, + 0.004174129106104374, + 0.005605035461485386, + 0.007077496033161879, + 0.008669471368193626, + 0.007201755419373512, + 0.009298373945057392, + 0.009547239169478416, + 0.0052953013218939304, + 0.004409097600728273, + 0.009393882937729359, + 0.006373928394168615, + 0.010316840372979641, + 0.01172943040728569, + 0.011830833740532398, + 0.013942614197731018, + 0.017710937187075615, + 0.011708452366292477, + 0.01464915368705988, + 0.009833330288529396, + 0.013232812285423279, + 0.01725068688392639, + 0.009547327645123005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.005946444813162088 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate moment of passionate conversation between two friends in a dimly lit, cozy bistro. A woman in her twenties, with wavy auburn hair loosely pulled back and wearing a soft cream-colored sweater, leans forward, a gentle smile on her face as she speaks. Across from her, a man of similar age, wearing a dark green button-down shirt with the sleeves casually rolled up to his forearms, listens intently, his head slightly tilted, a look of genuine engagement. Between them, on a small, round, dark wooden table, lie two elegant, large glasses half-filled with deep red wine. As the woman shares her story, her right hand gracefully and expressively gestures near the bowl of her glass, while the man's hand rests near the base of his own, his fingers leisurely tracing the polished wood. The entire scene is bathed in the warm golden light of a low-hanging lamp, casting soft shadows on their faces and causing the wine to gleam, creating a tranquil and welcoming atmosphere of easy friendship and shared confidence.", + "seed": 15305, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_5750_seed_15305_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": -0.026752868574112652, + "vlm_logprob": -0.026752868574112652, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.026752868574112652, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07464275509119034 + ], + "opposite_logprob_tokens": [ + -3.6996426582336426 + ], + "logprob_reward": -0.07464275509119034, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.07464275509119034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23282627761363983, + 0.16080772876739502, + 0.10403461009263992, + 0.08212008327245712, + 0.045827362686395645, + 0.11023228615522385, + 0.060769543051719666, + 0.03364283964037895, + 0.08083885163068771, + 0.08890042454004288 + ], + "credit_map": [ + 0.03275621682405472, + 0.026549670845270157, + 0.028879771009087563, + 0.024842821061611176, + 0.0275407861918211, + 0.02733752876520157, + 0.028474105522036552, + 0.03644535690546036, + 0.016445288434624672, + 0.02125524915754795, + 0.01685492880642414, + 0.013935689814388752, + 0.017762314528226852, + 0.02187521755695343, + 0.019009487703442574, + 0.033669546246528625, + 0.012390309944748878, + 0.01422849204391241, + 0.012610246427357197, + 0.013461390510201454, + 0.013128125108778477, + 0.01160262431949377, + 0.011994140222668648, + 0.014619281515479088, + 0.012468787841498852, + 0.01180370431393385, + 0.00883641466498375, + 0.00935181975364685, + 0.007482181303203106, + 0.009227443486452103, + 0.009355384856462479, + 0.013594349846243858, + 0.005689910613000393, + 0.004778075963258743, + 0.004841851070523262, + 0.0059511070139706135, + 0.006600824650377035, + 0.005421809386461973, + 0.005860263481736183, + 0.006683517247438431, + 0.011560356244444847, + 0.012527905404567719, + 0.010386338457465172, + 0.020323732867836952, + 0.011103590950369835, + 0.012686610221862793, + 0.01453899871557951, + 0.01710475981235504, + 0.006662487983703613, + 0.006624053698033094, + 0.006483652628958225, + 0.005833533126860857, + 0.008151820860803127, + 0.006141064688563347, + 0.011614738032221794, + 0.00925819668918848, + 0.003960431087762117, + 0.004543947987258434, + 0.004223550669848919, + 0.004043783061206341, + 0.0043471092358231544, + 0.003985019866377115, + 0.0038588419556617737, + 0.004680156707763672, + 0.009049802087247372, + 0.009905077517032623, + 0.009331396780908108, + 0.008849790319800377, + 0.013421092182397842, + 0.008483190089464188, + 0.01199747808277607, + 0.009801022708415985, + 0.010113521479070187, + 0.00977141410112381, + 0.01089052576571703, + 0.013987964950501919, + 0.010097917169332504, + 0.010695982724428177, + 0.012476646341383457, + 0.01086644921451807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07464275509119034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059780846349895 + ], + "opposite_logprob_tokens": [ + -5.130978107452393 + ], + "logprob_reward": -0.0059780846349895, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059780846349895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4095690846443176, + 0.26002880930900574, + 0.04656348377466202, + 0.058705393224954605, + 0.07624606043100357, + 0.026656778529286385, + 0.022256089374423027, + 0.02294214256107807, + 0.031743355095386505, + 0.045288801193237305 + ], + "credit_map": [ + 0.054613374173641205, + 0.046384166926145554, + 0.047853343188762665, + 0.03353894501924515, + 0.061092156916856766, + 0.048221662640571594, + 0.07678370922803879, + 0.04108172282576561, + 0.029949761927127838, + 0.03666942939162254, + 0.034242287278175354, + 0.025967242196202278, + 0.03552226722240448, + 0.030689064413309097, + 0.04297920688986778, + 0.02400955557823181, + 0.005871553905308247, + 0.006335533689707518, + 0.005628291983157396, + 0.004859504289925098, + 0.006698116194456816, + 0.005372466519474983, + 0.006808442063629627, + 0.0049895732663571835, + 0.007022188976407051, + 0.008266384713351727, + 0.009029723703861237, + 0.0056457361206412315, + 0.009938557632267475, + 0.005597394425421953, + 0.007177131716161966, + 0.006028278265148401, + 0.007991171441972256, + 0.009043294936418533, + 0.011464701034128666, + 0.008266860619187355, + 0.013224184513092041, + 0.008232352323830128, + 0.009145616553723812, + 0.00887787900865078, + 0.002727290615439415, + 0.0036628697998821735, + 0.004119557328522205, + 0.0054617756977677345, + 0.0025808964855968952, + 0.002982355421409011, + 0.0023700064048171043, + 0.0027520288713276386, + 0.0032562946435064077, + 0.0024298299103975296, + 0.00237436150200665, + 0.0020937398076057434, + 0.003066397039219737, + 0.002586620394140482, + 0.0032108791638165712, + 0.003237965516746044, + 0.0026028354186564684, + 0.003385829972103238, + 0.0027664387598633766, + 0.0028137564659118652, + 0.00397081021219492, + 0.0026381907518953085, + 0.0026246639899909496, + 0.002139616059139371, + 0.0033042272552847862, + 0.004447766579687595, + 0.0035666103940457106, + 0.00350617291405797, + 0.004415193106979132, + 0.0034021695610135794, + 0.003787808120250702, + 0.00531340716406703, + 0.004907097667455673, + 0.004973751027137041, + 0.0054551963694393635, + 0.004472440108656883, + 0.006930035073310137, + 0.005424941424280405, + 0.00870845653116703, + 0.004416882060468197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0059780846349895 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788726456463337 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788726456463337, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788726456463337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17213751375675201, + 0.1108035147190094, + 0.05275629833340645, + 0.12206093966960907, + 0.11435245722532272, + 0.21080519258975983, + 0.05577901378273964, + 0.03188814967870712, + 0.08786383271217346, + 0.0415530689060688 + ], + "credit_map": [ + 0.020316507667303085, + 0.017869800329208374, + 0.017729172483086586, + 0.031081708148121834, + 0.013688554055988789, + 0.02094154991209507, + 0.027891777455806732, + 0.022618455812335014, + 0.012646964751183987, + 0.013446057215332985, + 0.015531317330896854, + 0.009863248094916344, + 0.013062077574431896, + 0.013494901359081268, + 0.018777182325720787, + 0.013981766067445278, + 0.006066198460757732, + 0.006215632893145084, + 0.004120951984077692, + 0.004720534663647413, + 0.00516150938346982, + 0.0067172180861234665, + 0.009764990769326687, + 0.00998926442116499, + 0.01785876601934433, + 0.016556162387132645, + 0.01580003648996353, + 0.01203614566475153, + 0.019849151372909546, + 0.011529124341905117, + 0.015624345280230045, + 0.012807212769985199, + 0.012570839375257492, + 0.01507786475121975, + 0.015493297018110752, + 0.022679224610328674, + 0.0133344242349267, + 0.01279627624899149, + 0.011358784511685371, + 0.011041748337447643, + 0.027056055143475533, + 0.02335260435938835, + 0.030086718499660492, + 0.030036594718694687, + 0.027753986418247223, + 0.02386038564145565, + 0.02175435796380043, + 0.026904486119747162, + 0.0087131317704916, + 0.008280616253614426, + 0.005763179622590542, + 0.005705994088202715, + 0.00846357923001051, + 0.006066080182790756, + 0.006226574070751667, + 0.006559861823916435, + 0.003981371875852346, + 0.0039335633628070354, + 0.003969373647123575, + 0.00380972889252007, + 0.005495127756148577, + 0.003391019767150283, + 0.004556185565888882, + 0.002751777647063136, + 0.010462436825037003, + 0.012972214259207249, + 0.009917681105434895, + 0.009929394349455833, + 0.01447882130742073, + 0.00870126485824585, + 0.01076613087207079, + 0.01063589472323656, + 0.004745105747133493, + 0.004446381703019142, + 0.0060262843035161495, + 0.004200386814773083, + 0.005950589664280415, + 0.005113921593874693, + 0.006263833958655596, + 0.004806565120816231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.009788726456463337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104183971881866 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.038104183971881866, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104183971881866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25991764664649963, + 0.06427991390228271, + 0.028135664761066437, + 0.07515110820531845, + 0.12060092389583588, + 0.2164076864719391, + 0.04711013287305832, + 0.06950195133686066, + 0.09791329503059387, + 0.020981645211577415 + ], + "credit_map": [ + 0.03605397045612335, + 0.02500775270164013, + 0.03042576089501381, + 0.023608269169926643, + 0.0295370165258646, + 0.030110912397503853, + 0.044363923370838165, + 0.04081004485487938, + 0.007436876650899649, + 0.00872844085097313, + 0.008353152312338352, + 0.006225927732884884, + 0.008789518848061562, + 0.007700606714934111, + 0.011508269235491753, + 0.005537115968763828, + 0.003943494521081448, + 0.003215912263840437, + 0.0019657218363136053, + 0.002938701305538416, + 0.003349863225594163, + 0.0031277299858629704, + 0.004996597301214933, + 0.004597645252943039, + 0.011319957673549652, + 0.009637520648539066, + 0.009456419385969639, + 0.007512426469475031, + 0.012715945020318031, + 0.007423809729516506, + 0.007855805568397045, + 0.00922922883182764, + 0.01417500525712967, + 0.012829629704356194, + 0.021491317078471184, + 0.01808568648993969, + 0.013477656058967113, + 0.012112511321902275, + 0.01374432910233736, + 0.01468479074537754, + 0.02706017717719078, + 0.02602582797408104, + 0.03167502582073212, + 0.035152871161699295, + 0.02742300182580948, + 0.020719408988952637, + 0.02382349595427513, + 0.024527860805392265, + 0.0050085256807506084, + 0.005114123225212097, + 0.003939139191061258, + 0.005991945043206215, + 0.008285820484161377, + 0.005268174689263105, + 0.006411423906683922, + 0.007090980187058449, + 0.008736309595406055, + 0.00923903938382864, + 0.0069815474562346935, + 0.005686515010893345, + 0.009150692261755466, + 0.008438517339527607, + 0.010213199071586132, + 0.011056121438741684, + 0.010577831417322159, + 0.009588693268597126, + 0.012384505942463875, + 0.011494752950966358, + 0.013792670331895351, + 0.01109506469219923, + 0.014317603781819344, + 0.014662171714007854, + 0.0031325591262429953, + 0.0030410217586904764, + 0.0023271830286830664, + 0.002572843339294195, + 0.0023078827653080225, + 0.002357759280130267, + 0.002967849373817444, + 0.0022745467722415924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.038104183971881866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250592716038227 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250592716038227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250592716038227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3190650939941406, + 0.12848374247550964, + 0.061216242611408234, + 0.04979930818080902, + 0.0680299699306488, + 0.043963633477687836, + 0.08999831974506378, + 0.06149053946137428, + 0.1144661083817482, + 0.06348702311515808 + ], + "credit_map": [ + 0.04044996574521065, + 0.036682140082120895, + 0.034436821937561035, + 0.03202023729681969, + 0.030821586027741432, + 0.03972264379262924, + 0.04776344075798988, + 0.05716826394200325, + 0.01708361506462097, + 0.019919447600841522, + 0.01339825801551342, + 0.015294497832655907, + 0.013328670524060726, + 0.015886452049016953, + 0.02025269716978073, + 0.013320094905793667, + 0.008623183704912663, + 0.008345488458871841, + 0.0058801909908652306, + 0.006344188004732132, + 0.010165318846702576, + 0.00700869457796216, + 0.008142282254993916, + 0.006706892512738705, + 0.0061922818422317505, + 0.007471420802175999, + 0.005187726579606533, + 0.006487155798822641, + 0.008261754177510738, + 0.005299183540046215, + 0.005937755573540926, + 0.004962030332535505, + 0.008425261825323105, + 0.007637812290340662, + 0.010459796525537968, + 0.00826148223131895, + 0.009524109773337841, + 0.00735081871971488, + 0.00775174330919981, + 0.008618946187198162, + 0.0051110913045704365, + 0.005455558188259602, + 0.0036059918347746134, + 0.006619191262871027, + 0.005159095395356417, + 0.0048685092478990555, + 0.006961312144994736, + 0.006182883866131306, + 0.011867697350680828, + 0.010997938923537731, + 0.011452516540884972, + 0.009720779955387115, + 0.013588850386440754, + 0.010229956358671188, + 0.014097730629146099, + 0.008042847737669945, + 0.008238770067691803, + 0.00796548929065466, + 0.009702012874186039, + 0.007577213924378157, + 0.006617730017751455, + 0.007485788315534592, + 0.005500528495758772, + 0.00840300228446722, + 0.013022336177527905, + 0.017760299146175385, + 0.012957611121237278, + 0.014042199589312077, + 0.016387123614549637, + 0.011496547609567642, + 0.01337437890470028, + 0.01542561873793602, + 0.00829892698675394, + 0.008660119958221912, + 0.008596565574407578, + 0.009714496321976185, + 0.007893456146121025, + 0.007657669018954039, + 0.007148996461182833, + 0.005516787059605122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005250592716038227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures an intimate moment of passionate conversation between two friends in a dimly lit, cozy bistro. A woman in her twenties, with wavy auburn hair loosely pulled back and wearing a soft cream-colored sweater, leans forward, a gentle smile on her face as she speaks. Across from her, a man of similar age, wearing a dark green button-down shirt with the sleeves casually rolled up to his forearms, listens intently, his head slightly tilted, a look of genuine engagement. Between them, on a small, round, dark wooden table, lie two elegant, large glasses half-filled with deep red wine. As the woman shares her story, her right hand gracefully and expressively gestures near the bowl of her glass, while the man's hand rests near the base of his own, his fingers leisurely tracing the polished wood. The entire scene is bathed in the warm golden light of a low-hanging lamp, casting soft shadows on their faces and causing the wine to gleam, creating a tranquil and welcoming atmosphere of easy friendship and shared confidence.", + "seed": 15306, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_8/index_5750_seed_15306_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": -0.022708877827972172, + "vlm_logprob": -0.022708877827972172, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.022708877827972172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.058944184333086014 + ], + "opposite_logprob_tokens": [ + -4.183944225311279 + ], + "logprob_reward": -0.058944184333086014, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.058944184333086014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28314194083213806, + 0.15942804515361786, + 0.0702567845582962, + 0.021655065938830376, + 0.0957045778632164, + 0.03861670196056366, + 0.10719660669565201, + 0.05070706829428673, + 0.12223304063081741, + 0.05106018856167793 + ], + "credit_map": [ + 0.03636495769023895, + 0.027027107775211334, + 0.06343508511781693, + 0.019319582730531693, + 0.034134287387132645, + 0.026288533583283424, + 0.030067436397075653, + 0.04650494456291199, + 0.019647900015115738, + 0.007286299020051956, + 0.049093279987573624, + 0.010046835988759995, + 0.008432005532085896, + 0.01392898429185152, + 0.010068639181554317, + 0.04092409834265709, + 0.007410213351249695, + 0.004895031452178955, + 0.017818398773670197, + 0.00613817386329174, + 0.008374905213713646, + 0.00519496388733387, + 0.006754751782864332, + 0.013670345768332481, + 0.0021803106646984816, + 0.002038436010479927, + 0.0027233753353357315, + 0.0022635944187641144, + 0.0035533516202121973, + 0.0023366683162748814, + 0.004266264382749796, + 0.0022930640261620283, + 0.012949185445904732, + 0.01061099860817194, + 0.015460481867194176, + 0.008047858253121376, + 0.01521657407283783, + 0.008194883354008198, + 0.011267866007983685, + 0.013956736773252487, + 0.005624304059892893, + 0.005291660316288471, + 0.007846307009458542, + 0.0037986382376402617, + 0.004773908294737339, + 0.003310384461656213, + 0.0036252124700695276, + 0.004346287809312344, + 0.016336288303136826, + 0.00941302441060543, + 0.015849340707063675, + 0.010386206209659576, + 0.017663802951574326, + 0.011591866612434387, + 0.013428485952317715, + 0.012527598068118095, + 0.006433303467929363, + 0.004349429626017809, + 0.006301911547780037, + 0.005534073803573847, + 0.006610045675188303, + 0.006187694612890482, + 0.006505907513201237, + 0.008784704841673374, + 0.012361399829387665, + 0.00878870952874422, + 0.019008366391062737, + 0.009903348051011562, + 0.027946827933192253, + 0.010238861665129662, + 0.022087080404162407, + 0.011898454278707504, + 0.010217367671430111, + 0.006326996721327305, + 0.005714534781873226, + 0.004250014200806618, + 0.008794176392257214, + 0.004924096167087555, + 0.005211999174207449, + 0.00562100438401103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.058944184333086014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019656161312013865 + ], + "opposite_logprob_tokens": [ + -6.251965522766113 + ], + "logprob_reward": -0.0019656161312013865, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019656161312013865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24152220785617828, + 0.15347245335578918, + 0.02989852800965309, + 0.04341527074575424, + 0.10051742196083069, + 0.08196307718753815, + 0.14650419354438782, + 0.0533197745680809, + 0.06105252727866173, + 0.0883345752954483 + ], + "credit_map": [ + 0.025964686647057533, + 0.01306818425655365, + 0.06706032156944275, + 0.019898423925042152, + 0.022176586091518402, + 0.018460582941770554, + 0.022540835663676262, + 0.05235258862376213, + 0.017240623012185097, + 0.008142126724123955, + 0.043592292815446854, + 0.012866989709436893, + 0.00885283388197422, + 0.015361606143414974, + 0.010467980988323689, + 0.03694798797369003, + 0.003242927836254239, + 0.0036186284851282835, + 0.004278677515685558, + 0.0046733682975173, + 0.0052398680709302425, + 0.0033276870381087065, + 0.003225635038688779, + 0.0022917354945093393, + 0.004830512683838606, + 0.0024582238402217627, + 0.010082556866109371, + 0.004062206484377384, + 0.005117441061884165, + 0.004102788865566254, + 0.003765803761780262, + 0.00899573415517807, + 0.01341230608522892, + 0.012564467266201973, + 0.013543243519961834, + 0.00869552232325077, + 0.01425358559936285, + 0.008517980575561523, + 0.016039133071899414, + 0.013491189107298851, + 0.010038279928267002, + 0.008797296322882175, + 0.01870156079530716, + 0.007236080709844828, + 0.009113146923482418, + 0.00822583120316267, + 0.006796733010560274, + 0.013054152950644493, + 0.018167253583669662, + 0.013784383423626423, + 0.026488229632377625, + 0.015525003895163536, + 0.02285870350897312, + 0.01608957350254059, + 0.01738988421857357, + 0.016201162710785866, + 0.007042488548904657, + 0.004411501809954643, + 0.005494439974427223, + 0.005283968057483435, + 0.010503646917641163, + 0.0053487252444028854, + 0.009662380442023277, + 0.00557262497022748, + 0.008336079306900501, + 0.005307000130414963, + 0.007143104914575815, + 0.004977203439921141, + 0.012796525843441486, + 0.005756773054599762, + 0.011277901940047741, + 0.0054579367861151695, + 0.016128450632095337, + 0.008190793916583061, + 0.010278585366904736, + 0.006842435337603092, + 0.015051757916808128, + 0.00853223167359829, + 0.012454637326300144, + 0.010855686850845814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0019656161312013865 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956635810434818 + ], + "opposite_logprob_tokens": [ + -5.130956649780273 + ], + "logprob_reward": -0.005956635810434818, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956635810434818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13200771808624268, + 0.06827585399150848, + 0.039676424115896225, + 0.03998588025569916, + 0.4248828589916229, + 0.02604515850543976, + 0.0678863525390625, + 0.04243597015738487, + 0.05757060647010803, + 0.10123323649168015 + ], + "credit_map": [ + 0.014195438474416733, + 0.010712672956287861, + 0.01650766283273697, + 0.020237619057297707, + 0.018931889906525612, + 0.012219185940921307, + 0.026231668889522552, + 0.012971585616469383, + 0.00821240246295929, + 0.005436692386865616, + 0.009955629706382751, + 0.01017847377806902, + 0.012096927501261234, + 0.005808096844702959, + 0.010626972652971745, + 0.005960654933005571, + 0.00522534130141139, + 0.003775788238272071, + 0.005567335523664951, + 0.008216370828449726, + 0.004799363203346729, + 0.003343424992635846, + 0.004706470295786858, + 0.004042328335344791, + 0.004963480867445469, + 0.0038660780992358923, + 0.005385132040828466, + 0.004384480882436037, + 0.00628800131380558, + 0.005382454488426447, + 0.0066057476215064526, + 0.0031105035450309515, + 0.07155188918113708, + 0.07521816343069077, + 0.05790085718035698, + 0.028809471055865288, + 0.06947048753499985, + 0.03417982533574104, + 0.04339614883065224, + 0.04435601457953453, + 0.00302031054161489, + 0.002213010098785162, + 0.005101694259792566, + 0.0028708309400826693, + 0.0028315463569015265, + 0.0022681113332509995, + 0.0032802424393594265, + 0.004459413234144449, + 0.009848587214946747, + 0.008412363938987255, + 0.011290640570223331, + 0.007624203339219093, + 0.007656729780137539, + 0.007495962083339691, + 0.007783847860991955, + 0.007774014491587877, + 0.005512251053005457, + 0.0033370296005159616, + 0.005187699105590582, + 0.007476214785128832, + 0.0052991448901593685, + 0.0039240093901753426, + 0.006860104855149984, + 0.004839515313506126, + 0.006560970097780228, + 0.004788022954016924, + 0.007725521922111511, + 0.004905101377516985, + 0.012502072378993034, + 0.005504119209945202, + 0.010389333590865135, + 0.005195460747927427, + 0.01683313213288784, + 0.009302284568548203, + 0.01240982674062252, + 0.006737477146089077, + 0.015573524869978428, + 0.009148439392447472, + 0.01782800257205963, + 0.013400549069046974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.005956635810434818 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063629418611526 + ], + "opposite_logprob_tokens": [ + -3.1680636405944824 + ], + "logprob_reward": -0.043063629418611526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063629418611526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38952434062957764, + 0.12751156091690063, + 0.04401562735438347, + 0.05033740773797035, + 0.13478325307369232, + 0.04946344718337059, + 0.07309983670711517, + 0.03751622512936592, + 0.052577219903469086, + 0.04117107391357422 + ], + "credit_map": [ + 0.04069929197430611, + 0.014392590150237083, + 0.09926562011241913, + 0.027445493265986443, + 0.047971006482839584, + 0.03114914335310459, + 0.04569443315267563, + 0.08290677517652512, + 0.013855485245585442, + 0.005960669368505478, + 0.03487072512507439, + 0.008891572244465351, + 0.011363727040588856, + 0.009903649799525738, + 0.013649778440594673, + 0.02901596762239933, + 0.004392197355628014, + 0.0031160179059952497, + 0.006411245092749596, + 0.004600337706506252, + 0.007690968457609415, + 0.004344992805272341, + 0.008443030528724194, + 0.005016845650970936, + 0.005986163392663002, + 0.0044244034215807915, + 0.006120341829955578, + 0.004080884624272585, + 0.010111438110470772, + 0.00662165367975831, + 0.008820576593279839, + 0.004171950276941061, + 0.01866976171731949, + 0.016511360183358192, + 0.021258099004626274, + 0.01055710855871439, + 0.021555757150053978, + 0.010239478200674057, + 0.01977049559354782, + 0.0162211861461401, + 0.005985238589346409, + 0.004524781834334135, + 0.010678842663764954, + 0.004754412453621626, + 0.0053536556661129, + 0.004794752225279808, + 0.004912420175969601, + 0.008459345437586308, + 0.00971526000648737, + 0.005373873747885227, + 0.011281036771833897, + 0.00846889428794384, + 0.013103511184453964, + 0.007082771044224501, + 0.010771799832582474, + 0.007302696816623211, + 0.004407149273902178, + 0.0035451024305075407, + 0.0044721439480781555, + 0.004006636328995228, + 0.006083718966692686, + 0.004596584476530552, + 0.006173085421323776, + 0.0042318059131503105, + 0.006493725348263979, + 0.004207128658890724, + 0.01154288463294506, + 0.005997553933411837, + 0.006694621406495571, + 0.0047913240268826485, + 0.005540589336305857, + 0.007309395354241133, + 0.00409448379650712, + 0.003797536715865135, + 0.004822119604796171, + 0.005501923151314259, + 0.006033657118678093, + 0.005214530974626541, + 0.006863059941679239, + 0.004843767732381821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.043063629418611526 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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, eye-level medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are two friends, a woman with wavy auburn hair in a cream sweater and a man in a dark green button-down shirt, visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a small, round, dark wooden table with two large glasses half-filled with deep red wine between them?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman gesturing with her right hand near her glass while the man's hand rests on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden light from a low-hanging lamp casting soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036143234465271235 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036143234465271235, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036143234465271235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4508169889450073, + 0.13216720521450043, + 0.06520361453294754, + 0.03940301388502121, + 0.0650191381573677, + 0.09708204865455627, + 0.04885082319378853, + 0.04947880282998085, + 0.024727962911128998, + 0.02725040167570114 + ], + "credit_map": [ + 0.050714440643787384, + 0.030457928776741028, + 0.12795034050941467, + 0.04832710325717926, + 0.030273104086518288, + 0.03327320143580437, + 0.02535131759941578, + 0.10446957498788834, + 0.01536624040454626, + 0.00685834838077426, + 0.042271148413419724, + 0.010336917825043201, + 0.007677160669118166, + 0.008010965771973133, + 0.007041933014988899, + 0.03460448607802391, + 0.008482671342790127, + 0.005409335717558861, + 0.01751171052455902, + 0.006326364818960428, + 0.0047112759202718735, + 0.004406529478728771, + 0.004255322739481926, + 0.014100397936999798, + 0.0049368757754564285, + 0.004518652334809303, + 0.005387874785810709, + 0.0037375343963503838, + 0.00665940111503005, + 0.004844130482524633, + 0.005066105164587498, + 0.004252438433468342, + 0.009012145921587944, + 0.008579054847359657, + 0.00830936897546053, + 0.005963663104921579, + 0.009321726858615875, + 0.007454287260770798, + 0.008731488138437271, + 0.007647404447197914, + 0.011318235658109188, + 0.011085700243711472, + 0.021115371957421303, + 0.011992008425295353, + 0.00924491137266159, + 0.008126816712319851, + 0.008379599079489708, + 0.015819402411580086, + 0.005509631242603064, + 0.006807337515056133, + 0.005453732796013355, + 0.003659907728433609, + 0.011969132348895073, + 0.004928892012685537, + 0.005138082429766655, + 0.005384103860706091, + 0.005187053233385086, + 0.004499437287449837, + 0.0049840351566672325, + 0.0056817661970853806, + 0.00918105524033308, + 0.005677507258951664, + 0.008809828199446201, + 0.005458117928355932, + 0.0036714652087539434, + 0.002509199781343341, + 0.005292566493153572, + 0.0026285715866833925, + 0.0033047408796846867, + 0.002022691536694765, + 0.002471132669597864, + 0.0028275942895561457, + 0.003434550715610385, + 0.002794732805341482, + 0.004351482726633549, + 0.002303537679836154, + 0.0037345988675951958, + 0.0034562929067760706, + 0.0031075188890099525, + 0.004067684523761272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0036143234465271235 + } + ], + "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/0000158/group_9/index_tensor([5325, 5325])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_tensor([5325, 5325])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..99e5e1eab4cf2b1268c1855584707d628ab08967 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_tensor([5325, 5325])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a surreal 3D rendering, a low-angle medium shot frames a powerful construction warrior, his body huddled in a low crouch on a chaotic, smoke-filled battlefield at dusk. He wears worn, black leather armor over chain mail, with massive steel pauldrons protecting his broad shoulders. His face, with its chiseled features, square jaw, and resolute brows, is a mask of fierce loyalty, his eyes fixed on an unseen enemy. His hands, clasped in leather gloves, grip the shafts of a pair of halberds, knuckles white. Suddenly, the earth beneath him erupts violently, propelling him to the surface like an underground explosion. As the warrior is propelled into the air, the camera executes a rapid, tight zoom, focusing in realistic detail. This rapid movement captures the instantaneous shift in his expression from steely determination to pure shock, his eyes widening with pupils dilated. The zoom highlights the powerful ripples of his biceps and deltoids straining beneath his skin in response to the unexpected force, while a spray of dust and pebbles moves in slow-motion shockwaves around his rising body. The entire scene is bathed in the low, warm, orange glow of the setting sun, which glints off the armored edges of the sun, illuminating the swirling particles of smoke and debris, creating an atmosphere of tense and dramatic violence.", + "seed": 9445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_5325_seed_9445_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": -0.031718381028622386, + "vlm_logprob": -0.031718381028622386, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.031718381028622386, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06481971591711044 + ], + "opposite_logprob_tokens": [ + -2.814819812774658 + ], + "logprob_reward": -0.06481971591711044, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06481971591711044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1376764476299286, + 0.14288893342018127, + 0.09196699410676956, + 0.05527740344405174, + 0.015034410171210766, + 0.023707980290055275, + 0.138450488448143, + 0.15438827872276306, + 0.15462368726730347, + 0.08598531037569046 + ], + "credit_map": [ + 0.020799361169338226, + 0.01487007737159729, + 0.020825820043683052, + 0.01645226776599884, + 0.017740752547979355, + 0.01736740581691265, + 0.014655562117695808, + 0.014965218491852283, + 0.03150321915745735, + 0.014848587103188038, + 0.02002992294728756, + 0.026351483538746834, + 0.011230211704969406, + 0.012703168205916882, + 0.012954418547451496, + 0.01326792687177658, + 0.012493323534727097, + 0.008966629393398762, + 0.012576968409121037, + 0.009663726203143597, + 0.010132996365427971, + 0.015545089729130268, + 0.011960807256400585, + 0.01062745414674282, + 0.004694222938269377, + 0.006162039469927549, + 0.005435499828308821, + 0.0056169298477470875, + 0.01454126089811325, + 0.006207735277712345, + 0.005449066869914532, + 0.007170653436332941, + 0.0014236365677788854, + 0.0021973545663058758, + 0.001474150805734098, + 0.001699797692708671, + 0.0025763523299247026, + 0.0018081492744386196, + 0.001841049175709486, + 0.002013920107856393, + 0.0020257881842553616, + 0.0028147860430181026, + 0.0026304791681468487, + 0.004021819215267897, + 0.0023124790750443935, + 0.003026094986125827, + 0.0037073788698762655, + 0.003169155679643154, + 0.011580857448279858, + 0.018152400851249695, + 0.01349722035229206, + 0.012963905930519104, + 0.015816347673535347, + 0.025081230327486992, + 0.020525366067886353, + 0.02083316072821617, + 0.00749116251245141, + 0.005341666750609875, + 0.004087052308022976, + 0.005856535863131285, + 0.005969350226223469, + 0.052271075546741486, + 0.009706840850412846, + 0.06366460025310516, + 0.015500827692449093, + 0.021003825590014458, + 0.02030104026198387, + 0.01837240904569626, + 0.016146676614880562, + 0.028375744819641113, + 0.017184069380164146, + 0.017739107832312584, + 0.008851096965372562, + 0.012062598019838333, + 0.005856695119291544, + 0.011981021612882614, + 0.006824252661317587, + 0.012146410532295704, + 0.010758700780570507, + 0.01750454492866993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.06481971591711044 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.18949519097805023, + 0.06975176930427551, + 0.07569398730993271, + 0.026804521679878235, + 0.02921542525291443, + 0.02480863779783249, + 0.1409691423177719, + 0.2908027470111847, + 0.10122775286436081, + 0.0512307733297348 + ], + "credit_map": [ + 0.025937944650650024, + 0.021276354789733887, + 0.02239334210753441, + 0.024848952889442444, + 0.035491298884153366, + 0.021101653575897217, + 0.021486246958374977, + 0.016959430649876595, + 0.011028528213500977, + 0.009863562881946564, + 0.011461210437119007, + 0.009462960064411163, + 0.007750383112579584, + 0.0065405904315412045, + 0.007043278310447931, + 0.006601265165954828, + 0.007000182755291462, + 0.00886398646980524, + 0.011071224696934223, + 0.011911453679203987, + 0.009156190790235996, + 0.00951708760112524, + 0.009903816506266594, + 0.008270054124295712, + 0.002348752226680517, + 0.0024844252038747072, + 0.0020425445400178432, + 0.002259877510368824, + 0.004403479862958193, + 0.003566477680578828, + 0.0030570372473448515, + 0.00664193369448185, + 0.0026107500307261944, + 0.002895274432376027, + 0.005629457533359528, + 0.002842901973053813, + 0.003497759811580181, + 0.003446483751758933, + 0.003492234507575631, + 0.0048005664721131325, + 0.003340364433825016, + 0.002756834262982011, + 0.003371437080204487, + 0.0034917141310870647, + 0.002306690439581871, + 0.002601690124720335, + 0.0034519394394010305, + 0.0034879709128290415, + 0.011984269134700298, + 0.014373349025845528, + 0.01198552455753088, + 0.018662741407752037, + 0.016214480623602867, + 0.01734081096947193, + 0.029315924271941185, + 0.02109205536544323, + 0.015921849757432938, + 0.011528801172971725, + 0.009725399315357208, + 0.012024007737636566, + 0.013663495890796185, + 0.0890892744064331, + 0.021867869421839714, + 0.11698206514120102, + 0.011419513262808323, + 0.013812797144055367, + 0.012436647899448872, + 0.0126248924061656, + 0.010665621608495712, + 0.0159445833414793, + 0.013008476234972477, + 0.011315234936773777, + 0.005338088143616915, + 0.006920990068465471, + 0.011256009340286255, + 0.0068302880972623825, + 0.0039596352726221085, + 0.005417817737907171, + 0.005232457537204027, + 0.00627549085766077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.01108264084905386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738320458680391 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738320458680391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738320458680391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10914326459169388, + 0.02643703855574131, + 0.04832613468170166, + 0.02965371683239937, + 0.04054708778858185, + 0.025799056515097618, + 0.19668643176555634, + 0.04869350418448448, + 0.43117132782936096, + 0.043542396277189255 + ], + "credit_map": [ + 0.017239637672901154, + 0.018021324649453163, + 0.012576155364513397, + 0.010516884736716747, + 0.013988485559821129, + 0.013290710747241974, + 0.01013669092208147, + 0.013373380526900291, + 0.0030429090838879347, + 0.0033603431656956673, + 0.006124669220298529, + 0.003933507949113846, + 0.0024423690047115088, + 0.002457209164276719, + 0.0021258050110191107, + 0.00295022944919765, + 0.0043549202382564545, + 0.00711221294477582, + 0.006522598676383495, + 0.007648550905287266, + 0.0067674159072339535, + 0.004694752395153046, + 0.006472950801253319, + 0.004752737935632467, + 0.0021133513655513525, + 0.0031561790965497494, + 0.0029440177604556084, + 0.00288222823292017, + 0.00793319009244442, + 0.0030760294757783413, + 0.003498808480799198, + 0.004049915354698896, + 0.0034697663504630327, + 0.0053918082267045975, + 0.0044951592572033405, + 0.005457396153360605, + 0.0055405693128705025, + 0.006661797408014536, + 0.004370656795799732, + 0.005159936845302582, + 0.0027294072788208723, + 0.002711669309064746, + 0.004056911449879408, + 0.0031751617789268494, + 0.0026487091090530157, + 0.0047797346487641335, + 0.002858315361663699, + 0.0028391501400619745, + 0.01602594181895256, + 0.01578398607671261, + 0.014476485550403595, + 0.025173109024763107, + 0.021497095003724098, + 0.02595522254705429, + 0.049259766936302185, + 0.028514843434095383, + 0.005620080046355724, + 0.006052777171134949, + 0.0037607767153531313, + 0.004109942354261875, + 0.006603437475860119, + 0.008805356919765472, + 0.006910545285791159, + 0.0068305921740829945, + 0.022060178220272064, + 0.016882585361599922, + 0.02518288418650627, + 0.026758870109915733, + 0.016078030690550804, + 0.14320309460163116, + 0.023174062371253967, + 0.15783162415027618, + 0.004755985923111439, + 0.008188816718757153, + 0.0060810805298388, + 0.005173155106604099, + 0.003702998161315918, + 0.003668694756925106, + 0.005673948675394058, + 0.006297720596194267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.006738320458680391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947748199105263 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947748199105263, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947748199105263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05537702143192291, + 0.025715475901961327, + 0.07615631073713303, + 0.01488649845123291, + 0.006008592434227467, + 0.008470906876027584, + 0.042033564299345016, + 0.6747257113456726, + 0.059867002069950104, + 0.03675883263349533 + ], + "credit_map": [ + 0.009033787995576859, + 0.007013524882495403, + 0.007388708181679249, + 0.005008496809750795, + 0.010786272585391998, + 0.005520903039723635, + 0.006526034791022539, + 0.004099301993846893, + 0.0033217831514775753, + 0.003053990425541997, + 0.004836658947169781, + 0.004318051505833864, + 0.0024524787440896034, + 0.002066527958959341, + 0.0021328909788280725, + 0.0035331000108271837, + 0.007748080417513847, + 0.009888797998428345, + 0.011091244406998158, + 0.007264682557433844, + 0.00895786751061678, + 0.009777771309018135, + 0.008227049373090267, + 0.013200825080275536, + 0.0013180924579501152, + 0.0014633180107921362, + 0.0015365912113338709, + 0.0012617831816896796, + 0.0032640115823596716, + 0.0019235939253121614, + 0.0016620173119008541, + 0.0024570939131081104, + 0.0005414652405306697, + 0.0008521041600033641, + 0.0008392009767703712, + 0.0008895859937183559, + 0.0008546407334506512, + 0.000716830079909414, + 0.0005913404165767133, + 0.0007234254735521972, + 0.0007735663093626499, + 0.0007543743122369051, + 0.0009069076040759683, + 0.001202177256345749, + 0.0008348852279596031, + 0.0009107112418860197, + 0.0018977257423102856, + 0.0011905605206266046, + 0.004527784883975983, + 0.003611368825659156, + 0.002977921161800623, + 0.0045438483357429504, + 0.005292882677167654, + 0.0034380070865154266, + 0.01237424649298191, + 0.00526751484721899, + 0.041272882372140884, + 0.01846061460673809, + 0.018627895042300224, + 0.023353632539510727, + 0.0196148119866848, + 0.2264801412820816, + 0.04625987261533737, + 0.28065595030784607, + 0.0036204077769070864, + 0.002827928401529789, + 0.0028468570671975613, + 0.003584927646443248, + 0.002425546757876873, + 0.01963776908814907, + 0.0033729188144207, + 0.021550653502345085, + 0.004187399055808783, + 0.005030174273997545, + 0.00511113042011857, + 0.004813665524125099, + 0.0029866783879697323, + 0.0037982179783284664, + 0.003703418420627713, + 0.007128153461962938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.005947748199105263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000347971916199 + ], + "opposite_logprob_tokens": [ + -2.6950035095214844 + ], + "logprob_reward": -0.07000347971916199, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000347971916199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12733374536037445, + 0.11359404027462006, + 0.07175649702548981, + 0.03997079282999039, + 0.03970574215054512, + 0.062439464032649994, + 0.09765375405550003, + 0.17482218146324158, + 0.15065541863441467, + 0.1220683753490448 + ], + "credit_map": [ + 0.02481873333454132, + 0.0193751510232687, + 0.01730096898972988, + 0.015475835651159286, + 0.016530873253941536, + 0.012869981117546558, + 0.013464564457535744, + 0.007497629150748253, + 0.018637560307979584, + 0.01332719624042511, + 0.01532390434294939, + 0.024140041321516037, + 0.01002899743616581, + 0.010540659539401531, + 0.009865760803222656, + 0.011729918420314789, + 0.007094669621437788, + 0.011727516539394855, + 0.011431192979216576, + 0.007644519675523043, + 0.009254015982151031, + 0.00837548915296793, + 0.008009103126823902, + 0.008219987154006958, + 0.0034232516773045063, + 0.0038934173062443733, + 0.003562863217666745, + 0.003803297644481063, + 0.008992781862616539, + 0.00535624148324132, + 0.0037359287962317467, + 0.007203014567494392, + 0.004068917129188776, + 0.0049284398555755615, + 0.004208747763186693, + 0.0056978738866746426, + 0.006785422097891569, + 0.005495807155966759, + 0.004162553697824478, + 0.004357980098575354, + 0.006418395321816206, + 0.008538358844816685, + 0.008169940672814846, + 0.007587461266666651, + 0.0056752366945147514, + 0.010984552092850208, + 0.008329500444233418, + 0.00673602195456624, + 0.007490938529372215, + 0.012152929790318012, + 0.013422640971839428, + 0.009756975807249546, + 0.0127877714112401, + 0.015421449206769466, + 0.012505914084613323, + 0.014115135185420513, + 0.014200798235833645, + 0.024255001917481422, + 0.014301990158855915, + 0.012063417583703995, + 0.038656726479530334, + 0.02554604411125183, + 0.01359481830149889, + 0.03220338374376297, + 0.017386676743626595, + 0.015943923965096474, + 0.01668464206159115, + 0.02222624234855175, + 0.01626521162688732, + 0.02361179143190384, + 0.019771499559283257, + 0.018765442073345184, + 0.014451992698013783, + 0.016243815422058105, + 0.010082367807626724, + 0.017848394811153412, + 0.010306714102625847, + 0.012309989891946316, + 0.01336060743778944, + 0.027464495971798897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.07000347971916199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal 3D rendering, a low-angle medium shot frames a powerful construction warrior, his body huddled in a low crouch on a chaotic, smoke-filled battlefield at dusk. He wears worn, black leather armor over chain mail, with massive steel pauldrons protecting his broad shoulders. His face, with its chiseled features, square jaw, and resolute brows, is a mask of fierce loyalty, his eyes fixed on an unseen enemy. His hands, clasped in leather gloves, grip the shafts of a pair of halberds, knuckles white. Suddenly, the earth beneath him erupts violently, propelling him to the surface like an underground explosion. As the warrior is propelled into the air, the camera executes a rapid, tight zoom, focusing in realistic detail. This rapid movement captures the instantaneous shift in his expression from steely determination to pure shock, his eyes widening with pupils dilated. The zoom highlights the powerful ripples of his biceps and deltoids straining beneath his skin in response to the unexpected force, while a spray of dust and pebbles moves in slow-motion shockwaves around his rising body. The entire scene is bathed in the low, warm, orange glow of the setting sun, which glints off the armored edges of the sun, illuminating the swirling particles of smoke and debris, creating an atmosphere of tense and dramatic violence.", + "seed": 9446, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_5325_seed_9446_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": -0.02699742205440998, + "vlm_logprob": -0.02699742205440998, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.02699742205440998, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03386717289686203 + ], + "opposite_logprob_tokens": [ + -3.533867120742798 + ], + "logprob_reward": -0.03386717289686203, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03386717289686203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09291128069162369, + 0.0322023406624794, + 0.044545114040374756, + 0.035746682435274124, + 0.5448722839355469, + 0.038317713886499405, + 0.06768740713596344, + 0.035598285496234894, + 0.042354609817266464, + 0.06576418876647949 + ], + "credit_map": [ + 0.015007058158516884, + 0.011254668235778809, + 0.012043607421219349, + 0.008995037525892258, + 0.00897557009011507, + 0.011182177811861038, + 0.005649459082633257, + 0.01980370469391346, + 0.003552481299266219, + 0.004682379309087992, + 0.004847585689276457, + 0.0034414553083479404, + 0.0055322544649243355, + 0.0023731207475066185, + 0.003381160320714116, + 0.0043919021263718605, + 0.0065298243425786495, + 0.004707339685410261, + 0.005491282790899277, + 0.0037767640314996243, + 0.00774637283757329, + 0.0055337511003017426, + 0.005992045626044273, + 0.004767735954374075, + 0.003613656386733055, + 0.0048867580480873585, + 0.0048195431008934975, + 0.0042188712395727634, + 0.004852290730923414, + 0.004308235365897417, + 0.005263776518404484, + 0.0037835510447621346, + 0.023603009060025215, + 0.06351695954799652, + 0.017444631084799767, + 0.022787367925047874, + 0.033494967967271805, + 0.043714966624975204, + 0.2139783501625061, + 0.12633204460144043, + 0.002128027845174074, + 0.00479188933968544, + 0.0049874321557581425, + 0.004456640221178532, + 0.005186029709875584, + 0.004623359069228172, + 0.005921407137066126, + 0.006222928408533335, + 0.004373657517135143, + 0.012110029347240925, + 0.008604207076132298, + 0.008376515470445156, + 0.008098171092569828, + 0.00548179354518652, + 0.012016508728265762, + 0.008626523427665234, + 0.0024200084153562784, + 0.0033697488252073526, + 0.004555373452603817, + 0.005439838394522667, + 0.0030453396029770374, + 0.006066371686756611, + 0.007909896783530712, + 0.0027917122934013605, + 0.004623491782695055, + 0.002943244995549321, + 0.004737745504826307, + 0.005282003432512283, + 0.0038671623915433884, + 0.008579814806580544, + 0.004605418536812067, + 0.00771572720259428, + 0.0064478968270123005, + 0.008314583450555801, + 0.007661568466573954, + 0.010603180155158043, + 0.005422842688858509, + 0.00713688088580966, + 0.009618280455470085, + 0.010558951646089554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.03386717289686203 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2201724648475647, + 0.06992728263139725, + 0.1686088591814041, + 0.057428549975156784, + 0.19636258482933044, + 0.045272547751665115, + 0.06243327632546425, + 0.04049410670995712, + 0.07274790853261948, + 0.06655246764421463 + ], + "credit_map": [ + 0.028520651161670685, + 0.024976659566164017, + 0.03334704786539078, + 0.027015510946512222, + 0.02538003772497177, + 0.024118419736623764, + 0.023132650181651115, + 0.033681467175483704, + 0.01078796200454235, + 0.009270096197724342, + 0.009383194148540497, + 0.008347825147211552, + 0.009780865162611008, + 0.005749712698161602, + 0.00839759036898613, + 0.008210024796426296, + 0.00977308489382267, + 0.006142403930425644, + 0.051606032997369766, + 0.007058162707835436, + 0.017894545570015907, + 0.005544258747249842, + 0.0629148781299591, + 0.007675481028854847, + 0.005671306978911161, + 0.008175364695489407, + 0.007299529854208231, + 0.005159749183803797, + 0.00833162385970354, + 0.009095967747271061, + 0.007787107024341822, + 0.005907895043492317, + 0.012317147105932236, + 0.024301212280988693, + 0.011385265737771988, + 0.013260694220662117, + 0.01937294378876686, + 0.016304846853017807, + 0.060604214668273926, + 0.03881625086069107, + 0.003279769327491522, + 0.008889986202120781, + 0.006200266536325216, + 0.004937542136758566, + 0.005400074180215597, + 0.003997317980974913, + 0.0071828970685601234, + 0.0053846868686378, + 0.004029677715152502, + 0.00766333844512701, + 0.008394870907068253, + 0.007313860114663839, + 0.006536464206874371, + 0.006744626443833113, + 0.011273689568042755, + 0.010476738214492798, + 0.0038419764023274183, + 0.0036366856656968594, + 0.005339848343282938, + 0.0065690441988408566, + 0.00371514237485826, + 0.008407353423535824, + 0.006514308508485556, + 0.002469744998961687, + 0.013398255221545696, + 0.00468958867713809, + 0.007199764717370272, + 0.006077926605939865, + 0.006557457614690065, + 0.018607448786497116, + 0.006880775094032288, + 0.009336686693131924, + 0.00672411173582077, + 0.007199267391115427, + 0.007065337151288986, + 0.010046896524727345, + 0.005066649056971073, + 0.012570780701935291, + 0.0074646687135100365, + 0.010414747521281242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "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 main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951540544629097 + ], + "opposite_logprob_tokens": [ + -5.130951404571533 + ], + "logprob_reward": -0.005951540544629097, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951540544629097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06578733772039413, + 0.037502069026231766, + 0.12049324065446854, + 0.1024499461054802, + 0.1227322593331337, + 0.05773704871535301, + 0.10836564004421234, + 0.02025340124964714, + 0.32807567715644836, + 0.03660346567630768 + ], + "credit_map": [ + 0.00789814256131649, + 0.008051019161939621, + 0.007605398539453745, + 0.005745825357735157, + 0.009463771246373653, + 0.007362962234765291, + 0.007829923182725906, + 0.011830284260213375, + 0.005320762284100056, + 0.006269835401326418, + 0.004244097508490086, + 0.0054159886203706264, + 0.006927156820893288, + 0.0025999480858445168, + 0.0038278663996607065, + 0.0028964094817638397, + 0.008466548286378384, + 0.005841766484081745, + 0.02559439279139042, + 0.005782127380371094, + 0.016745558008551598, + 0.0070012654177844524, + 0.04522496089339256, + 0.005836620461195707, + 0.01556695532053709, + 0.010758885182440281, + 0.015497530810534954, + 0.012973194941878319, + 0.011689819395542145, + 0.008919638581573963, + 0.014057167805731297, + 0.012986741028726101, + 0.008327585645020008, + 0.01637927070260048, + 0.008527562953531742, + 0.006564945448189974, + 0.011436926200985909, + 0.01016947254538536, + 0.034553855657577515, + 0.026772629469633102, + 0.006883396301418543, + 0.009759975597262383, + 0.0053309365175664425, + 0.00660464633256197, + 0.005683038849383593, + 0.005528590176254511, + 0.010209762491285801, + 0.007736698258668184, + 0.007821938954293728, + 0.014503413811326027, + 0.013452311046421528, + 0.008431665599346161, + 0.00932356622070074, + 0.010029730387032032, + 0.016335539519786835, + 0.028467470780014992, + 0.0014982816064730287, + 0.001980568515136838, + 0.0022710319608449936, + 0.0035364821087569, + 0.0016767509514465928, + 0.003598319599404931, + 0.004288756754249334, + 0.0014032081235200167, + 0.06158822402358055, + 0.017477242276072502, + 0.03184639289975166, + 0.030462607741355896, + 0.013522100634872913, + 0.036072179675102234, + 0.03658946603536606, + 0.10051743686199188, + 0.003896544221788645, + 0.005228990223258734, + 0.006046683993190527, + 0.006386501714587212, + 0.002647622488439083, + 0.003657048800960183, + 0.00453488202765584, + 0.004205187316983938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.005951540544629097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012546183541417122 + ], + "opposite_logprob_tokens": [ + -4.387546062469482 + ], + "logprob_reward": -0.012546183541417122, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012546183541417122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20164652168750763, + 0.1538538634777069, + 0.06731437146663666, + 0.050273872911930084, + 0.10871317237615585, + 0.03362097218632698, + 0.10743465274572372, + 0.11411717534065247, + 0.06759687513113022, + 0.09542852640151978 + ], + "credit_map": [ + 0.022943193092942238, + 0.017361490055918694, + 0.029105396941304207, + 0.02459958754479885, + 0.024953318759799004, + 0.024008849635720253, + 0.01581026427447796, + 0.042864423245191574, + 0.019511666148900986, + 0.0246506966650486, + 0.025051185861229897, + 0.014128392562270164, + 0.030500927940011024, + 0.010579912923276424, + 0.014564353972673416, + 0.014866744168102741, + 0.009424122981727123, + 0.006854061037302017, + 0.009098591282963753, + 0.006485668942332268, + 0.010596385225653648, + 0.007007961627095938, + 0.011454897001385689, + 0.006392677314579487, + 0.003832900896668434, + 0.0065472484566271305, + 0.008054357022047043, + 0.0044013988226652145, + 0.008050813339650631, + 0.007416474167257547, + 0.006488955579698086, + 0.005481726955622435, + 0.00611050333827734, + 0.014705364592373371, + 0.006712120026350021, + 0.007796372752636671, + 0.011245563626289368, + 0.009477971121668816, + 0.03696756437420845, + 0.015697713941335678, + 0.0028613030444830656, + 0.005536473356187344, + 0.004810213111341, + 0.003981861751526594, + 0.0036099073477089405, + 0.0033989844378083944, + 0.004919103346765041, + 0.004503125790506601, + 0.007039892952889204, + 0.01846180111169815, + 0.013235576450824738, + 0.009912203997373581, + 0.009694154374301434, + 0.010461805388331413, + 0.016850421205163002, + 0.021778792142868042, + 0.010028257966041565, + 0.012535933405160904, + 0.017894065007567406, + 0.014152351766824722, + 0.011099613271653652, + 0.024685125797986984, + 0.01670895703136921, + 0.007012873888015747, + 0.015796925872564316, + 0.004691948648542166, + 0.005669211037456989, + 0.0063458834774792194, + 0.005772527307271957, + 0.01678304187953472, + 0.005759201478213072, + 0.006778130307793617, + 0.010500725358724594, + 0.011894508264958858, + 0.010341651737689972, + 0.013651852495968342, + 0.007453169673681259, + 0.014393663965165615, + 0.013044299557805061, + 0.014148651622235775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.012546183541417122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203103065490723 + ], + "opposite_logprob_tokens": [ + -2.8120310306549072 + ], + "logprob_reward": -0.06203103065490723, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203103065490723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07648056745529175, + 0.048457056283950806, + 0.03220849856734276, + 0.08839669823646545, + 0.12295280396938324, + 0.06349112093448639, + 0.0707005187869072, + 0.052373167127370834, + 0.15965864062309265, + 0.28528088331222534 + ], + "credit_map": [ + 0.009820999577641487, + 0.009729430079460144, + 0.010456816293299198, + 0.00826970487833023, + 0.011729232966899872, + 0.007908671163022518, + 0.008270219899713993, + 0.010295494459569454, + 0.00590226287022233, + 0.0066611273214221, + 0.009390869177877903, + 0.0038516416680067778, + 0.009025584906339645, + 0.0034866430796682835, + 0.005156714003533125, + 0.0049822148866951466, + 0.0027976413257420063, + 0.002943123923614621, + 0.005961332470178604, + 0.0032654309179633856, + 0.0034821878653019667, + 0.0038460837677121162, + 0.007708391174674034, + 0.0022043129429221153, + 0.011492029763758183, + 0.008872227743268013, + 0.013794613070786, + 0.011277996003627777, + 0.010652453638613224, + 0.00887787714600563, + 0.012397699058055878, + 0.011031812988221645, + 0.00650441600009799, + 0.01558217778801918, + 0.004976538475602865, + 0.006362607702612877, + 0.008089965209364891, + 0.010879609733819962, + 0.04506275802850723, + 0.025494743138551712, + 0.006986261811107397, + 0.008500313386321068, + 0.008406220003962517, + 0.006567338481545448, + 0.008719808422029018, + 0.007027429062873125, + 0.009644954465329647, + 0.007638801354914904, + 0.005166742950677872, + 0.00887745339423418, + 0.00900780875235796, + 0.007053944747895002, + 0.00715402327477932, + 0.008613151498138905, + 0.008686388842761517, + 0.016141008585691452, + 0.004677755758166313, + 0.0059882113710045815, + 0.006708299275487661, + 0.007882203906774521, + 0.004513303749263287, + 0.008947087451815605, + 0.009016405791044235, + 0.004639904014766216, + 0.02284531109035015, + 0.009591718204319477, + 0.016121957451105118, + 0.015604554675519466, + 0.008997652679681778, + 0.022827114909887314, + 0.017580239102244377, + 0.046090103685855865, + 0.02220018021762371, + 0.040483105927705765, + 0.05139114707708359, + 0.06243237107992172, + 0.020585857331752777, + 0.020031118765473366, + 0.035382118076086044, + 0.032774996012449265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.06203103065490723 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal 3D rendering, a low-angle medium shot frames a powerful construction warrior, his body huddled in a low crouch on a chaotic, smoke-filled battlefield at dusk. He wears worn, black leather armor over chain mail, with massive steel pauldrons protecting his broad shoulders. His face, with its chiseled features, square jaw, and resolute brows, is a mask of fierce loyalty, his eyes fixed on an unseen enemy. His hands, clasped in leather gloves, grip the shafts of a pair of halberds, knuckles white. Suddenly, the earth beneath him erupts violently, propelling him to the surface like an underground explosion. As the warrior is propelled into the air, the camera executes a rapid, tight zoom, focusing in realistic detail. This rapid movement captures the instantaneous shift in his expression from steely determination to pure shock, his eyes widening with pupils dilated. The zoom highlights the powerful ripples of his biceps and deltoids straining beneath his skin in response to the unexpected force, while a spray of dust and pebbles moves in slow-motion shockwaves around his rising body. The entire scene is bathed in the low, warm, orange glow of the setting sun, which glints off the armored edges of the sun, illuminating the swirling particles of smoke and debris, creating an atmosphere of tense and dramatic violence.", + "seed": 9447, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_5325_seed_9447_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": -0.030101233907043935, + "vlm_logprob": -0.030101233907043935, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.030101233907043935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038299817591905594 + ], + "opposite_logprob_tokens": [ + -3.413299798965454 + ], + "logprob_reward": -0.038299817591905594, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.038299817591905594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13262197375297546, + 0.06626950949430466, + 0.047985389828681946, + 0.1436668485403061, + 0.21232163906097412, + 0.05044311285018921, + 0.07396268844604492, + 0.04080980271100998, + 0.027806367725133896, + 0.20411263406276703 + ], + "credit_map": [ + 0.01362574566155672, + 0.01594952493906021, + 0.015094039961695671, + 0.02107343263924122, + 0.02416010946035385, + 0.011684236116707325, + 0.016574211418628693, + 0.014460686594247818, + 0.007485654670745134, + 0.009328071027994156, + 0.01020051259547472, + 0.005218695383518934, + 0.013425704091787338, + 0.006368093658238649, + 0.007364867255091667, + 0.006877914536744356, + 0.006034629885107279, + 0.005609449464827776, + 0.005438121035695076, + 0.004159979522228241, + 0.00728148128837347, + 0.005154198035597801, + 0.009270346723496914, + 0.005037184804677963, + 0.013565354980528355, + 0.013472938910126686, + 0.01113643404096365, + 0.014279487542808056, + 0.01955084688961506, + 0.015341163612902164, + 0.047587815672159195, + 0.008732803165912628, + 0.01905694045126438, + 0.07046842575073242, + 0.019419880583882332, + 0.028994625434279442, + 0.01723690703511238, + 0.019183548167347908, + 0.02541598491370678, + 0.012545325793325901, + 0.0039081815630197525, + 0.005179372150450945, + 0.008231121115386486, + 0.007751803379505873, + 0.006006450857967138, + 0.005086245480924845, + 0.006536719389259815, + 0.00774322310462594, + 0.008794727735221386, + 0.011182778514921665, + 0.0060230037197470665, + 0.007736993953585625, + 0.011055521667003632, + 0.007033332716673613, + 0.009378083981573582, + 0.012758246622979641, + 0.0039043077267706394, + 0.006792285013943911, + 0.005124332383275032, + 0.0036351983435451984, + 0.006522404495626688, + 0.0058985683135688305, + 0.004587233066558838, + 0.004345475230365992, + 0.0031195117626339197, + 0.00439595989882946, + 0.003600087482482195, + 0.0026529415044933558, + 0.00495936581864953, + 0.003943540621548891, + 0.0031094111036509275, + 0.002025549067184329, + 0.020247792825102806, + 0.018609492108225822, + 0.029369300231337547, + 0.044395048171281815, + 0.014800721779465675, + 0.03193977475166321, + 0.014742556028068066, + 0.03000795841217041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.038299817591905594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197020515799522 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197020515799522, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197020515799522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14043967425823212, + 0.10250397771596909, + 0.09472991526126862, + 0.07860387116670609, + 0.14065593481063843, + 0.07465136796236038, + 0.05936075747013092, + 0.07092247903347015, + 0.1574467122554779, + 0.08068530261516571 + ], + "credit_map": [ + 0.015295644290745258, + 0.016907505691051483, + 0.020850691944360733, + 0.02066671848297119, + 0.02484344132244587, + 0.013832163996994495, + 0.01483503170311451, + 0.013208480551838875, + 0.0109475739300251, + 0.012795664370059967, + 0.014011630788445473, + 0.008303358219563961, + 0.016662755981087685, + 0.011654474772512913, + 0.012819718569517136, + 0.015308799222111702, + 0.01348201371729374, + 0.012110333889722824, + 0.011704151518642902, + 0.011466875672340393, + 0.01046677865087986, + 0.008957005105912685, + 0.01720895990729332, + 0.009333798661828041, + 0.005445184651762247, + 0.007312655448913574, + 0.011927313171327114, + 0.008516145870089531, + 0.011021954007446766, + 0.0073541984893381596, + 0.022084495052695274, + 0.004941917955875397, + 0.011465044692158699, + 0.03795536607503891, + 0.041993893682956696, + 0.011310460977256298, + 0.009208873845636845, + 0.011042707599699497, + 0.011172554455697536, + 0.006507045589387417, + 0.005340456496924162, + 0.011637084186077118, + 0.014688349328935146, + 0.008769899606704712, + 0.01121545024216175, + 0.00765289505943656, + 0.00633295439183712, + 0.009014279581606388, + 0.007186912931501865, + 0.004914230201393366, + 0.00537845678627491, + 0.0052823335863649845, + 0.007790718227624893, + 0.0061391680501401424, + 0.0074195400811731815, + 0.015249399468302727, + 0.009063147008419037, + 0.008526258170604706, + 0.008011494763195515, + 0.008056470192968845, + 0.01146569475531578, + 0.008691782131791115, + 0.008935666643083096, + 0.008171968162059784, + 0.01097105722874403, + 0.012928023934364319, + 0.009253055788576603, + 0.007350347936153412, + 0.016812527552247047, + 0.044770482927560806, + 0.045783836394548416, + 0.009577381424605846, + 0.008250609040260315, + 0.010021277703344822, + 0.008935225196182728, + 0.012019411660730839, + 0.008696982637047768, + 0.012693936005234718, + 0.009225179441273212, + 0.010842685587704182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.018197020515799522 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.0940372496843338, + 0.04470377787947655, + 0.03607327491044998, + 0.06425774842500687, + 0.2941049635410309, + 0.13092583417892456, + 0.0851505845785141, + 0.12588568031787872, + 0.035391442477703094, + 0.08946942538022995 + ], + "credit_map": [ + 0.009515523910522461, + 0.012425719760358334, + 0.009149233810603619, + 0.011727649718523026, + 0.021763678640127182, + 0.008321440778672695, + 0.011793317273259163, + 0.009340684860944748, + 0.005134528502821922, + 0.006151610054075718, + 0.008127572946250439, + 0.003954329993575811, + 0.008102130144834518, + 0.0033081865403801203, + 0.005638867150992155, + 0.004286549985408783, + 0.004435000941157341, + 0.004011563025414944, + 0.004482184071093798, + 0.004637802019715309, + 0.003722903784364462, + 0.002833128906786442, + 0.008846450597047806, + 0.0031042415648698807, + 0.004706797190010548, + 0.006106771063059568, + 0.009289340116083622, + 0.006957052741199732, + 0.010021043010056019, + 0.0037833747919648886, + 0.019277401268482208, + 0.004115974064916372, + 0.02455288916826248, + 0.09847927838563919, + 0.04904836788773537, + 0.025560416281223297, + 0.02182343415915966, + 0.018736714497208595, + 0.026180503889918327, + 0.029723362997174263, + 0.01140670571476221, + 0.011685183271765709, + 0.022353922948241234, + 0.017506912350654602, + 0.014736371114850044, + 0.013154997490346432, + 0.020604826509952545, + 0.019476911053061485, + 0.006739422213286161, + 0.010224112309515476, + 0.010150118730962276, + 0.008895667269825935, + 0.013468543998897076, + 0.009926185943186283, + 0.009777682833373547, + 0.01596885174512863, + 0.013039127923548222, + 0.020424749702215195, + 0.01263835746794939, + 0.011021854355931282, + 0.01976725459098816, + 0.017879461869597435, + 0.016421077772974968, + 0.014693792909383774, + 0.0035835029557347298, + 0.0038181529380381107, + 0.002850675256922841, + 0.001861006603576243, + 0.0030445426236838102, + 0.007771440781652927, + 0.009005601517856121, + 0.0034565208479762077, + 0.009386765770614147, + 0.010977613739669323, + 0.011871268972754478, + 0.012866544537246227, + 0.013099897652864456, + 0.009639488533139229, + 0.010366187430918217, + 0.011261660605669022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 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 main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08415228128433228, + 0.16600069403648376, + 0.1662670075893402, + 0.16716542840003967, + 0.1525898277759552, + 0.04012575373053551, + 0.029835568740963936, + 0.058733414858579636, + 0.03848007321357727, + 0.09664996713399887 + ], + "credit_map": [ + 0.01138767134398222, + 0.008740435354411602, + 0.013236600905656815, + 0.01287265494465828, + 0.009525422006845474, + 0.0077406675554811954, + 0.00924096442759037, + 0.011407863348722458, + 0.017642708495259285, + 0.01904817298054695, + 0.018173441290855408, + 0.01225439552217722, + 0.02894502505660057, + 0.01364218257367611, + 0.02020926959812641, + 0.03608549386262894, + 0.028731347993016243, + 0.016327960416674614, + 0.02098543383181095, + 0.02276618406176567, + 0.01553487777709961, + 0.012702284380793571, + 0.035571321845054626, + 0.013647600077092648, + 0.0156779233366251, + 0.012246428057551384, + 0.014721124432981014, + 0.013182587921619415, + 0.02227773889899254, + 0.011657253839075565, + 0.0687946230173111, + 0.008607749827206135, + 0.010539252310991287, + 0.04233318194746971, + 0.044346485286951065, + 0.011624708771705627, + 0.009535769000649452, + 0.010463990271091461, + 0.012598736211657524, + 0.011147703975439072, + 0.0032944981940090656, + 0.003257851582020521, + 0.005680022295564413, + 0.006302881985902786, + 0.005316638387739658, + 0.0037089576944708824, + 0.0064516072161495686, + 0.006113295443356037, + 0.0045060450211167336, + 0.0030144108459353447, + 0.002583123277872801, + 0.0022884742356836796, + 0.0029620237182825804, + 0.002507696161046624, + 0.003838716773316264, + 0.00813507754355669, + 0.00653211772441864, + 0.009627229534089565, + 0.006086469162255526, + 0.005533585324883461, + 0.007904035039246082, + 0.008780144155025482, + 0.006543032359331846, + 0.007726800628006458, + 0.0024646820966154337, + 0.0038430020213127136, + 0.0032269463408738375, + 0.0027774658519774675, + 0.0037325373850762844, + 0.009750261902809143, + 0.010181695222854614, + 0.002503482159227133, + 0.008739236742258072, + 0.008003505878150463, + 0.012432971969246864, + 0.01964409090578556, + 0.009375878609716892, + 0.017802204936742783, + 0.007474659476429224, + 0.013177413493394852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.009787900373339653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896443456411362 + ], + "opposite_logprob_tokens": [ + -2.5789644718170166 + ], + "logprob_reward": -0.07896443456411362, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896443456411362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20715229213237762, + 0.04974847286939621, + 0.07968387007713318, + 0.05787890404462814, + 0.20379507541656494, + 0.05421857908368111, + 0.027156557887792587, + 0.06491166353225708, + 0.06618445366621017, + 0.18927015364170074 + ], + "credit_map": [ + 0.020757177844643593, + 0.021679310128092766, + 0.029494309797883034, + 0.027123814448714256, + 0.0416528657078743, + 0.019205480813980103, + 0.023943446576595306, + 0.02329585887491703, + 0.006314021535217762, + 0.007520109415054321, + 0.007229417562484741, + 0.004471082240343094, + 0.010565845295786858, + 0.003801565384492278, + 0.005342807620763779, + 0.004503617063164711, + 0.008156366646289825, + 0.01294534932821989, + 0.006988434586673975, + 0.009663033299148083, + 0.008746318519115448, + 0.014131688512861729, + 0.010137807577848434, + 0.008914867416024208, + 0.0050166817381978035, + 0.006379156839102507, + 0.006280946079641581, + 0.0056783403269946575, + 0.010862366296350956, + 0.006077498197555542, + 0.014319033361971378, + 0.003264880971983075, + 0.014129004441201687, + 0.06677223742008209, + 0.04855387285351753, + 0.017649926245212555, + 0.014781446196138859, + 0.010135827586054802, + 0.01777569204568863, + 0.013997048139572144, + 0.004012790508568287, + 0.006737783085554838, + 0.0079871267080307, + 0.007345535326749086, + 0.009211226366460323, + 0.0054037906229496, + 0.004161621909588575, + 0.009358697570860386, + 0.0029459178913384676, + 0.0030937239062041044, + 0.003262521233409643, + 0.0032083140686154366, + 0.00410045450553298, + 0.002379949437454343, + 0.003288023406639695, + 0.004877651110291481, + 0.006658852566033602, + 0.00775239709764719, + 0.008371223695576191, + 0.00651172548532486, + 0.00910759624093771, + 0.009103045798838139, + 0.00713957566767931, + 0.010267241857945919, + 0.003979294560849667, + 0.007081157993525267, + 0.0055331820622086525, + 0.003432232653722167, + 0.0059935590252280235, + 0.017309032380580902, + 0.018694015219807625, + 0.004161969758570194, + 0.016932493075728416, + 0.019840342923998833, + 0.019256383180618286, + 0.026432959362864494, + 0.031823672354221344, + 0.022168954834342003, + 0.028782222419977188, + 0.024033116176724434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.07896443456411362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a surreal 3D rendering, a low-angle medium shot frames a powerful construction warrior, his body huddled in a low crouch on a chaotic, smoke-filled battlefield at dusk. He wears worn, black leather armor over chain mail, with massive steel pauldrons protecting his broad shoulders. His face, with its chiseled features, square jaw, and resolute brows, is a mask of fierce loyalty, his eyes fixed on an unseen enemy. His hands, clasped in leather gloves, grip the shafts of a pair of halberds, knuckles white. Suddenly, the earth beneath him erupts violently, propelling him to the surface like an underground explosion. As the warrior is propelled into the air, the camera executes a rapid, tight zoom, focusing in realistic detail. This rapid movement captures the instantaneous shift in his expression from steely determination to pure shock, his eyes widening with pupils dilated. The zoom highlights the powerful ripples of his biceps and deltoids straining beneath his skin in response to the unexpected force, while a spray of dust and pebbles moves in slow-motion shockwaves around his rising body. The entire scene is bathed in the low, warm, orange glow of the setting sun, which glints off the armored edges of the sun, illuminating the swirling particles of smoke and debris, creating an atmosphere of tense and dramatic violence.", + "seed": 9448, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000158/group_9/index_5325_seed_9448_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": -0.02014419510960579, + "vlm_logprob": -0.02014419510960579, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.02014419510960579, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015193680301308632 + ], + "opposite_logprob_tokens": [ + -4.515193462371826 + ], + "logprob_reward": -0.015193680301308632, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015193680301308632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3722386658191681, + 0.027266860008239746, + 0.06270000338554382, + 0.03779495507478714, + 0.04635686054825783, + 0.03086545690894127, + 0.07651185244321823, + 0.06081339716911316, + 0.12095388025045395, + 0.16449803113937378 + ], + "credit_map": [ + 0.11538699269294739, + 0.035762958228588104, + 0.012717518955469131, + 0.015865230932831764, + 0.028732705861330032, + 0.1329777091741562, + 0.014988142997026443, + 0.015807459130883217, + 0.0030359074007719755, + 0.004693374037742615, + 0.0028249064926058054, + 0.00238327425904572, + 0.006006742361932993, + 0.003819181816652417, + 0.0020730167161673307, + 0.0024304569233208895, + 0.007068268023431301, + 0.00762057863175869, + 0.0051536038517951965, + 0.004832149483263493, + 0.010891788639128208, + 0.015086792409420013, + 0.004610802978277206, + 0.007436019368469715, + 0.0028261342085897923, + 0.008914860896766186, + 0.002632246585562825, + 0.003972200211137533, + 0.006036802660673857, + 0.0059213219210505486, + 0.003224803600460291, + 0.004266584757715464, + 0.004331968724727631, + 0.009029640816152096, + 0.006451250985264778, + 0.00465775839984417, + 0.0068491315469145775, + 0.004574475809931755, + 0.005431530997157097, + 0.005031106527894735, + 0.0029909182339906693, + 0.00425884872674942, + 0.0022571985609829426, + 0.004189484752714634, + 0.004190054256469011, + 0.006012078374624252, + 0.004803173243999481, + 0.0021637030877172947, + 0.007718631066381931, + 0.012368287891149521, + 0.005651975050568581, + 0.011439148336648941, + 0.010301767848432064, + 0.009008496068418026, + 0.012152643874287605, + 0.007870902307331562, + 0.004861814435571432, + 0.0039938827976584435, + 0.007635877467691898, + 0.005744052119553089, + 0.010750003159046173, + 0.00952319335192442, + 0.008560332469642162, + 0.009744248352944851, + 0.010913968086242676, + 0.014595461077988148, + 0.010914932936429977, + 0.016914820298552513, + 0.019233573228120804, + 0.01848941296339035, + 0.014036528766155243, + 0.01585518568754196, + 0.018473317846655846, + 0.021882513538002968, + 0.03787757083773613, + 0.02433035336434841, + 0.01591314934194088, + 0.017588600516319275, + 0.011348755098879337, + 0.017083769664168358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.015193680301308632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204702340066433 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204702340066433, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204702340066433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36783888936042786, + 0.12414874881505966, + 0.04991396516561508, + 0.033999525010585785, + 0.03842208534479141, + 0.08502087742090225, + 0.03251882642507553, + 0.10309015214443207, + 0.10771169513463974, + 0.05733529478311539 + ], + "credit_map": [ + 0.10168475657701492, + 0.03680873289704323, + 0.013778469525277615, + 0.018348069861531258, + 0.04412955790758133, + 0.11223353445529938, + 0.020052514970302582, + 0.02080325409770012, + 0.014015881344676018, + 0.023363588377833366, + 0.010181330144405365, + 0.010126987472176552, + 0.031813934445381165, + 0.013263649307191372, + 0.01106616947799921, + 0.010317201726138592, + 0.005321255419403315, + 0.00746254762634635, + 0.004664235282689333, + 0.005463017616420984, + 0.008480888791382313, + 0.010199379175901413, + 0.0036437821108847857, + 0.004678858444094658, + 0.0029900637455284595, + 0.009402446448802948, + 0.002692085225135088, + 0.0038197175599634647, + 0.005482125096023083, + 0.004359147045761347, + 0.0028552147559821606, + 0.002398725366219878, + 0.004073294810950756, + 0.007958920672535896, + 0.004554365295916796, + 0.004553846083581448, + 0.003920520655810833, + 0.004417485557496548, + 0.00438123382627964, + 0.004562418442219496, + 0.006977638695389032, + 0.015226414427161217, + 0.006557703949511051, + 0.010107229463756084, + 0.013369912281632423, + 0.009707177989184856, + 0.014749865047633648, + 0.008324939757585526, + 0.0018326494609937072, + 0.006493792403489351, + 0.0022696293890476227, + 0.004431372042745352, + 0.004619764629751444, + 0.005454187281429768, + 0.004184698686003685, + 0.0032327338121831417, + 0.006896741222590208, + 0.007856200449168682, + 0.012399173341691494, + 0.01261543482542038, + 0.01791563257575035, + 0.012944189831614494, + 0.01740143448114395, + 0.015061342157423496, + 0.008120075799524784, + 0.013222218491137028, + 0.009993034414947033, + 0.01417539082467556, + 0.018150117248296738, + 0.01358711440116167, + 0.014690057374536991, + 0.01577368564903736, + 0.00695117749273777, + 0.008131791837513447, + 0.010794022120535374, + 0.007607370149344206, + 0.005123710259795189, + 0.006283025722950697, + 0.005033030174672604, + 0.007411168422549963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014204702340066433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643371794372797 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.004643371794372797, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643371794372797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5793133974075317, + 0.02935415878891945, + 0.022688375785946846, + 0.047039467841386795, + 0.05178800970315933, + 0.08131474256515503, + 0.07841096073389053, + 0.06693487614393234, + 0.028544604778289795, + 0.014611393213272095 + ], + "credit_map": [ + 0.1889035552740097, + 0.05226760730147362, + 0.017230743542313576, + 0.02203105390071869, + 0.03355126827955246, + 0.21723616123199463, + 0.021949253976345062, + 0.02614375576376915, + 0.0030065872706472874, + 0.005390326026827097, + 0.0029324579518288374, + 0.0030239690095186234, + 0.005886439234018326, + 0.0034110688138753176, + 0.0025806038174778223, + 0.0031227085273712873, + 0.0023909565061330795, + 0.004984008148312569, + 0.0024323195684701204, + 0.002190881408751011, + 0.0025071396958082914, + 0.00406296830624342, + 0.0017872934695333242, + 0.002332808915525675, + 0.004010013770312071, + 0.00696680648252368, + 0.002695126459002495, + 0.005303237121552229, + 0.006404506042599678, + 0.01203268300741911, + 0.0039780111983418465, + 0.005649082828313112, + 0.003359084250405431, + 0.01114046573638916, + 0.008099092170596123, + 0.005460543092340231, + 0.007723012939095497, + 0.005980824586004019, + 0.006080842111259699, + 0.003944141790270805, + 0.006597063038498163, + 0.012555570341646671, + 0.006318952422589064, + 0.010805056430399418, + 0.010773037560284138, + 0.011383323930203915, + 0.014776604250073433, + 0.008105129934847355, + 0.006304644048213959, + 0.01642623171210289, + 0.004155738744884729, + 0.009449317120015621, + 0.010231374762952328, + 0.01171811856329441, + 0.011350897140800953, + 0.008774641901254654, + 0.005974631756544113, + 0.005917984060943127, + 0.010989303700625896, + 0.006281473208218813, + 0.010692614130675793, + 0.011121958494186401, + 0.008779147639870644, + 0.007177760824561119, + 0.002473236061632633, + 0.0030242549255490303, + 0.002834110986441374, + 0.004653720185160637, + 0.004269740078598261, + 0.003180854953825474, + 0.004304791335016489, + 0.003803896252065897, + 0.0011262515326961875, + 0.0014270002720877528, + 0.001694986829534173, + 0.0016390595119446516, + 0.001954294042661786, + 0.00319497543387115, + 0.001656128792092204, + 0.001918695867061615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.004643371794372797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46399983763694763, + 0.053073808550834656, + 0.037655510008335114, + 0.014125064015388489, + 0.157906174659729, + 0.04965198040008545, + 0.042904552072286606, + 0.062425218522548676, + 0.06499315798282623, + 0.053264718502759933 + ], + "credit_map": [ + 0.14015315473079681, + 0.046934086829423904, + 0.02009526640176773, + 0.019197724759578705, + 0.02940177358686924, + 0.16099590063095093, + 0.020545903593301773, + 0.026676040142774582, + 0.006447232328355312, + 0.0087608452886343, + 0.005585716105997562, + 0.004399283789098263, + 0.008540174923837185, + 0.010467504151165485, + 0.00436124624684453, + 0.004511803388595581, + 0.005192629061639309, + 0.005344476085156202, + 0.003111525671556592, + 0.0027780013624578714, + 0.004246974363923073, + 0.009946409612894058, + 0.0024811874609440565, + 0.004554304294288158, + 0.0011468837037682533, + 0.0024963703472167253, + 0.001138353138230741, + 0.0013925072271376848, + 0.002080590231344104, + 0.002800058340653777, + 0.0015595107106491923, + 0.0015107908984646201, + 0.030049733817577362, + 0.021582432091236115, + 0.01617858000099659, + 0.01774006336927414, + 0.012702349573373795, + 0.025871211662888527, + 0.01595962420105934, + 0.017822185531258583, + 0.003678318113088608, + 0.00776353944092989, + 0.004012067336589098, + 0.0061329626478254795, + 0.007395206019282341, + 0.010418171994388103, + 0.0060258121229708195, + 0.004225906450301409, + 0.002677035517990589, + 0.009055296890437603, + 0.0027920228894799948, + 0.007377853151410818, + 0.006672597955912352, + 0.004259410779923201, + 0.0061998525634408, + 0.0038704839535057545, + 0.00695780711248517, + 0.0062752957455813885, + 0.009706908836960793, + 0.005911606829613447, + 0.008297263644635677, + 0.010725383646786213, + 0.00683901272714138, + 0.0077119432389736176, + 0.0061498722061514854, + 0.007627381477504969, + 0.005045274738222361, + 0.009864001534879208, + 0.00870082713663578, + 0.010583187453448772, + 0.008824949152767658, + 0.008197657763957977, + 0.006341899745166302, + 0.006819530390202999, + 0.009423363953828812, + 0.007355062756687403, + 0.005863902624696493, + 0.004888051189482212, + 0.0063283247873187065, + 0.00624458771198988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.00464135454967618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a construction warrior wearing worn black leather armor over chain mail with massive steel pauldrons?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the warrior start in a low crouch on a chaotic, smoke-filled battlefield at dusk before being propelled upward by an eruption from the earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera execute a rapid, tight zoom focusing on the warrior's face and upper body as he is propelled into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a low, warm, orange glow of a setting sun that glints off the armored edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the warrior's expression shifting from determination to shock with widened eyes and dilated pupils while dust and pebbles move in slow-motion shockwaves around him?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037866562604904 + ], + "opposite_logprob_tokens": [ + -2.812037944793701 + ], + "logprob_reward": -0.062037866562604904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037866562604904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6202802062034607, + 0.07106219232082367, + 0.02213568612933159, + 0.018446607515215874, + 0.018384424969553947, + 0.042019229382276535, + 0.03751134127378464, + 0.05995989590883255, + 0.06848357617855072, + 0.04171686992049217 + ], + "credit_map": [ + 0.2127012312412262, + 0.04434339329600334, + 0.016394183039665222, + 0.020563190802931786, + 0.03861113637685776, + 0.24202148616313934, + 0.02268318273127079, + 0.022962359711527824, + 0.007733202539384365, + 0.015058734454214573, + 0.004867961164563894, + 0.005677387118339539, + 0.018544211983680725, + 0.007519538048654795, + 0.006369594484567642, + 0.005291555542498827, + 0.0015581361949443817, + 0.0047075278125703335, + 0.0026096501387655735, + 0.0023964366409927607, + 0.0034311125054955482, + 0.003674768377095461, + 0.0016380284214392304, + 0.002120023127645254, + 0.0016350829973816872, + 0.0047504776157438755, + 0.0011744053335860372, + 0.002280645305290818, + 0.002755071735009551, + 0.003069006372243166, + 0.0016052345745265484, + 0.0011766821844503284, + 0.0013124204706400633, + 0.004126259591430426, + 0.002217136789113283, + 0.0022743698209524155, + 0.0024880783166736364, + 0.0019742026925086975, + 0.002062530256807804, + 0.0019294259836897254, + 0.0038188460748642683, + 0.006553633604198694, + 0.0028968642000108957, + 0.005427399184554815, + 0.005793303716927767, + 0.007705105002969503, + 0.005874612368643284, + 0.00394946476444602, + 0.0031911565456539392, + 0.006623847410082817, + 0.002829532139003277, + 0.004627797286957502, + 0.0053657786920666695, + 0.0040449961088597775, + 0.007908979430794716, + 0.0029192548245191574, + 0.004051389172673225, + 0.004991541150957346, + 0.006320641376078129, + 0.006470552645623684, + 0.011592667549848557, + 0.00744327949360013, + 0.010393931530416012, + 0.00869588740170002, + 0.006410365458577871, + 0.007211253512650728, + 0.0062456983141601086, + 0.009462602436542511, + 0.01284843496978283, + 0.005793681368231773, + 0.011355525813996792, + 0.009156009182333946, + 0.004514588043093681, + 0.0050932057201862335, + 0.005201965570449829, + 0.005964057520031929, + 0.005953256040811539, + 0.005129718687385321, + 0.0040098619647324085, + 0.005850215442478657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.062037866562604904 + } + ], + "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/0000160/group_0/index_tensor([4677, 4677])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_tensor([4677, 4677])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1b5866cfa93c282f2d5f986adea46534040386f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_tensor([4677, 4677])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow and graceful 360-degree pan shot, from a slightly bird's-eye view, slowly unveils a sumptuous Sichuan hot pot feast. In the center of the frame, an exquisitely carved copper pot stands atop a dark, round wooden table. Its body gleams with a warm, patina of bronze, intricately carved with vintage patterns of auspicious clouds and dragons. Within, a vibrant, translucent red oil broth simmers, bubbling with countless plump dried chilies and Sichuan peppercorns. Wisps of white steam rise, blurring the background and adding a dreamy, soft-focus effect. Circling the hot pot, plates of premium, thinly sliced, marbling beef spread out, its red and white marbled veins clearly visible in the light, revealing its exquisite tenderness. Next to them, juicy green spinach, neatly arranged white jade mushrooms, and tender lotus root slices cut into petals, all arranged like stars surrounding the moon. A Chinese palace lantern suspended above the table casts a warm, soft yellow light, illuminating the splashes of oil from the broth, making them shimmer with golden glints and coating each ingredient with an exceptionally alluring sheen. The entire scene, dominated by rich reds and warm yellows, creates an atmosphere of tranquility and abundance, reminiscent of a classical poetic painting, imbued with the warmth and abundant joy of everyday life.", + "seed": 31510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_4677_seed_31510_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": -0.032394619192928076, + "vlm_logprob": -0.032394619192928076, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.032394619192928076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.05777806416153908 + ], + "opposite_logprob_tokens": [ + -2.9327781200408936 + ], + "logprob_reward": -0.05777806416153908, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05777806416153908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1175413727760315, + 0.06364911049604416, + 0.032548513263463974, + 0.4497450292110443, + 0.10927817970514297, + 0.0424613393843174, + 0.0386061929166317, + 0.030350632965564728, + 0.06706792861223221, + 0.04875169321894646 + ], + "credit_map": [ + 0.016679922118782997, + 0.015395433641970158, + 0.01399440597742796, + 0.009145236574113369, + 0.015992898494005203, + 0.007359782699495554, + 0.022479554638266563, + 0.016494140028953552, + 0.006492390297353268, + 0.008230315521359444, + 0.004013757687062025, + 0.005540368612855673, + 0.007646402809768915, + 0.012296881526708603, + 0.013756273314356804, + 0.005672721192240715, + 0.0039264061488211155, + 0.004001488909125328, + 0.0022280781995505095, + 0.0021200599148869514, + 0.00443279929459095, + 0.002305619651451707, + 0.010575126856565475, + 0.0029589354526251554, + 0.06865371763706207, + 0.019593844190239906, + 0.019625604152679443, + 0.01831158436834812, + 0.06902261078357697, + 0.04424881562590599, + 0.1812622845172882, + 0.029026588425040245, + 0.009988718666136265, + 0.008078278042376041, + 0.006749795284122229, + 0.007825898006558418, + 0.01704271510243416, + 0.012659178115427494, + 0.03964017331600189, + 0.007293420843780041, + 0.0056418986059725285, + 0.003955888096243143, + 0.0031527576502412558, + 0.002873998833820224, + 0.008235782384872437, + 0.007797651924192905, + 0.007628134451806545, + 0.0031752288341522217, + 0.0045463754795491695, + 0.00986540224403143, + 0.0025858539156615734, + 0.003305034479126334, + 0.004024825524538755, + 0.005062573589384556, + 0.005570362787693739, + 0.0036457646638154984, + 0.0030551378149539232, + 0.004922959487885237, + 0.002183803590014577, + 0.0027561532333493233, + 0.004296321887522936, + 0.0035209355410188437, + 0.006602801848202944, + 0.0030125193297863007, + 0.008089704439043999, + 0.006124445237219334, + 0.0050361789762973785, + 0.005726669915020466, + 0.011054972186684608, + 0.0074527389369904995, + 0.015036772936582565, + 0.008546446450054646, + 0.004316722508519888, + 0.004496271722018719, + 0.0031279672402888536, + 0.003293737769126892, + 0.015308992937207222, + 0.006062360946089029, + 0.005911897402256727, + 0.006233741994947195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.05777806416153908 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.006747674662619829 + ], + "opposite_logprob_tokens": [ + -5.006747722625732 + ], + "logprob_reward": -0.006747674662619829, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006747674662619829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21259987354278564, + 0.05432809144258499, + 0.11420972645282745, + 0.16513608396053314, + 0.05233440175652504, + 0.09030449390411377, + 0.06541718542575836, + 0.04679965600371361, + 0.06078953668475151, + 0.13808095455169678 + ], + "credit_map": [ + 0.028290562331676483, + 0.019529294222593307, + 0.015403705649077892, + 0.016311677172780037, + 0.03927164897322655, + 0.013496791943907738, + 0.0495750829577446, + 0.030721133574843407, + 0.00427052890881896, + 0.006194127723574638, + 0.0038284321781247854, + 0.004920911975204945, + 0.006625551730394363, + 0.00951207522302866, + 0.011807634495198727, + 0.007168835494667292, + 0.013265066780149937, + 0.015894891694188118, + 0.006566724274307489, + 0.007591461297124624, + 0.027402399107813835, + 0.008438742719590664, + 0.0248742513358593, + 0.010176200419664383, + 0.021971967071294785, + 0.007363433949649334, + 0.006591022480279207, + 0.007103225216269493, + 0.025918224826455116, + 0.016692230477929115, + 0.06772158294916153, + 0.011774413287639618, + 0.0049167596735060215, + 0.006465380545705557, + 0.0038995197974145412, + 0.004793954081833363, + 0.012969011440873146, + 0.007165398448705673, + 0.008597709238529205, + 0.003526669694110751, + 0.008203345350921154, + 0.010242918506264687, + 0.006287966854870319, + 0.006456280127167702, + 0.017272355034947395, + 0.01420537382364273, + 0.02026975154876709, + 0.007366515230387449, + 0.013829694129526615, + 0.00426447344943881, + 0.005692261271178722, + 0.003061063587665558, + 0.016903338953852654, + 0.005564313381910324, + 0.00797342974692583, + 0.008128623478114605, + 0.005464387591928244, + 0.0054712737910449505, + 0.0036330767907202244, + 0.00540952431038022, + 0.005236923694610596, + 0.008903839625418186, + 0.008501658216118813, + 0.0041789766401052475, + 0.005876118317246437, + 0.006037815008312464, + 0.0049425880424678326, + 0.005913200788199902, + 0.011795450001955032, + 0.008229859173297882, + 0.010766774415969849, + 0.007227735593914986, + 0.014660598710179329, + 0.016523869708180428, + 0.007026694715023041, + 0.010076925158500671, + 0.02810910902917385, + 0.01750296540558338, + 0.029620546847581863, + 0.014560249634087086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.006747674662619829 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.012542770244181156 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542770244181156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542770244181156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15518923103809357, + 0.029874611645936966, + 0.06915801763534546, + 0.24533335864543915, + 0.038344115018844604, + 0.14131397008895874, + 0.04185165837407112, + 0.030259259045124054, + 0.11429475992918015, + 0.1343809962272644 + ], + "credit_map": [ + 0.017259616404771805, + 0.012624604627490044, + 0.012263832613825798, + 0.012524677440524101, + 0.022409914061427116, + 0.010931297205388546, + 0.04084322601556778, + 0.026332072913646698, + 0.0028857477009296417, + 0.004418330267071724, + 0.0037070773541927338, + 0.0038048941642045975, + 0.002870656084269285, + 0.002998233074322343, + 0.005227999296039343, + 0.003961673937737942, + 0.008142581209540367, + 0.011440024711191654, + 0.00467512896284461, + 0.00573571166023612, + 0.007886446081101894, + 0.0048094578087329865, + 0.02076653018593788, + 0.005702135618776083, + 0.025771774351596832, + 0.010093793272972107, + 0.010919344611465931, + 0.010667218826711178, + 0.040204789489507675, + 0.0270131453871727, + 0.10164286196231842, + 0.019020438194274902, + 0.0038469505961984396, + 0.0036237703170627356, + 0.00247741281054914, + 0.002776568057015538, + 0.006534032057970762, + 0.00433149142190814, + 0.012195783667266369, + 0.0025581070221960545, + 0.020471228286623955, + 0.01195064838975668, + 0.00904831849038601, + 0.0102988975122571, + 0.017710763961076736, + 0.03682657703757286, + 0.023814594373106956, + 0.011192947626113892, + 0.00700012082234025, + 0.00915513001382351, + 0.0034990168642252684, + 0.002966927830129862, + 0.006494278088212013, + 0.0031486849766224623, + 0.004509628750383854, + 0.005077873356640339, + 0.003981465008109808, + 0.005749983247369528, + 0.0015933268005028367, + 0.003019105177372694, + 0.003674115752801299, + 0.003964285831898451, + 0.005251593422144651, + 0.003025384619832039, + 0.010801508091390133, + 0.01044981088489294, + 0.013653391040861607, + 0.010706565342843533, + 0.020773213356733322, + 0.01712089031934738, + 0.018739880993962288, + 0.01204950176179409, + 0.011908956803381443, + 0.009508873336017132, + 0.004948113579303026, + 0.009001573547720909, + 0.04065549373626709, + 0.011669868603348732, + 0.03056934103369713, + 0.01611877977848053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.012542770244181156 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + 0.16811935603618622, + 0.06913937628269196, + 0.10034824907779694, + 0.11099297553300858, + 0.04771408066153526, + 0.22274766862392426, + 0.057686951011419296, + 0.019677497446537018, + 0.13253428041934967, + 0.07103955745697021 + ], + "credit_map": [ + 0.017060913145542145, + 0.016928138211369514, + 0.012825420126318932, + 0.01748828776180744, + 0.029286760836839676, + 0.011465742252767086, + 0.0408342182636261, + 0.022229880094528198, + 0.006448956672102213, + 0.007337109185755253, + 0.005061200819909573, + 0.0052200681529939175, + 0.009538335725665092, + 0.016028674319386482, + 0.009274918586015701, + 0.010230114683508873, + 0.015910010784864426, + 0.013326632790267467, + 0.009147350676357746, + 0.00791931338608265, + 0.018128447234630585, + 0.01254421565681696, + 0.013356534764170647, + 0.010015740059316158, + 0.01273114513605833, + 0.005178527440875769, + 0.005427144002169371, + 0.005149281583726406, + 0.017131295055150986, + 0.011885825544595718, + 0.04496614634990692, + 0.008523611351847649, + 0.006856036838144064, + 0.005106254946440458, + 0.004731764085590839, + 0.003641820512712002, + 0.007543737534433603, + 0.007820972241461277, + 0.006517750211060047, + 0.005495745688676834, + 0.03046909160912037, + 0.03011269122362137, + 0.017619663849473, + 0.021050579845905304, + 0.034407712519168854, + 0.02319963648915291, + 0.047262560576200485, + 0.018625745549798012, + 0.005801072344183922, + 0.019507354125380516, + 0.003275851020589471, + 0.004395109601318836, + 0.007419866975396872, + 0.0042961351573467255, + 0.007937614805996418, + 0.005053945351392031, + 0.002129405504092574, + 0.0029103467240929604, + 0.0014445980777963996, + 0.0021113657858222723, + 0.0018370089819654822, + 0.004119016695767641, + 0.003295441158115864, + 0.0018303130054846406, + 0.018509425222873688, + 0.015247175469994545, + 0.009843503125011921, + 0.014793976210057735, + 0.02168498933315277, + 0.016113152727484703, + 0.022670812904834747, + 0.013671253807842731, + 0.006081685423851013, + 0.005558269564062357, + 0.0029034498147666454, + 0.005910134874284267, + 0.017214490100741386, + 0.006151390261948109, + 0.0192823875695467, + 0.007937749847769737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.005944667384028435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.07895991951227188 + ], + "opposite_logprob_tokens": [ + -2.5789599418640137 + ], + "logprob_reward": -0.07895991951227188, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895991951227188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.178794726729393, + 0.13782356679439545, + 0.12483150511980057, + 0.1475769579410553, + 0.07553070783615112, + 0.13978137075901031, + 0.06035997346043587, + 0.03859876096248627, + 0.024088677018880844, + 0.07261376827955246 + ], + "credit_map": [ + 0.016988417133688927, + 0.017641089856624603, + 0.01059713400900364, + 0.015135169960558414, + 0.032053932547569275, + 0.011480842716991901, + 0.052742838859558105, + 0.022155309095978737, + 0.013157390058040619, + 0.012196186929941177, + 0.009333744645118713, + 0.011021139100193977, + 0.01610284484922886, + 0.03153102472424507, + 0.018819181248545647, + 0.025662047788500786, + 0.014750314876437187, + 0.014037356711924076, + 0.009202091954648495, + 0.008742949925363064, + 0.01756477914750576, + 0.007775891572237015, + 0.04067213088274002, + 0.01208597794175148, + 0.031302984803915024, + 0.007848786190152168, + 0.007404845207929611, + 0.005812889896333218, + 0.020049240440130234, + 0.012829151935875416, + 0.05446898937225342, + 0.007860067300498486, + 0.009233719669282436, + 0.008664139546453953, + 0.006185231730341911, + 0.006152177695184946, + 0.012803037650883198, + 0.009278633631765842, + 0.01789049617946148, + 0.005323274526745081, + 0.03254973143339157, + 0.006863063666969538, + 0.006788062863051891, + 0.005034772213548422, + 0.014597323723137379, + 0.053039271384477615, + 0.013430120423436165, + 0.00747902225703001, + 0.0106127317994833, + 0.011404985561966896, + 0.005448996089398861, + 0.0034174819011241198, + 0.012786177918314934, + 0.0054206522181630135, + 0.007726997137069702, + 0.0035419513005763292, + 0.0037741565611213446, + 0.007866190746426582, + 0.002739895833656192, + 0.003932818304747343, + 0.0038400853518396616, + 0.006286310963332653, + 0.006847120355814695, + 0.003312187734991312, + 0.0035625447053462267, + 0.00385094340890646, + 0.002642047358676791, + 0.0021404337603598833, + 0.0028565120883286, + 0.003388331737369299, + 0.0032948304433375597, + 0.0023530335165560246, + 0.006290959194302559, + 0.00492358673363924, + 0.0025772168301045895, + 0.0033635548315942287, + 0.02829344943165779, + 0.0080918800085783, + 0.007742059882730246, + 0.011331056244671345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.07895991951227188 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful 360-degree pan shot, from a slightly bird's-eye view, slowly unveils a sumptuous Sichuan hot pot feast. In the center of the frame, an exquisitely carved copper pot stands atop a dark, round wooden table. Its body gleams with a warm, patina of bronze, intricately carved with vintage patterns of auspicious clouds and dragons. Within, a vibrant, translucent red oil broth simmers, bubbling with countless plump dried chilies and Sichuan peppercorns. Wisps of white steam rise, blurring the background and adding a dreamy, soft-focus effect. Circling the hot pot, plates of premium, thinly sliced, marbling beef spread out, its red and white marbled veins clearly visible in the light, revealing its exquisite tenderness. Next to them, juicy green spinach, neatly arranged white jade mushrooms, and tender lotus root slices cut into petals, all arranged like stars surrounding the moon. A Chinese palace lantern suspended above the table casts a warm, soft yellow light, illuminating the splashes of oil from the broth, making them shimmer with golden glints and coating each ingredient with an exceptionally alluring sheen. The entire scene, dominated by rich reds and warm yellows, creates an atmosphere of tranquility and abundance, reminiscent of a classical poetic painting, imbued with the warmth and abundant joy of everyday life.", + "seed": 31511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_4677_seed_31511_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": -0.042768625915050505, + "vlm_logprob": -0.042768625915050505, + "vlm_yes_no_margin": 3.499999976158142, + "vlm_token_credit": -0.042768625915050505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.0274268239736557 + ], + "opposite_logprob_tokens": [ + -3.7774267196655273 + ], + "logprob_reward": -0.0274268239736557, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0274268239736557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09347449988126755, + 0.022724168375134468, + 0.07748785614967346, + 0.17270395159721375, + 0.16608262062072754, + 0.1803583800792694, + 0.04378877952694893, + 0.15304677188396454, + 0.036861687898635864, + 0.05347128584980965 + ], + "credit_map": [ + 0.007935743778944016, + 0.015097317285835743, + 0.00817504245787859, + 0.009740255773067474, + 0.013095196336507797, + 0.014258528128266335, + 0.01394426729530096, + 0.011228138580918312, + 0.002086173975840211, + 0.002278108848258853, + 0.004396241158246994, + 0.0025111595168709755, + 0.0033082966692745686, + 0.0023795512970536947, + 0.0034790285862982273, + 0.0022856066934764385, + 0.005828544031828642, + 0.008909452706575394, + 0.0032805316150188446, + 0.002991972491145134, + 0.03312774747610092, + 0.012281732633709908, + 0.007361546158790588, + 0.003706333227455616, + 0.013997999019920826, + 0.019606657326221466, + 0.021009426563978195, + 0.011407787911593914, + 0.038249820470809937, + 0.020724430680274963, + 0.03246121108531952, + 0.01524660550057888, + 0.020826926454901695, + 0.020052462816238403, + 0.02069806680083275, + 0.013019229285418987, + 0.028284601867198944, + 0.014372753910720348, + 0.024762731045484543, + 0.02406584843993187, + 0.03608029708266258, + 0.012713797390460968, + 0.06619784981012344, + 0.014067737385630608, + 0.019218875095248222, + 0.01440854649990797, + 0.009647092781960964, + 0.008024181239306927, + 0.003857847535982728, + 0.004252247978001833, + 0.006372688803821802, + 0.00630447780713439, + 0.007980478927493095, + 0.004104372579604387, + 0.0076957340352237225, + 0.003220930928364396, + 0.01073791179805994, + 0.010561643168330193, + 0.04557612165808678, + 0.009544669650495052, + 0.016141539439558983, + 0.007983664982020855, + 0.04266615957021713, + 0.00983505044132471, + 0.0030589131638407707, + 0.005306025501340628, + 0.003907680511474609, + 0.0030008729081600904, + 0.006087697576731443, + 0.002692609094083309, + 0.010104998014867306, + 0.0027028946205973625, + 0.004124144557863474, + 0.006617874838411808, + 0.008874546736478806, + 0.0047821770422160625, + 0.008829372934997082, + 0.007342213299125433, + 0.008227723650634289, + 0.004673230927437544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.0274268239736557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.01820380985736847 + ], + "opposite_logprob_tokens": [ + -4.0182037353515625 + ], + "logprob_reward": -0.01820380985736847, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820380985736847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13986748456954956, + 0.039662592113018036, + 0.17181098461151123, + 0.10348796099424362, + 0.3151625990867615, + 0.03392906114459038, + 0.03757151588797569, + 0.079988993704319, + 0.04361840710043907, + 0.03490036725997925 + ], + "credit_map": [ + 0.012453497387468815, + 0.019554903730750084, + 0.00948767177760601, + 0.01565156690776348, + 0.0191888939589262, + 0.013104149140417576, + 0.03678568825125694, + 0.013641123659908772, + 0.0034024056512862444, + 0.005673683248460293, + 0.00582262733951211, + 0.004880147986114025, + 0.005655790213495493, + 0.004514549393206835, + 0.005521236453205347, + 0.004192157182842493, + 0.01641525886952877, + 0.021225955337285995, + 0.007004142738878727, + 0.00479920906946063, + 0.07610786706209183, + 0.02136935479938984, + 0.017616739496588707, + 0.007272472605109215, + 0.008896737359464169, + 0.01207740418612957, + 0.006196310743689537, + 0.005901230499148369, + 0.03334612771868706, + 0.011595242656767368, + 0.009197845123708248, + 0.016277072951197624, + 0.032304197549819946, + 0.05161181837320328, + 0.03724918141961098, + 0.02414880506694317, + 0.05032898113131523, + 0.039148274809122086, + 0.04345938190817833, + 0.03691194951534271, + 0.004534328822046518, + 0.003943289164453745, + 0.007251905743032694, + 0.0033441486302763224, + 0.005247886758297682, + 0.0027413801290094852, + 0.004747252445667982, + 0.002118869684636593, + 0.0027918608393520117, + 0.0033005152363330126, + 0.004601361230015755, + 0.0036956430412828922, + 0.00685892766341567, + 0.004190114792436361, + 0.009358802810311317, + 0.0027742930687963963, + 0.006016128230839968, + 0.00492287240922451, + 0.024238426238298416, + 0.005188646726310253, + 0.007993219420313835, + 0.0032826135866343975, + 0.02313382178544998, + 0.005213264375925064, + 0.005190613679587841, + 0.0053331987001001835, + 0.005987039301544428, + 0.003568546148017049, + 0.005999511573463678, + 0.0047917538322508335, + 0.009938397444784641, + 0.0028093508444726467, + 0.004148613195866346, + 0.0036166426725685596, + 0.004656062927097082, + 0.003403368638828397, + 0.004738382529467344, + 0.005305428057909012, + 0.005305714905261993, + 0.003726158756762743 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.01820380985736847 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.02058802917599678 + ], + "opposite_logprob_tokens": [ + -3.895587921142578 + ], + "logprob_reward": -0.02058802917599678, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058802917599678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09592462331056595, + 0.048080943524837494, + 0.08905535936355591, + 0.04292060807347298, + 0.5488889217376709, + 0.03449971601366997, + 0.024841928854584694, + 0.047043923288583755, + 0.013998954556882381, + 0.054745081812143326 + ], + "credit_map": [ + 0.008629404939711094, + 0.014219893142580986, + 0.00751129537820816, + 0.01102950144559145, + 0.012391588650643826, + 0.009795996360480785, + 0.022742871195077896, + 0.009604071266949177, + 0.004062075167894363, + 0.006397200282663107, + 0.005488402210175991, + 0.006081418599933386, + 0.006682104896754026, + 0.005433894228190184, + 0.007950365543365479, + 0.0059854816645383835, + 0.009421270340681076, + 0.011160824447870255, + 0.004167275968939066, + 0.003070294391363859, + 0.037420064210891724, + 0.008593359962105751, + 0.010600630193948746, + 0.004621643107384443, + 0.004850026685744524, + 0.004421223420649767, + 0.0048172022216022015, + 0.0032025498803704977, + 0.007237710524350405, + 0.00474851718172431, + 0.010015686973929405, + 0.003627693746238947, + 0.07256027311086655, + 0.06446750462055206, + 0.06499719619750977, + 0.041655030101537704, + 0.0723021999001503, + 0.056396983563899994, + 0.08082829415798187, + 0.09568140655755997, + 0.004529919009655714, + 0.00398815143853426, + 0.00788711104542017, + 0.0031148262787610292, + 0.005718605127185583, + 0.002677575219422579, + 0.004593431483954191, + 0.0019901003688573837, + 0.003242210950702429, + 0.0025145101826637983, + 0.004958658013492823, + 0.0028000075835734606, + 0.0035599733237177134, + 0.0024308129213750362, + 0.0033121537417173386, + 0.002023601671680808, + 0.006016595754772425, + 0.0040105190128088, + 0.009847485460340977, + 0.003439942840486765, + 0.005682194605469704, + 0.0024601356126368046, + 0.012212145142257214, + 0.0033749029971659184, + 0.0015050420770421624, + 0.0017443107208237052, + 0.002016837242990732, + 0.0014004787662997842, + 0.001886253710836172, + 0.0015740785747766495, + 0.0028454286511987448, + 0.001026524812914431, + 0.004707958083599806, + 0.006972855422645807, + 0.006135478615760803, + 0.005159706808626652, + 0.00680196238681674, + 0.008910998702049255, + 0.008304438553750515, + 0.007751681376248598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.02058802917599678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.02058732882142067 + ], + "opposite_logprob_tokens": [ + -3.89558744430542 + ], + "logprob_reward": -0.02058732882142067, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058732882142067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2655350863933563, + 0.04025290906429291, + 0.06435586512088776, + 0.09399063885211945, + 0.19820229709148407, + 0.05203493684530258, + 0.13526204228401184, + 0.07382160425186157, + 0.04715337976813316, + 0.02939119003713131 + ], + "credit_map": [ + 0.02213379740715027, + 0.038880832493305206, + 0.022411642596125603, + 0.028822122141718864, + 0.03994842246174812, + 0.025149421766400337, + 0.06147105246782303, + 0.02671780250966549, + 0.00335111771710217, + 0.0031177119817584753, + 0.007124810013920069, + 0.003437067149206996, + 0.006778437178581953, + 0.002820358145982027, + 0.010482118465006351, + 0.0031412874814122915, + 0.00732421176508069, + 0.007311374880373478, + 0.004909386392682791, + 0.00394735811278224, + 0.022974994033575058, + 0.006850320380181074, + 0.006711764261126518, + 0.00432645482942462, + 0.007129472214728594, + 0.009157990105450153, + 0.009048597887158394, + 0.007319600321352482, + 0.011773478239774704, + 0.008505682460963726, + 0.022596439346671104, + 0.018459385260939598, + 0.022615736350417137, + 0.026791146025061607, + 0.026255346834659576, + 0.01571633294224739, + 0.030068840831518173, + 0.01966409757733345, + 0.030071819201111794, + 0.027018969878554344, + 0.007373750675469637, + 0.004312167409807444, + 0.017521610483527184, + 0.00467303954064846, + 0.006981171667575836, + 0.004511757288128138, + 0.0038042122032493353, + 0.0028572326991707087, + 0.01890101097524166, + 0.007957159541547298, + 0.025653600692749023, + 0.010331696830689907, + 0.013141418807208538, + 0.007959958165884018, + 0.039452988654375076, + 0.011864208616316319, + 0.00821715872734785, + 0.005089758429676294, + 0.0203477144241333, + 0.0062479134649038315, + 0.0064417156390845776, + 0.0029556434601545334, + 0.019468095153570175, + 0.005053611937910318, + 0.007484192959964275, + 0.0037151838187128305, + 0.011128165759146214, + 0.004951027221977711, + 0.005831026006489992, + 0.003927646204829216, + 0.0069988444447517395, + 0.0031172949820756912, + 0.00442096870392561, + 0.00330067565664649, + 0.004221160896122456, + 0.003243526443839073, + 0.003693818813189864, + 0.004137847572565079, + 0.003856782102957368, + 0.0025164098478853703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.02058732882142067 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.1270371377468109 + ], + "opposite_logprob_tokens": [ + -2.1270370483398438 + ], + "logprob_reward": -0.1270371377468109, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270371377468109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14126409590244293, + 0.0283732358366251, + 0.05081610381603241, + 0.180804044008255, + 0.06692232191562653, + 0.16255110502243042, + 0.04742535948753357, + 0.22971373796463013, + 0.06773851811885834, + 0.02439151704311371 + ], + "credit_map": [ + 0.011903722770512104, + 0.020078234374523163, + 0.009988733567297459, + 0.016042808070778847, + 0.018290184438228607, + 0.015242239460349083, + 0.034305375069379807, + 0.015412784181535244, + 0.0037063495256006718, + 0.004369593691080809, + 0.004777819849550724, + 0.003341068048030138, + 0.00391780398786068, + 0.002635158598423004, + 0.0033847708255052567, + 0.0022406692150980234, + 0.004869877360761166, + 0.00613769143819809, + 0.0034214770421385765, + 0.002162312623113394, + 0.020477402955293655, + 0.0064000519923865795, + 0.004570785444229841, + 0.0027764979749917984, + 0.014321599155664444, + 0.02329847775399685, + 0.01048543956130743, + 0.009623050689697266, + 0.06696765124797821, + 0.011700311675667763, + 0.017026683315634727, + 0.027380822226405144, + 0.006408875808119774, + 0.008630703203380108, + 0.008575583808124065, + 0.0054200077429413795, + 0.010741528123617172, + 0.008984477259218693, + 0.00985648762434721, + 0.008304652757942677, + 0.032481640577316284, + 0.008904533460736275, + 0.06676740199327469, + 0.012407837435603142, + 0.014890000224113464, + 0.013488810509443283, + 0.0069874003529548645, + 0.006623462773859501, + 0.005491666961461306, + 0.004482667427510023, + 0.006225096993148327, + 0.006727340631186962, + 0.0058568804524838924, + 0.004188219550997019, + 0.009361648932099342, + 0.00509183295071125, + 0.014204462990164757, + 0.013131988234817982, + 0.06415683031082153, + 0.012763961218297482, + 0.02323758602142334, + 0.0067162285558879375, + 0.07911038398742676, + 0.016392288729548454, + 0.00756118493154645, + 0.008392700925469398, + 0.009539862163364887, + 0.006401797290891409, + 0.01146672759205103, + 0.004254613537341356, + 0.015474230982363224, + 0.004647392313927412, + 0.003608883125707507, + 0.0026987912133336067, + 0.00243581784889102, + 0.0021679021883755922, + 0.003271405817940831, + 0.0033971744123846292, + 0.003480972722172737, + 0.0033305666875094175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.1270371377468109 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful 360-degree pan shot, from a slightly bird's-eye view, slowly unveils a sumptuous Sichuan hot pot feast. In the center of the frame, an exquisitely carved copper pot stands atop a dark, round wooden table. Its body gleams with a warm, patina of bronze, intricately carved with vintage patterns of auspicious clouds and dragons. Within, a vibrant, translucent red oil broth simmers, bubbling with countless plump dried chilies and Sichuan peppercorns. Wisps of white steam rise, blurring the background and adding a dreamy, soft-focus effect. Circling the hot pot, plates of premium, thinly sliced, marbling beef spread out, its red and white marbled veins clearly visible in the light, revealing its exquisite tenderness. Next to them, juicy green spinach, neatly arranged white jade mushrooms, and tender lotus root slices cut into petals, all arranged like stars surrounding the moon. A Chinese palace lantern suspended above the table casts a warm, soft yellow light, illuminating the splashes of oil from the broth, making them shimmer with golden glints and coating each ingredient with an exceptionally alluring sheen. The entire scene, dominated by rich reds and warm yellows, creates an atmosphere of tranquility and abundance, reminiscent of a classical poetic painting, imbued with the warmth and abundant joy of everyday life.", + "seed": 31512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_4677_seed_31512_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": -0.048724221903830765, + "vlm_logprob": -0.048724221903830765, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.048724221903830765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.036926016211509705 + ], + "opposite_logprob_tokens": [ + -3.411926031112671 + ], + "logprob_reward": -0.036926016211509705, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.036926016211509705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07860088348388672, + 0.16002053022384644, + 0.0787830501794815, + 0.028044365346431732, + 0.08476796746253967, + 0.02985009364783764, + 0.02501494251191616, + 0.05350106582045555, + 0.42157918214797974, + 0.03983789682388306 + ], + "credit_map": [ + 0.022659145295619965, + 0.006278459448367357, + 0.006498228292912245, + 0.004333553835749626, + 0.007929909974336624, + 0.01052253320813179, + 0.008406161330640316, + 0.011972892098128796, + 0.04779598489403725, + 0.007637252099812031, + 0.0062324851751327515, + 0.009919744916260242, + 0.007413850631564856, + 0.007120897527784109, + 0.0676020085811615, + 0.00629831850528717, + 0.012884547002613544, + 0.003167164046317339, + 0.005003242287784815, + 0.010213082656264305, + 0.0034124397207051516, + 0.009069214574992657, + 0.02409263700246811, + 0.01094071939587593, + 0.0037532025016844273, + 0.0027931167278438807, + 0.004563538823276758, + 0.0029809828847646713, + 0.003507356159389019, + 0.0033165160566568375, + 0.003987450618296862, + 0.0031422025058418512, + 0.003690341953188181, + 0.004059006460011005, + 0.004156562965363264, + 0.009406467899680138, + 0.005502998363226652, + 0.010472634807229042, + 0.03177208453416824, + 0.015707865357398987, + 0.00546020595356822, + 0.003271891037002206, + 0.002202797681093216, + 0.0070425234735012054, + 0.0031321654096245766, + 0.0024528938811272383, + 0.0036462198477238417, + 0.002641398459672928, + 0.0020774409640580416, + 0.0029885501135140657, + 0.0031290275510400534, + 0.004239985719323158, + 0.0025466091465204954, + 0.0034141237847507, + 0.0034397472627460957, + 0.0031794579699635506, + 0.005055706948041916, + 0.004753407556563616, + 0.006808414123952389, + 0.006211215630173683, + 0.004678046330809593, + 0.008337320759892464, + 0.00715611083433032, + 0.01050084549933672, + 0.03138550743460655, + 0.012951040640473366, + 0.15354004502296448, + 0.023982200771570206, + 0.01315916609019041, + 0.10693781822919846, + 0.029897058382630348, + 0.04972635954618454, + 0.0053825294598937035, + 0.004337346646934748, + 0.003797424491494894, + 0.006994928698986769, + 0.003716760315001011, + 0.004775402136147022, + 0.005538430530577898, + 0.005295077338814735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.036926016211509705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.003204094711691141 + ], + "opposite_logprob_tokens": [ + -5.753203868865967 + ], + "logprob_reward": -0.003204094711691141, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003204094711691141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13628536462783813, + 0.13869862258434296, + 0.2603711783885956, + 0.03961116448044777, + 0.10565485060214996, + 0.06691785156726837, + 0.026937447488307953, + 0.03535164147615433, + 0.089024618268013, + 0.10114727169275284 + ], + "credit_map": [ + 0.02820991352200508, + 0.013459338806569576, + 0.011526155285537243, + 0.009250611066818237, + 0.016176650300621986, + 0.015667976811528206, + 0.02216806448996067, + 0.01982663944363594, + 0.029555946588516235, + 0.006552043836563826, + 0.005081369075924158, + 0.011320557445287704, + 0.00747328158468008, + 0.00840443093329668, + 0.05935844033956528, + 0.010952543467283249, + 0.024801457300782204, + 0.008593956008553505, + 0.014343930408358574, + 0.03563229739665985, + 0.013198391534388065, + 0.021430518478155136, + 0.11086355894804001, + 0.031507059931755066, + 0.0040939911268651485, + 0.0040075392462313175, + 0.003773084841668606, + 0.005506246816366911, + 0.004379777703434229, + 0.00565393827855587, + 0.008379741571843624, + 0.003816843731328845, + 0.006583304610103369, + 0.006257028318941593, + 0.004714393988251686, + 0.0072937216609716415, + 0.008072038181126118, + 0.04189744591712952, + 0.015964450314641, + 0.014872469939291477, + 0.01031448133289814, + 0.006603697314858437, + 0.005757986102253199, + 0.02323313243687153, + 0.0055403197184205055, + 0.004190286621451378, + 0.007877795957028866, + 0.0034001495223492384, + 0.0027525676414370537, + 0.003926761448383331, + 0.003911933861672878, + 0.004072191193699837, + 0.003798369551077485, + 0.0032740687020123005, + 0.003045515390112996, + 0.002156041096895933, + 0.003979226108640432, + 0.00474432110786438, + 0.0027855620719492435, + 0.005845940671861172, + 0.004273338709026575, + 0.004348986782133579, + 0.004239479545503855, + 0.005134785082191229, + 0.010525175370275974, + 0.009266417473554611, + 0.008856769651174545, + 0.0104643190279603, + 0.006292429752647877, + 0.015400977805256844, + 0.012608972378075123, + 0.01560955960303545, + 0.015718715265393257, + 0.009744081646203995, + 0.011153827421367168, + 0.016828618943691254, + 0.012216516770422459, + 0.012400845997035503, + 0.014685886912047863, + 0.008398777805268764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.003204094711691141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.011074741370975971 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074741370975971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074741370975971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19285868108272552, + 0.15797515213489532, + 0.08256671577692032, + 0.02257176861166954, + 0.037178680300712585, + 0.01481410302221775, + 0.0098288394510746, + 0.008501200936734676, + 0.43310609459877014, + 0.04059877619147301 + ], + "credit_map": [ + 0.06334622949361801, + 0.01394509244710207, + 0.0135126281529665, + 0.008819060400128365, + 0.01334426924586296, + 0.014180154539644718, + 0.03307417035102844, + 0.03263707831501961, + 0.04491132125258446, + 0.007496309000998735, + 0.0058312928304076195, + 0.010341892018914223, + 0.007397548761218786, + 0.006523303221911192, + 0.06808196753263474, + 0.007391529623419046, + 0.0077044484205543995, + 0.0028444614727050066, + 0.00431663217023015, + 0.011506564915180206, + 0.00419200723990798, + 0.006937738507986069, + 0.034475285559892654, + 0.010589581914246082, + 0.0018107348587363958, + 0.0034979775082319975, + 0.0019225487485527992, + 0.0032680791337043047, + 0.0036061815917491913, + 0.0025770864449441433, + 0.003076074877753854, + 0.0028130831196904182, + 0.0028713068459182978, + 0.0031288405880331993, + 0.003537760116159916, + 0.004511976148933172, + 0.0037542758509516716, + 0.01069906447082758, + 0.004333921242505312, + 0.004341534338891506, + 0.004260744899511337, + 0.001231681671924889, + 0.001165623078122735, + 0.0033759898506104946, + 0.0014145723544061184, + 0.0010911888675764203, + 0.001095456420443952, + 0.0011788442498072982, + 0.0007419136818498373, + 0.0009123753407038748, + 0.0012625714298337698, + 0.0013902231585234404, + 0.0008306782110594213, + 0.0018197967438027263, + 0.0013768919743597507, + 0.001494389260187745, + 0.0010642941342666745, + 0.0007689602207392454, + 0.0011167953489348292, + 0.0011100330157205462, + 0.0007647696183994412, + 0.0010099193314090371, + 0.0009306595311500132, + 0.001735770027153194, + 0.03552067279815674, + 0.016927743330597878, + 0.16449102759361267, + 0.02333991229534149, + 0.013547062873840332, + 0.10016383230686188, + 0.029529955238103867, + 0.04958592355251312, + 0.006171839311718941, + 0.004052331205457449, + 0.004662638530135155, + 0.005412694066762924, + 0.004706209991127253, + 0.006947665475308895, + 0.003814229741692543, + 0.004831165540963411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.011074741370975971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.012532881461083889 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532881461083889, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532881461083889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09218260645866394, + 0.3744712471961975, + 0.06831058114767075, + 0.014370164833962917, + 0.037688758224248886, + 0.014840587042272091, + 0.015951192006468773, + 0.0130494749173522, + 0.3046969175338745, + 0.06443845480680466 + ], + "credit_map": [ + 0.02264680527150631, + 0.006236804183572531, + 0.007964993827044964, + 0.006297409534454346, + 0.005894347093999386, + 0.011831650510430336, + 0.010045836679637432, + 0.021264759823679924, + 0.12898744642734528, + 0.018013013526797295, + 0.014604954048991203, + 0.01647784374654293, + 0.015198413282632828, + 0.014849048107862473, + 0.14554955065250397, + 0.020791005343198776, + 0.007537295576184988, + 0.0026822853833436966, + 0.005742026027292013, + 0.007165033835917711, + 0.003823285223916173, + 0.007718779146671295, + 0.027248863130807877, + 0.0063930172473192215, + 0.0015622212085872889, + 0.0014970123302191496, + 0.0011231002863496542, + 0.0017862700624391437, + 0.0020682814065366983, + 0.002506938064470887, + 0.0022466524969786406, + 0.0015796881634742022, + 0.0026497971266508102, + 0.002246409421786666, + 0.004402328748255968, + 0.004882048349827528, + 0.003237426048144698, + 0.0017908575246110559, + 0.012594004161655903, + 0.0058858864940702915, + 0.0024356406647711992, + 0.0013748897472396493, + 0.0012726961867883801, + 0.003990182653069496, + 0.0012377548264339566, + 0.0011458441149443388, + 0.0018514955881983042, + 0.0015320831444114447, + 0.001969732344150543, + 0.00219130446203053, + 0.0011449402663856745, + 0.002402503276243806, + 0.0023588419426232576, + 0.002462651114910841, + 0.002313237404450774, + 0.0011079811956733465, + 0.0011077726958319545, + 0.0014792438596487045, + 0.0016092201694846153, + 0.0017897575162351131, + 0.0016551667358726263, + 0.0016196861397475004, + 0.0016051473794505, + 0.002183479955419898, + 0.025315189734101295, + 0.01187690906226635, + 0.11671479046344757, + 0.016891658306121826, + 0.00975228101015091, + 0.0655505433678627, + 0.01982574537396431, + 0.03876980394124985, + 0.009402542375028133, + 0.005535251460969448, + 0.009359829127788544, + 0.006225203163921833, + 0.007790052331984043, + 0.01101401075720787, + 0.0056366268545389175, + 0.00947494339197874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.012532881461083889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.17988337576389313 + ], + "opposite_logprob_tokens": [ + -1.804883360862732 + ], + "logprob_reward": -0.17988337576389313, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988337576389313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712529718875885, + 0.16298864781856537, + 0.07841993123292923, + 0.024974027648568153, + 0.10267113149166107, + 0.08917512744665146, + 0.09351535141468048, + 0.09690956771373749, + 0.20371940732002258, + 0.07637381553649902 + ], + "credit_map": [ + 0.008969597518444061, + 0.0074598114006221294, + 0.007877304218709469, + 0.007405800744891167, + 0.0078754723072052, + 0.014268829487264156, + 0.008913171477615833, + 0.008482989855110645, + 0.05354719236493111, + 0.008080213330686092, + 0.0066412328742444515, + 0.008219360373914242, + 0.007168352138251066, + 0.0062156543135643005, + 0.06653600186109543, + 0.00658064242452383, + 0.026282908394932747, + 0.004819671623408794, + 0.004475422669202089, + 0.008926577866077423, + 0.0042160083539783955, + 0.010486194863915443, + 0.007506257388740778, + 0.011706892400979996, + 0.0032004406675696373, + 0.002670341869816184, + 0.0025432920083403587, + 0.003957475069910288, + 0.0025136731564998627, + 0.0042871409095823765, + 0.003060217248275876, + 0.002741446951404214, + 0.004511553328484297, + 0.005451667122542858, + 0.005104591138660908, + 0.011109673418104649, + 0.00633555743843317, + 0.016620442271232605, + 0.03757145628333092, + 0.015966197475790977, + 0.010583057068288326, + 0.009394519031047821, + 0.0054834396578371525, + 0.02216113917529583, + 0.012488134205341339, + 0.007945913821458817, + 0.012895121239125729, + 0.008223798125982285, + 0.008270157501101494, + 0.011240411549806595, + 0.009039092808961868, + 0.012323005124926567, + 0.010339096188545227, + 0.01775156520307064, + 0.015553437173366547, + 0.008998582139611244, + 0.01397060789167881, + 0.011796753853559494, + 0.007929510436952114, + 0.013538587838411331, + 0.010350069962441921, + 0.014845632016658783, + 0.013690073043107986, + 0.0107883270829916, + 0.01682138256728649, + 0.009804151952266693, + 0.07707688957452774, + 0.011331775225698948, + 0.0067326161079108715, + 0.0433899462223053, + 0.013554850593209267, + 0.025007780641317368, + 0.009050232358276844, + 0.007723080925643444, + 0.009298103861510754, + 0.011004895903170109, + 0.007340733427554369, + 0.014764967374503613, + 0.006304876413196325, + 0.010886919684708118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 4 + ], + "selected_reward": -0.17988337576389313 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow and graceful 360-degree pan shot, from a slightly bird's-eye view, slowly unveils a sumptuous Sichuan hot pot feast. In the center of the frame, an exquisitely carved copper pot stands atop a dark, round wooden table. Its body gleams with a warm, patina of bronze, intricately carved with vintage patterns of auspicious clouds and dragons. Within, a vibrant, translucent red oil broth simmers, bubbling with countless plump dried chilies and Sichuan peppercorns. Wisps of white steam rise, blurring the background and adding a dreamy, soft-focus effect. Circling the hot pot, plates of premium, thinly sliced, marbling beef spread out, its red and white marbled veins clearly visible in the light, revealing its exquisite tenderness. Next to them, juicy green spinach, neatly arranged white jade mushrooms, and tender lotus root slices cut into petals, all arranged like stars surrounding the moon. A Chinese palace lantern suspended above the table casts a warm, soft yellow light, illuminating the splashes of oil from the broth, making them shimmer with golden glints and coating each ingredient with an exceptionally alluring sheen. The entire scene, dominated by rich reds and warm yellows, creates an atmosphere of tranquility and abundance, reminiscent of a classical poetic painting, imbued with the warmth and abundant joy of everyday life.", + "seed": 31513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_0/index_4677_seed_31513_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": -0.05127526530995965, + "vlm_logprob": -0.05127526530995965, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05127526530995965, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.10388629883527756 + ], + "opposite_logprob_tokens": [ + -2.353886365890503 + ], + "logprob_reward": -0.10388629883527756, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10388629883527756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4734700620174408, + 0.048154957592487335, + 0.04437337815761566, + 0.02769572287797928, + 0.03579021245241165, + 0.04898252710700035, + 0.15523895621299744, + 0.06157098338007927, + 0.04116485267877579, + 0.0635584369301796 + ], + "credit_map": [ + 0.03265496343374252, + 0.024642150849103928, + 0.01699495129287243, + 0.016248874366283417, + 0.21222791075706482, + 0.1207924336194992, + 0.03607025370001793, + 0.01383854541927576, + 0.0060797506012022495, + 0.005806790664792061, + 0.005745702888816595, + 0.004479110706597567, + 0.008838275447487831, + 0.004393807612359524, + 0.006719788536429405, + 0.00609173160046339, + 0.005483358167111874, + 0.005115029867738485, + 0.0047414666041731834, + 0.005376425571739674, + 0.005981877911835909, + 0.0038322368636727333, + 0.008320854045450687, + 0.005522129125893116, + 0.0045863911509513855, + 0.0035891986917704344, + 0.0028811246156692505, + 0.0048738871701061726, + 0.0031312103383243084, + 0.002471508691087365, + 0.004347701091319323, + 0.0018147022929042578, + 0.002921532606706023, + 0.0036103632301092148, + 0.004397553391754627, + 0.0036667166277766228, + 0.007143479306250811, + 0.005861599463969469, + 0.004208256024867296, + 0.003980710636824369, + 0.006245179567486048, + 0.0055645969696342945, + 0.008090391755104065, + 0.0043844240717589855, + 0.009355141781270504, + 0.004140072036534548, + 0.0067956699058413506, + 0.004407049622386694, + 0.019041365012526512, + 0.020071545615792274, + 0.023392006754875183, + 0.015345651656389236, + 0.02957734651863575, + 0.016111768782138824, + 0.019074898213148117, + 0.012624365277588367, + 0.007352518383413553, + 0.011523423716425896, + 0.0049797543324530125, + 0.005189405754208565, + 0.008753860369324684, + 0.0067423018626868725, + 0.009919672273099422, + 0.007110049482434988, + 0.004364746622741222, + 0.005244921892881393, + 0.004204500000923872, + 0.004639863967895508, + 0.008291387930512428, + 0.0044920798391103745, + 0.0045221103355288506, + 0.005405242554843426, + 0.014148383401334286, + 0.007534892298281193, + 0.004608706571161747, + 0.006872231140732765, + 0.006259381305426359, + 0.004782949108630419, + 0.015221841633319855, + 0.004130059387534857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.10388629883527756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.0067512271925807 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.0067512271925807, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067512271925807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5106526613235474, + 0.0739934965968132, + 0.028137080371379852, + 0.028299229219555855, + 0.054154008626937866, + 0.033573102205991745, + 0.027558045461773872, + 0.14667531847953796, + 0.011713310144841671, + 0.08524379879236221 + ], + "credit_map": [ + 0.03407934308052063, + 0.024194980040192604, + 0.016896113753318787, + 0.015141258016228676, + 0.230647012591362, + 0.13926240801811218, + 0.03598593920469284, + 0.01444556936621666, + 0.00522199273109436, + 0.004361690487712622, + 0.0037519189063459635, + 0.004449822939932346, + 0.024167465046048164, + 0.022311298176646233, + 0.007115586195141077, + 0.0026137116365134716, + 0.0035172472707927227, + 0.004075041506439447, + 0.0020915812347084284, + 0.004743257537484169, + 0.003925707191228867, + 0.002164642559364438, + 0.005265415646135807, + 0.002354186261072755, + 0.0029298949521034956, + 0.003765000496059656, + 0.0023639744613319635, + 0.005778191145509481, + 0.004788363818079233, + 0.002001530723646283, + 0.004994469694793224, + 0.0016778029967099428, + 0.006588970310986042, + 0.002805328695103526, + 0.004645922686904669, + 0.0029135379008948803, + 0.010778114199638367, + 0.014045077376067638, + 0.006066595204174519, + 0.00631046062335372, + 0.008690671995282173, + 0.0023679062724113464, + 0.0044244020245969296, + 0.0018804817227646708, + 0.002347000641748309, + 0.008459639735519886, + 0.002762060845270753, + 0.0026409358251839876, + 0.0018762116087600589, + 0.003957455512136221, + 0.001696841442026198, + 0.002539264503866434, + 0.0058279079385101795, + 0.005620258394628763, + 0.0037090149708092213, + 0.0023310882970690727, + 0.011819549836218357, + 0.046774815768003464, + 0.010096222162246704, + 0.011731280013918877, + 0.024883152917027473, + 0.00982751790434122, + 0.024694524705410004, + 0.006848248653113842, + 0.0017078492091968656, + 0.0010843273485079408, + 0.001022682641632855, + 0.0016336517874151468, + 0.0021042916923761368, + 0.0012630014680325985, + 0.0019713169895112514, + 0.0009261877276003361, + 0.021768340840935707, + 0.006715922150760889, + 0.004585453309118748, + 0.007129105739295483, + 0.006852788385003805, + 0.004031758289784193, + 0.029542621225118637, + 0.00461780047044158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.0067512271925807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.011077924631536007 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011077924631536007, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077924631536007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6469178795814514, + 0.05317535623908043, + 0.012335547246038914, + 0.009639118798077106, + 0.06531141698360443, + 0.13891218602657318, + 0.025841396301984787, + 0.017013240605592728, + 0.010618239641189575, + 0.020235642790794373 + ], + "credit_map": [ + 0.04182751476764679, + 0.030351463705301285, + 0.02002079412341118, + 0.017733652144670486, + 0.2910139262676239, + 0.1843838393688202, + 0.04373644292354584, + 0.01785021647810936, + 0.004107639659196138, + 0.003331966931000352, + 0.002812966238707304, + 0.0038740970194339752, + 0.014999900944530964, + 0.0164786484092474, + 0.0057388730347156525, + 0.001831264584325254, + 0.001498541096225381, + 0.0015355008654296398, + 0.0009681818191893399, + 0.001944358111359179, + 0.0020640138536691666, + 0.0012148171663284302, + 0.0019713847432285547, + 0.0011387497652322054, + 0.0007884286460466683, + 0.0014360063942149282, + 0.0006838017725385725, + 0.00220506452023983, + 0.0015500764129683375, + 0.0007117594941519201, + 0.0015947014326229692, + 0.000669279950670898, + 0.006403861567378044, + 0.0034364627208560705, + 0.005950392223894596, + 0.0032533679623156786, + 0.01333031989634037, + 0.019833289086818695, + 0.003999629989266396, + 0.009104100987315178, + 0.01546845305711031, + 0.013180664740502834, + 0.030419887974858284, + 0.010671545751392841, + 0.023519275709986687, + 0.012087895534932613, + 0.013038409873843193, + 0.02052605338394642, + 0.0025060821790248156, + 0.0033847226295620203, + 0.0029858637135475874, + 0.0023456907365471125, + 0.005686394404619932, + 0.0035799487959593534, + 0.003248365130275488, + 0.002104330575093627, + 0.0025571677833795547, + 0.004274336155503988, + 0.0017975553637370467, + 0.0014150572242215276, + 0.0021887372713536024, + 0.0015628660330548882, + 0.001974615501239896, + 0.0012429052731022239, + 0.0013834548881277442, + 0.0011382066877558827, + 0.0011193815153092146, + 0.0014747710665687919, + 0.0019072641152888536, + 0.0010460151825100183, + 0.0013910976704210043, + 0.0011580486316233873, + 0.0059533496387302876, + 0.0012402915162965655, + 0.0008821325027383864, + 0.0014711827971041203, + 0.0011709824902936816, + 0.0011361300712451339, + 0.007168683223426342, + 0.0012128883972764015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.011077924631536007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.007626465056091547 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626465056091547, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626465056091547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7774182558059692, + 0.029633035883307457, + 0.014203748665750027, + 0.032290756702423096, + 0.02284015156328678, + 0.04586338624358177, + 0.014936842024326324, + 0.03108743205666542, + 0.018078168854117393, + 0.013648276217281818 + ], + "credit_map": [ + 0.04982880502939224, + 0.03679868206381798, + 0.02486308105289936, + 0.021795814856886864, + 0.3495267629623413, + 0.22168275713920593, + 0.05143655091524124, + 0.021485771983861923, + 0.0035539313685148954, + 0.0029855158645659685, + 0.002031086478382349, + 0.0027194370049983263, + 0.005405488424003124, + 0.0068397242575883865, + 0.0037103972863405943, + 0.002387454267591238, + 0.0022994172759354115, + 0.001329396734945476, + 0.001536339521408081, + 0.0016823721816763282, + 0.002620853018015623, + 0.0017869432922452688, + 0.0013553553726524115, + 0.0015930697554722428, + 0.0072958809323608875, + 0.003281007055193186, + 0.0029363706707954407, + 0.0035943896509706974, + 0.005290647968649864, + 0.0035786102525889874, + 0.00345950061455369, + 0.002854348858818412, + 0.0032056912314146757, + 0.002105870982632041, + 0.0016500739147886634, + 0.002033090451732278, + 0.0043546040542423725, + 0.003768897382542491, + 0.0037411011289805174, + 0.001980822766199708, + 0.005055414978414774, + 0.0034737917594611645, + 0.0056541310623288155, + 0.0029807728715240955, + 0.012441849336028099, + 0.00891182478517294, + 0.004389692563563585, + 0.002955906791612506, + 0.0018056187545880675, + 0.0018776399083435535, + 0.002580473432317376, + 0.0012487738858908415, + 0.0021601025946438313, + 0.002596479607746005, + 0.0011906358413398266, + 0.0014771182322874665, + 0.002230976475402713, + 0.007088686805218458, + 0.0027117806021124125, + 0.0024731236044317484, + 0.004805712029337883, + 0.0027942820452153683, + 0.00564305717125535, + 0.0033398109953850508, + 0.0026659376453608274, + 0.001777655677869916, + 0.0016697581158950925, + 0.002224898897111416, + 0.0028150680009275675, + 0.0018108284566551447, + 0.003785467939451337, + 0.0013285528402775526, + 0.0027031442150473595, + 0.0014754117000848055, + 0.0008043894194997847, + 0.0012513003312051296, + 0.0013843872584402561, + 0.0008032171754166484, + 0.004476331640034914, + 0.0007500943029299378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.007626465056091547 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 360-degree pan from a slightly bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an intricately carved copper pot with vintage patterns of auspicious clouds and dragons visible in the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the pot contain a vibrant red oil broth simmering with dried chilies and Sichuan peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are plates of marbled beef, green spinach, white jade mushrooms, and lotus root slices arranged around the pot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a Chinese palace lantern suspended above the table casting warm yellow 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": [ + -0.12703441083431244 + ], + "opposite_logprob_tokens": [ + -2.1270344257354736 + ], + "logprob_reward": -0.12703441083431244, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703441083431244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.740626871585846, + 0.04243188723921776, + 0.014030016958713531, + 0.011318051256239414, + 0.016388114541769028, + 0.09864762425422668, + 0.0296435859054327, + 0.017776161432266235, + 0.01643499732017517, + 0.012702629901468754 + ], + "credit_map": [ + 0.046406470239162445, + 0.03500603884458542, + 0.022443097084760666, + 0.02155364863574505, + 0.3301588296890259, + 0.21655234694480896, + 0.04803891107439995, + 0.02046751044690609, + 0.0034380091819912195, + 0.004183325916528702, + 0.0031934119760990143, + 0.004497172776609659, + 0.013057813979685307, + 0.00468421308323741, + 0.006672431714832783, + 0.0027055074460804462, + 0.00160312675870955, + 0.0013058335753157735, + 0.0015086188213899732, + 0.0017365168314427137, + 0.003489996772259474, + 0.0016487805405631661, + 0.0017155577661469579, + 0.0010215844959020615, + 0.0015443188603967428, + 0.0015257310587912798, + 0.0010755067924037576, + 0.0019735570531338453, + 0.0012688004644587636, + 0.0012044330360367894, + 0.002056695753708482, + 0.0006690085283480585, + 0.0016503739170730114, + 0.0012479252181947231, + 0.0021639687474817038, + 0.0012857296969741583, + 0.0038627309259027243, + 0.0016071299323812127, + 0.002322772052139044, + 0.002247484168037772, + 0.009726282209157944, + 0.006036267150193453, + 0.00902740191668272, + 0.00447932630777359, + 0.029301216825842857, + 0.027544524520635605, + 0.008363659493625164, + 0.004168945364654064, + 0.003678403329104185, + 0.0036401108372956514, + 0.0034921420738101006, + 0.0032742884941399097, + 0.005262379534542561, + 0.004362399689853191, + 0.003593772416934371, + 0.002340089064091444, + 0.001990116434171796, + 0.0057131582871079445, + 0.0018029740313068032, + 0.001377591979689896, + 0.001934098661877215, + 0.0012464661849662662, + 0.0026607001200318336, + 0.0010510552674531937, + 0.0020617858972400427, + 0.0024970336817204952, + 0.0011821932857856154, + 0.0016058481996878982, + 0.0027304310351610184, + 0.0015467770863324404, + 0.0032150433398783207, + 0.0015958861913532019, + 0.003630894934758544, + 0.0008064443827606738, + 0.0006684763357043266, + 0.0009813280776143074, + 0.0007354178233072162, + 0.000583856541197747, + 0.004264622461050749, + 0.001031589461490512 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.12703441083431244 + } + ], + "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/0000160/group_10/index_tensor([3321, 3321])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_tensor([3321, 3321])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cee474e37a2d6515860710713f9d30542f80da5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_tensor([3321, 3321])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium close-up depicts a man with a neatly trimmed dark brown beard, wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap. He stands in the center of a sun-drenched meadow, the soft-focus background revealing tall green grass and scattered yellow wildflowers in the warm, golden afternoon light. His gaze is intently directed downward at a small metal ring delicately held between the thumb and index finger of each hand. He carefully wipes away specks of dark dirt from the surface, his lips moving as he speaks, narrating his actions with an air of focused discovery. He fluidly turns his head momentarily to the left, as if to greet someone off-screen, before quickly returning his gaze to the precious object in his hands, his expression a mixture of concentration and quiet excitement.", + "seed": 14241, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14241_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": -0.022725538816303015, + "vlm_logprob": -0.022725538816303015, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.022725538816303015, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04392101615667343 + ], + "opposite_logprob_tokens": [ + -4.293920993804932 + ], + "logprob_reward": -0.04392101615667343, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.04392101615667343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15916085243225098, + 0.05831784009933472, + 0.2603086233139038, + 0.04693737253546715, + 0.040632180869579315, + 0.021993983536958694, + 0.0612667016685009, + 0.03511768579483032, + 0.15442922711372375, + 0.16183549165725708 + ], + "credit_map": [ + 0.024046974256634712, + 0.011918894946575165, + 0.01157347671687603, + 0.01572093553841114, + 0.0203398410230875, + 0.01832704246044159, + 0.020850932225584984, + 0.036382757127285004, + 0.009551914408802986, + 0.004556314088404179, + 0.0038168507162481546, + 0.007554848678410053, + 0.005966330412775278, + 0.01175648532807827, + 0.005298670846968889, + 0.009816423058509827, + 0.022205784916877747, + 0.01944972388446331, + 0.011711329221725464, + 0.06905742734670639, + 0.010306117124855518, + 0.08024632930755615, + 0.023915970697999, + 0.023415930569171906, + 0.005171117838472128, + 0.004691462963819504, + 0.005767126567661762, + 0.007790518458932638, + 0.004362907260656357, + 0.0054998318664729595, + 0.007635646499693394, + 0.00601876201108098, + 0.004088073503226042, + 0.0036740724463015795, + 0.0032254166435450315, + 0.003348300466313958, + 0.007025501690804958, + 0.006196399684995413, + 0.004711299203336239, + 0.008363117463886738, + 0.0031068213284015656, + 0.002829192439094186, + 0.0013673027278855443, + 0.0020510717295110226, + 0.004459386225789785, + 0.003112160600721836, + 0.0020168758928775787, + 0.0030511729419231415, + 0.010732210241258144, + 0.005133137106895447, + 0.008298885077238083, + 0.0076820687390863895, + 0.00621699308976531, + 0.009992658160626888, + 0.007225626148283482, + 0.005985126364976168, + 0.007314730901271105, + 0.003154620761051774, + 0.003924541641026735, + 0.00230239424854517, + 0.004071976523846388, + 0.008015824481844902, + 0.003580257296562195, + 0.0027533387765288353, + 0.029483921825885773, + 0.012055404484272003, + 0.010907000862061977, + 0.017777778208255768, + 0.015384105034172535, + 0.03681192174553871, + 0.0166721623390913, + 0.01533693540841341, + 0.02506173588335514, + 0.01634743995964527, + 0.009382989257574081, + 0.047831568866968155, + 0.01291703525930643, + 0.01500788889825344, + 0.022054137662053108, + 0.013232681900262833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.04392101615667343 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560192495584488 + ], + "opposite_logprob_tokens": [ + -4.3875603675842285 + ], + "logprob_reward": -0.012560192495584488, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560192495584488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16338993608951569, + 0.07834011316299438, + 0.39572960138320923, + 0.013754512183368206, + 0.03495386615395546, + 0.025009220466017723, + 0.04048025608062744, + 0.013562980107963085, + 0.11568141728639603, + 0.11909817159175873 + ], + "credit_map": [ + 0.023583419620990753, + 0.01626507379114628, + 0.011552460491657257, + 0.013654266484081745, + 0.0271439328789711, + 0.01872609183192253, + 0.019929109141230583, + 0.032535575330257416, + 0.01006736047565937, + 0.0075605446472764015, + 0.008105524815618992, + 0.008364856243133545, + 0.01619221456348896, + 0.010853475891053677, + 0.009752054698765278, + 0.007444076240062714, + 0.027890091761946678, + 0.026919402182102203, + 0.021649975329637527, + 0.09246496856212616, + 0.031063267961144447, + 0.08111762255430222, + 0.07784391939640045, + 0.03678031265735626, + 0.001804151339456439, + 0.001457828562706709, + 0.001251501846127212, + 0.0018489814829081297, + 0.0029971685726195574, + 0.0015643672086298466, + 0.0018270673463121057, + 0.0010034440783783793, + 0.005095343571156263, + 0.0035891584120690823, + 0.0033796222414821386, + 0.007614342495799065, + 0.004504767712205648, + 0.004360816441476345, + 0.003221973543986678, + 0.0031878401059657335, + 0.004424704238772392, + 0.003843252779915929, + 0.001975885359570384, + 0.0034109645057469606, + 0.0032860981300473213, + 0.0033315306063741446, + 0.0023366035893559456, + 0.0024001768324524164, + 0.006602078210562468, + 0.003926255274564028, + 0.0035766216460615396, + 0.003462193300947547, + 0.0034903367049992085, + 0.006764499470591545, + 0.005681072361767292, + 0.006977197714149952, + 0.001743379165418446, + 0.0017567855538800359, + 0.0008095605880953372, + 0.00199287012219429, + 0.0019027305534109473, + 0.0017127413302659988, + 0.001780788879841566, + 0.0018641226924955845, + 0.019851865246891975, + 0.00865720584988594, + 0.019570806995034218, + 0.013654034584760666, + 0.008447589352726936, + 0.02609168365597725, + 0.009817824698984623, + 0.009590392000973225, + 0.021236825734376907, + 0.01114443875849247, + 0.007976762019097805, + 0.039296865463256836, + 0.006932784337550402, + 0.008781896904110909, + 0.017610028386116028, + 0.006118553690612316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.012560192495584488 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198776990175247 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.018198776990175247, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198776990175247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20169515907764435, + 0.055222608149051666, + 0.11661266535520554, + 0.036237768828868866, + 0.07041216641664505, + 0.12661342322826385, + 0.04702581837773323, + 0.14427809417247772, + 0.1010674387216568, + 0.10083481669425964 + ], + "credit_map": [ + 0.032973192632198334, + 0.014725836925208569, + 0.02185140550136566, + 0.015070267021656036, + 0.0222944226115942, + 0.027298493310809135, + 0.02369551546871662, + 0.04378603398799896, + 0.006322503089904785, + 0.005281988065689802, + 0.0029706989880651236, + 0.008988995105028152, + 0.009004366584122181, + 0.006348694209009409, + 0.007860423997044563, + 0.008444941602647305, + 0.010920662432909012, + 0.007855048403143883, + 0.007714799605309963, + 0.021950090304017067, + 0.009779909625649452, + 0.02476414293050766, + 0.022995661944150925, + 0.010632355697453022, + 0.0035862205550074577, + 0.003358344081789255, + 0.003332996740937233, + 0.006487190257757902, + 0.0046897707507014275, + 0.005471347831189632, + 0.0057915146462619305, + 0.0035203900188207626, + 0.009720203466713428, + 0.007394807878881693, + 0.005417900625616312, + 0.006784662138670683, + 0.013849332928657532, + 0.009101496078073978, + 0.009828565642237663, + 0.008315203711390495, + 0.010994712822139263, + 0.016597650945186615, + 0.009940946474671364, + 0.01207608263939619, + 0.01880674995481968, + 0.016787223517894745, + 0.017235388979315758, + 0.024174682796001434, + 0.006753950379788876, + 0.003380365902557969, + 0.005004169885069132, + 0.005293629132211208, + 0.005612519569694996, + 0.006231876090168953, + 0.0070325895212590694, + 0.007716724183410406, + 0.033903542906045914, + 0.008969283662736416, + 0.0055083646439015865, + 0.01098861638456583, + 0.01646585389971733, + 0.03536016866564751, + 0.020834336057305336, + 0.012247934006154537, + 0.015305209904909134, + 0.007397888228297234, + 0.018171874806284904, + 0.010249697603285313, + 0.008812944404780865, + 0.02229267917573452, + 0.010867753066122532, + 0.007969397120177746, + 0.014176367782056332, + 0.011420870199799538, + 0.007488974370062351, + 0.028921829536557198, + 0.006923336070030928, + 0.007353230845183134, + 0.013695066794753075, + 0.010855150409042835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.018198776990175247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052565219812095165 + ], + "opposite_logprob_tokens": [ + -5.255256652832031 + ], + "logprob_reward": -0.0052565219812095165, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052565219812095165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1125103309750557, + 0.023650499060750008, + 0.03668424114584923, + 0.053971607238054276, + 0.025936681777238846, + 0.02797095663845539, + 0.1618678867816925, + 0.03957103565335274, + 0.37442776560783386, + 0.14340901374816895 + ], + "credit_map": [ + 0.01911807619035244, + 0.008502564392983913, + 0.009599602781236172, + 0.01004739012569189, + 0.011174116283655167, + 0.015139678493142128, + 0.014125112444162369, + 0.02480378746986389, + 0.0044409530237317085, + 0.0021602651104331017, + 0.002375203650444746, + 0.0024352173786610365, + 0.0032264026813209057, + 0.003392620477825403, + 0.0022991420701146126, + 0.003320696298032999, + 0.005122463684529066, + 0.002798982197418809, + 0.0027459105476737022, + 0.005577339790761471, + 0.0032010655850172043, + 0.007648986764252186, + 0.003953045234084129, + 0.00563644478097558, + 0.007360232528299093, + 0.005895874463021755, + 0.0056639243848621845, + 0.009105680510401726, + 0.004777204245328903, + 0.00626168679445982, + 0.007552911061793566, + 0.007354096043854952, + 0.004804485943168402, + 0.0031460479367524385, + 0.002322461223229766, + 0.0023392117582261562, + 0.0033385083079338074, + 0.0036989094223827124, + 0.0026393686421215534, + 0.0036476864479482174, + 0.0029565496370196342, + 0.0031095831654965878, + 0.0022939839400351048, + 0.0026698154397308826, + 0.005186412949115038, + 0.003741097403690219, + 0.0029464473482221365, + 0.005067066289484501, + 0.02876632660627365, + 0.012279696762561798, + 0.027685299515724182, + 0.014699206687510014, + 0.012690187431871891, + 0.03290782868862152, + 0.012518838047981262, + 0.020320499315857887, + 0.01102258451282978, + 0.002206125296652317, + 0.0016804477199912071, + 0.0019415614660829306, + 0.004345038905739784, + 0.012333993799984455, + 0.0035075112245976925, + 0.0025337734259665012, + 0.05380106344819069, + 0.030324067920446396, + 0.09635600447654724, + 0.04066558554768562, + 0.02718588337302208, + 0.06303931772708893, + 0.0324234701693058, + 0.03063235431909561, + 0.02429521456360817, + 0.01231454685330391, + 0.01340437587350607, + 0.03890358284115791, + 0.009849742986261845, + 0.016065094619989395, + 0.01790495775640011, + 0.010671497322618961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0052565219812095165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.17865341901779175, + 0.040443308651447296, + 0.1738993525505066, + 0.039143577218055725, + 0.05888925865292549, + 0.044205643236637115, + 0.07962843775749207, + 0.054170265793800354, + 0.13927766680717468, + 0.1916891485452652 + ], + "credit_map": [ + 0.026404930278658867, + 0.014750758185982704, + 0.018743639811873436, + 0.014959752559661865, + 0.024043450132012367, + 0.020060978829860687, + 0.02051178365945816, + 0.039178118109703064, + 0.006608161143958569, + 0.002484505297616124, + 0.0027285816613584757, + 0.003553696209564805, + 0.005900357849895954, + 0.006834999658167362, + 0.004097072873264551, + 0.00823593232780695, + 0.012721975333988667, + 0.011426475830376148, + 0.011266082525253296, + 0.0434972420334816, + 0.010370391421020031, + 0.03323117271065712, + 0.03546465188264847, + 0.015921350568532944, + 0.004625872243195772, + 0.004454636014997959, + 0.00330803613178432, + 0.005711480975151062, + 0.005689313635230064, + 0.005399335641413927, + 0.005276132375001907, + 0.004678771831095219, + 0.007802782580256462, + 0.005606465972959995, + 0.005843486171215773, + 0.007067364640533924, + 0.008021805435419083, + 0.00824944581836462, + 0.007392905652523041, + 0.008904999122023582, + 0.005530551075935364, + 0.005808583926409483, + 0.003946146462112665, + 0.004201104398816824, + 0.006719698663800955, + 0.005591899156570435, + 0.006137289106845856, + 0.006270367652177811, + 0.0133058475330472, + 0.0072290790267288685, + 0.00919215939939022, + 0.007733436767011881, + 0.00865971390157938, + 0.013351867906749249, + 0.010354330763220787, + 0.009801993146538734, + 0.008990079164505005, + 0.004315026570111513, + 0.004964585416018963, + 0.004218309186398983, + 0.006126797292381525, + 0.01024337112903595, + 0.008423276245594025, + 0.006888817995786667, + 0.029009800404310226, + 0.012221291661262512, + 0.014400131069123745, + 0.01821613684296608, + 0.01463186927139759, + 0.024626143276691437, + 0.015110590495169163, + 0.011061692610383034, + 0.03900722414255142, + 0.01789718307554722, + 0.009171118959784508, + 0.05884291231632233, + 0.017486531287431717, + 0.012759046629071236, + 0.02576051838696003, + 0.01076460350304842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.03369118645787239 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up depicts a man with a neatly trimmed dark brown beard, wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap. He stands in the center of a sun-drenched meadow, the soft-focus background revealing tall green grass and scattered yellow wildflowers in the warm, golden afternoon light. His gaze is intently directed downward at a small metal ring delicately held between the thumb and index finger of each hand. He carefully wipes away specks of dark dirt from the surface, his lips moving as he speaks, narrating his actions with an air of focused discovery. He fluidly turns his head momentarily to the left, as if to greet someone off-screen, before quickly returning his gaze to the precious object in his hands, his expression a mixture of concentration and quiet excitement.", + "seed": 14242, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14242_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": -0.028371470561251045, + "vlm_logprob": -0.028371470561251045, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.028371470561251045, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0447470061480999 + ], + "opposite_logprob_tokens": [ + -4.5447468757629395 + ], + "logprob_reward": -0.0447470061480999, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0447470061480999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04848669841885567, + 0.0227302648127079, + 0.10079449415206909, + 0.20148925483226776, + 0.032618358731269836, + 0.13765907287597656, + 0.030863020569086075, + 0.026942983269691467, + 0.04588809981942177, + 0.35252776741981506 + ], + "credit_map": [ + 0.010459036566317081, + 0.003561859019100666, + 0.007711641024798155, + 0.007387196645140648, + 0.0048530735075473785, + 0.0063252318650484085, + 0.004522007890045643, + 0.0036666537635028362, + 0.003872802248224616, + 0.0022653082851320505, + 0.0033186511136591434, + 0.003094568382948637, + 0.002096960786730051, + 0.0025685718283057213, + 0.00375050469301641, + 0.0017628960777074099, + 0.01852181926369667, + 0.009001541882753372, + 0.017821909859776497, + 0.0075806924141943455, + 0.008221070282161236, + 0.021792927756905556, + 0.005070138722658157, + 0.012784390710294247, + 0.06848248839378357, + 0.012592324055731297, + 0.005822582170367241, + 0.0075743733905255795, + 0.008226998150348663, + 0.07334129512310028, + 0.004550368990749121, + 0.02089882269501686, + 0.0047616735100746155, + 0.0033697285689413548, + 0.004380899015814066, + 0.004648908507078886, + 0.0031469562090933323, + 0.006061555352061987, + 0.003472263226285577, + 0.0027763729449361563, + 0.04564541578292847, + 0.009782923385500908, + 0.00649674516171217, + 0.007087897043675184, + 0.008128677494823933, + 0.0491596981883049, + 0.005477142054587603, + 0.005880566313862801, + 0.004052557051181793, + 0.002312650205567479, + 0.0023810393176972866, + 0.00550595298409462, + 0.0026524525601416826, + 0.003334679873660207, + 0.007290747947990894, + 0.003332938766106963, + 0.0041894433088600636, + 0.0032510352320969105, + 0.0031804197933524847, + 0.0035994804929941893, + 0.0028798982966691256, + 0.0026499649975448847, + 0.004735050722956657, + 0.0024576897267252207, + 0.005661125294864178, + 0.007084397599101067, + 0.0065057966858148575, + 0.005692483391612768, + 0.007471089717000723, + 0.004225395154207945, + 0.00475873239338398, + 0.004489079583436251, + 0.052310191094875336, + 0.021973714232444763, + 0.07890333980321884, + 0.014386212453246117, + 0.02123594470322132, + 0.1325991153717041, + 0.012986138463020325, + 0.01813311129808426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.0447470061480999 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038130003958940506 + ], + "opposite_logprob_tokens": [ + -3.288130044937134 + ], + "logprob_reward": -0.038130003958940506, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038130003958940506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06476067751646042, + 0.09681779891252518, + 0.014925376512110233, + 0.1251721829175949, + 0.044016383588314056, + 0.05714384838938713, + 0.07215026766061783, + 0.04516945034265518, + 0.07023247331380844, + 0.4096114933490753 + ], + "credit_map": [ + 0.01028760802000761, + 0.006011906545609236, + 0.011922965757548809, + 0.01618114486336708, + 0.004835417028516531, + 0.006909010466188192, + 0.004966037813574076, + 0.003646597033366561, + 0.020285144448280334, + 0.010748426429927349, + 0.0122337955981493, + 0.008900165557861328, + 0.007467614021152258, + 0.012905308045446873, + 0.009924895130097866, + 0.014352460391819477, + 0.003193741198629141, + 0.001132696750573814, + 0.003664958057925105, + 0.0013852986739948392, + 0.0013983655953779817, + 0.0024590047542005777, + 0.0007823628839105368, + 0.0009089498198591173, + 0.04237496107816696, + 0.007577911950647831, + 0.0037383344024419785, + 0.0034552307333797216, + 0.005477581173181534, + 0.04574095085263252, + 0.0036531889345496893, + 0.013154033571481705, + 0.006862990092486143, + 0.005232695024460554, + 0.004139030817896128, + 0.006341478321701288, + 0.004701254423707724, + 0.007794720120728016, + 0.003548487788066268, + 0.005395729560405016, + 0.021510222926735878, + 0.004086161497980356, + 0.0018553908448666334, + 0.0018203911604359746, + 0.0018786563305184245, + 0.02260930836200714, + 0.0014892497565597296, + 0.0018944720504805446, + 0.018543293699622154, + 0.004818346817046404, + 0.008884022012352943, + 0.005293128080666065, + 0.003984703682363033, + 0.019447635859251022, + 0.006195997819304466, + 0.004983150400221348, + 0.006346455309540033, + 0.0033543321769684553, + 0.0050558908842504025, + 0.007439008913934231, + 0.004711423069238663, + 0.003897636430338025, + 0.007460787892341614, + 0.0069039189256727695, + 0.008678467944264412, + 0.009040468372404575, + 0.010838430374860764, + 0.009378672577440739, + 0.010096879675984383, + 0.005476297345012426, + 0.007788433693349361, + 0.008934828452765942, + 0.05643443390727043, + 0.02730761095881462, + 0.09453880041837692, + 0.017221499234437943, + 0.025015980005264282, + 0.15089216828346252, + 0.018113330006599426, + 0.020087692886590958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.038130003958940506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361288502812386 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361288502812386, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361288502812386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060404568910598755, + 0.04025715962052345, + 0.03875066712498665, + 0.3758261501789093, + 0.02763495221734047, + 0.06033586710691452, + 0.10107606649398804, + 0.04736290127038956, + 0.033806513994932175, + 0.2145450860261917 + ], + "credit_map": [ + 0.008322480134665966, + 0.005870691034942865, + 0.007985789328813553, + 0.008523407392203808, + 0.0076394593343138695, + 0.009031090885400772, + 0.0064079915173351765, + 0.006623662542551756, + 0.006657359190285206, + 0.0035650681238621473, + 0.004344056826084852, + 0.0037520502228289843, + 0.0047830380499362946, + 0.005060478113591671, + 0.004306057933717966, + 0.007789056748151779, + 0.006880690343677998, + 0.0050672260113060474, + 0.003340404247865081, + 0.003398803062736988, + 0.00418986426666379, + 0.007993736304342747, + 0.0031046399381011724, + 0.0047753071412444115, + 0.12351161241531372, + 0.02274148538708687, + 0.015560613013803959, + 0.010914788581430912, + 0.018470002338290215, + 0.13358861207962036, + 0.009879197925329208, + 0.041159894317388535, + 0.005270576570183039, + 0.00274603976868093, + 0.0029220927972346544, + 0.004160794895142317, + 0.002757360925897956, + 0.003412219462916255, + 0.004304460249841213, + 0.0020614115055650473, + 0.01653626747429371, + 0.004475289955735207, + 0.003062922041863203, + 0.007063151802867651, + 0.0039978595450520515, + 0.016668330878019333, + 0.006139758974313736, + 0.0023922903928905725, + 0.028477584943175316, + 0.006640944629907608, + 0.015140664763748646, + 0.005570050794631243, + 0.005041430238634348, + 0.029161691665649414, + 0.005785831715911627, + 0.0052578807808458805, + 0.01059749536216259, + 0.004006017930805683, + 0.004116219934076071, + 0.005522787570953369, + 0.004779667127877474, + 0.005259868688881397, + 0.008599143475294113, + 0.004481709562242031, + 0.0057435729540884495, + 0.004150261636823416, + 0.006795048248022795, + 0.0035521152894943953, + 0.004009872674942017, + 0.0030135365668684244, + 0.0028647631406784058, + 0.0036773455794900656, + 0.03391437977552414, + 0.012996180914342403, + 0.044765062630176544, + 0.008296684361994267, + 0.012425016611814499, + 0.08410412818193436, + 0.007534972857683897, + 0.010508676059544086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.026361288502812386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028225842397660017 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028225842397660017, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028225842397660017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05589187145233154, + 0.01886652410030365, + 0.15988247096538544, + 0.16995875537395477, + 0.03425049036741257, + 0.2916201055049896, + 0.031480979174375534, + 0.0716862604022026, + 0.021889280527830124, + 0.14447328448295593 + ], + "credit_map": [ + 0.011648244224488735, + 0.005069531034678221, + 0.007161951623857021, + 0.0059769949875772, + 0.006191193591803312, + 0.006947418209165335, + 0.005049056373536587, + 0.007847479544579983, + 0.0028586559928953648, + 0.001648718025535345, + 0.00277563720010221, + 0.0031465585343539715, + 0.001905213575810194, + 0.002456365618854761, + 0.0023454686161130667, + 0.0017299065366387367, + 0.02957184799015522, + 0.014553978107869625, + 0.02353908121585846, + 0.016081124544143677, + 0.013711445033550262, + 0.03568965196609497, + 0.006774030160158873, + 0.019961310550570488, + 0.05853240564465523, + 0.010508737526834011, + 0.004662648309022188, + 0.00499604269862175, + 0.006870310287922621, + 0.06319219619035721, + 0.0035489776637405157, + 0.01764744333922863, + 0.0054017421789467335, + 0.003467830829322338, + 0.004608869086951017, + 0.004275802988559008, + 0.003411170793697238, + 0.00649284478276968, + 0.0027095910627394915, + 0.0038826398085802794, + 0.10388193279504776, + 0.021221397444605827, + 0.011750631034374237, + 0.012198603712022305, + 0.013191059231758118, + 0.11169683188199997, + 0.009208962321281433, + 0.008470668457448483, + 0.0054673096165061, + 0.003584320889785886, + 0.007299025543034077, + 0.0033752748277038336, + 0.0026531510520726442, + 0.0036331149749457836, + 0.003504994325339794, + 0.0019637891091406345, + 0.015136269852519035, + 0.004862363915890455, + 0.015602263621985912, + 0.005890767555683851, + 0.006175247021019459, + 0.010537915863096714, + 0.006807334721088409, + 0.006674095988273621, + 0.0025525607634335756, + 0.0033969527576118708, + 0.003121940651908517, + 0.002823806367814541, + 0.0029188513290137053, + 0.0022587720304727554, + 0.002414089161902666, + 0.002402308164164424, + 0.020246803760528564, + 0.00949257705360651, + 0.03261104226112366, + 0.0071796150878071785, + 0.008553189225494862, + 0.05202139541506767, + 0.007008392829447985, + 0.007360267918556929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.0028225842397660017 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979646995663643 + ], + "opposite_logprob_tokens": [ + -3.5297963619232178 + ], + "logprob_reward": -0.02979646995663643, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979646995663643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0647001713514328, + 0.033098477870225906, + 0.04822179302573204, + 0.14761225879192352, + 0.044369302690029144, + 0.3789004683494568, + 0.07533480226993561, + 0.031550705432891846, + 0.05391133949160576, + 0.12230075150728226 + ], + "credit_map": [ + 0.00897764042019844, + 0.006755258422344923, + 0.007889529690146446, + 0.00878234300762415, + 0.0071630594320595264, + 0.008918607607483864, + 0.00871567614376545, + 0.007498057093471289, + 0.0061691682785749435, + 0.0033243231009691954, + 0.002978081116452813, + 0.0033648281823843718, + 0.003440394066274166, + 0.005121761001646519, + 0.003808738198131323, + 0.00489118043333292, + 0.00950910896062851, + 0.005457952618598938, + 0.007590191904455423, + 0.0035155678633600473, + 0.005292385816574097, + 0.010079371742904186, + 0.0022497999016195536, + 0.004527411889284849, + 0.048750508576631546, + 0.008805432356894016, + 0.005288388580083847, + 0.005906836595386267, + 0.007255586329847574, + 0.05266665294766426, + 0.0036537565756589174, + 0.015285089612007141, + 0.006031821947544813, + 0.0053436728194355965, + 0.004650592338293791, + 0.009873418137431145, + 0.0038024685345590115, + 0.005880279932171106, + 0.003922057803720236, + 0.004864987917244434, + 0.13955549895763397, + 0.027742214500904083, + 0.013910339213907719, + 0.013370622880756855, + 0.014502210542559624, + 0.1493767648935318, + 0.010517646558582783, + 0.009925097227096558, + 0.014600598253309727, + 0.005506997928023338, + 0.011870862916111946, + 0.00867472030222416, + 0.006066327448934317, + 0.010859155096113682, + 0.011437804438173771, + 0.006318330764770508, + 0.004651776980608702, + 0.0035155070945620537, + 0.003882534569129348, + 0.0038784502539783716, + 0.0038153284694999456, + 0.0034929763060063124, + 0.0044129397720098495, + 0.0039011905901134014, + 0.005950599908828735, + 0.0072201830334961414, + 0.007437044754624367, + 0.006548697594553232, + 0.008523317985236645, + 0.004151565488427877, + 0.00743781216442585, + 0.006642113905400038, + 0.014570514671504498, + 0.007443837355822325, + 0.03030932880938053, + 0.007227907422930002, + 0.006989524699747562, + 0.04140983894467354, + 0.0076314243488013744, + 0.00671837292611599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 9 + ], + "selected_reward": -0.02979646995663643 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up depicts a man with a neatly trimmed dark brown beard, wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap. He stands in the center of a sun-drenched meadow, the soft-focus background revealing tall green grass and scattered yellow wildflowers in the warm, golden afternoon light. His gaze is intently directed downward at a small metal ring delicately held between the thumb and index finger of each hand. He carefully wipes away specks of dark dirt from the surface, his lips moving as he speaks, narrating his actions with an air of focused discovery. He fluidly turns his head momentarily to the left, as if to greet someone off-screen, before quickly returning his gaze to the precious object in his hands, his expression a mixture of concentration and quiet excitement.", + "seed": 14243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14243_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": -0.029241559468209742, + "vlm_logprob": -0.029241559468209742, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.029241559468209742, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06660810112953186 + ], + "opposite_logprob_tokens": [ + -4.06660795211792 + ], + "logprob_reward": -0.06660810112953186, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.06660810112953186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2860138416290283, + 0.1775885671377182, + 0.05559132620692253, + 0.048079390078783035, + 0.027935657650232315, + 0.025714781135320663, + 0.041761238127946854, + 0.06878812611103058, + 0.10544943064451218, + 0.16307762265205383 + ], + "credit_map": [ + 0.07478984445333481, + 0.10426866263151169, + 0.023452430963516235, + 0.017028270289301872, + 0.0210659671574831, + 0.02173938788473606, + 0.009207184426486492, + 0.014462074264883995, + 0.02697257697582245, + 0.05350811034440994, + 0.03588394820690155, + 0.011826399713754654, + 0.010724881663918495, + 0.01918225921690464, + 0.008323289453983307, + 0.011167105287313461, + 0.0062926230020821095, + 0.005706985481083393, + 0.0048305275849998, + 0.009172530844807625, + 0.005047988146543503, + 0.007024703081697226, + 0.007435958832502365, + 0.010080013424158096, + 0.006888743955641985, + 0.004655039869248867, + 0.005589640699326992, + 0.004413101822137833, + 0.007164192385971546, + 0.005252728704363108, + 0.007343885488808155, + 0.006772055756300688, + 0.0040011340752244, + 0.002181822434067726, + 0.00245039165019989, + 0.003939629532396793, + 0.0045153615064918995, + 0.003727325703948736, + 0.00404608016833663, + 0.00307391001842916, + 0.0034024144988507032, + 0.002301017986610532, + 0.004018939565867186, + 0.002127370098605752, + 0.004408948123455048, + 0.0033793780021369457, + 0.0035800496116280556, + 0.0024966630153357983, + 0.007620537653565407, + 0.005686694756150246, + 0.00491239782422781, + 0.004976141732186079, + 0.004641533829271793, + 0.005975889042019844, + 0.004382561892271042, + 0.0035654804669320583, + 0.01572575233876705, + 0.005398179870098829, + 0.006755637936294079, + 0.006033151410520077, + 0.007110715843737125, + 0.016385087743401527, + 0.005061975214630365, + 0.0063176280818879604, + 0.012101911939680576, + 0.007766077760607004, + 0.017466098070144653, + 0.00969034992158413, + 0.015221410430967808, + 0.021735981106758118, + 0.01231811661273241, + 0.009149487130343914, + 0.01673242449760437, + 0.013810696080327034, + 0.042762745171785355, + 0.0073286183178424835, + 0.01728413626551628, + 0.04783289134502411, + 0.01060386374592781, + 0.006722241174429655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.06660810112953186 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014226444996893406 + ], + "opposite_logprob_tokens": [ + -4.26422643661499 + ], + "logprob_reward": -0.014226444996893406, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014226444996893406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5421493053436279, + 0.07822176814079285, + 0.03512200340628624, + 0.017225397750735283, + 0.06212568283081055, + 0.015859032049775124, + 0.022520048543810844, + 0.02713554911315441, + 0.036820217967033386, + 0.16282100975513458 + ], + "credit_map": [ + 0.15506581962108612, + 0.21757687628269196, + 0.026504984125494957, + 0.024301905184984207, + 0.035115811973810196, + 0.05026553198695183, + 0.01368844136595726, + 0.019629916176199913, + 0.010784815065562725, + 0.011837693862617016, + 0.017711574211716652, + 0.009615277871489525, + 0.007965964265167713, + 0.008752563036978245, + 0.004836617037653923, + 0.006717254873365164, + 0.00509257335215807, + 0.003786774817854166, + 0.004389753099530935, + 0.007007536944001913, + 0.003134037833660841, + 0.0049375006929039955, + 0.002657205332070589, + 0.0041166204027831554, + 0.0021482505835592747, + 0.0023989989422261715, + 0.002361824968829751, + 0.0031489909160882235, + 0.0015713910106569529, + 0.0020893816836178303, + 0.0013119030045345426, + 0.00219465559348464, + 0.007064847741276026, + 0.005448170937597752, + 0.013650798238813877, + 0.012020450085401535, + 0.004981100559234619, + 0.009202760644257069, + 0.0050862254574894905, + 0.004671324975788593, + 0.0027931772638112307, + 0.001528348890133202, + 0.002682590624317527, + 0.0017258450388908386, + 0.0015801744302734733, + 0.0023195925168693066, + 0.001251390902325511, + 0.0019779126159846783, + 0.00556176109239459, + 0.0018600225448608398, + 0.002449972555041313, + 0.0024218882899731398, + 0.00251955958083272, + 0.003730611875653267, + 0.0020780563354492188, + 0.001898174057714641, + 0.005864443723112345, + 0.001745490008033812, + 0.0033197326119989157, + 0.002801191294565797, + 0.0030010933987796307, + 0.005474371369928122, + 0.002547871321439743, + 0.002381355734542012, + 0.004802322015166283, + 0.0036085147876292467, + 0.004067534580826759, + 0.003431546501815319, + 0.006215644534677267, + 0.0073899454437196255, + 0.004466775339096785, + 0.002837936393916607, + 0.013099818490445614, + 0.007450934965163469, + 0.043733853846788406, + 0.007984488271176815, + 0.009026715531945229, + 0.06263837963342667, + 0.006951746065169573, + 0.011935072019696236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.014226444996893406 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594803616404533 + ], + "opposite_logprob_tokens": [ + -3.895594835281372 + ], + "logprob_reward": -0.020594803616404533, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594803616404533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5495521426200867, + 0.20181304216384888, + 0.03578365221619606, + 0.022423742339015007, + 0.019801242277026176, + 0.014848453924059868, + 0.04988006874918938, + 0.023318832740187645, + 0.052220944315195084, + 0.030357863754034042 + ], + "credit_map": [ + 0.14979082345962524, + 0.20668627321720123, + 0.02876037172973156, + 0.0291487704962492, + 0.03838399052619934, + 0.057730671018362045, + 0.015677345916628838, + 0.023373929783701897, + 0.025565044954419136, + 0.03238196298480034, + 0.047312937676906586, + 0.01719672605395317, + 0.01874084584414959, + 0.023532574996352196, + 0.014588642865419388, + 0.022494329139590263, + 0.004515452776104212, + 0.003869706764817238, + 0.0030273005831986666, + 0.004702817648649216, + 0.003960970789194107, + 0.004679540637880564, + 0.004509555641561747, + 0.006518309935927391, + 0.004069173242896795, + 0.002401738427579403, + 0.002150805201381445, + 0.0017310349503532052, + 0.0029338528402149677, + 0.002116081304848194, + 0.0024773620534688234, + 0.004543696530163288, + 0.002767528872936964, + 0.0018654785817489028, + 0.0020256894640624523, + 0.002823330694809556, + 0.0021159006282687187, + 0.0030840460676699877, + 0.0027146199718117714, + 0.0024046499747782946, + 0.002099438104778528, + 0.0018827550811693072, + 0.0015973766567185521, + 0.0012678328203037381, + 0.0021069387439638376, + 0.0021817798260599375, + 0.0019993928726762533, + 0.0017129415646195412, + 0.011224538087844849, + 0.003895337460562587, + 0.0061799087561666965, + 0.004468899220228195, + 0.006951275747269392, + 0.007581871002912521, + 0.005310069769620895, + 0.004268174059689045, + 0.00506684510037303, + 0.001681370078586042, + 0.0019994250033050776, + 0.0020789559930562973, + 0.002274606842547655, + 0.005665204953402281, + 0.0023044603876769543, + 0.0022479670587927103, + 0.007623244542628527, + 0.0038179061375558376, + 0.0079825883731246, + 0.005434268619865179, + 0.006171193905174732, + 0.009059654548764229, + 0.005679342895746231, + 0.00645274855196476, + 0.0031452265102416277, + 0.00400736415758729, + 0.00386948324739933, + 0.0026571352500468493, + 0.005262037273496389, + 0.006048406939953566, + 0.003350716084241867, + 0.0020174975506961346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.020594803616404533 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2951536178588867, + 0.1448753923177719, + 0.05616915225982666, + 0.01608503982424736, + 0.0067702727392315865, + 0.006272242404520512, + 0.09754480421543121, + 0.07348557561635971, + 0.022192535921931267, + 0.28145137429237366 + ], + "credit_map": [ + 0.08586236089468002, + 0.11902033537626266, + 0.01451873779296875, + 0.011082060635089874, + 0.019953660666942596, + 0.027722598984837532, + 0.006894445978105068, + 0.010099402628839016, + 0.021444320678710938, + 0.02704203501343727, + 0.035108909010887146, + 0.015187220647931099, + 0.011555043049156666, + 0.01566203124821186, + 0.008685024455189705, + 0.010190804488956928, + 0.01163931842893362, + 0.00626174034550786, + 0.007669453043490648, + 0.007039661053568125, + 0.004856111481785774, + 0.007465718779712915, + 0.002660993253812194, + 0.008576154708862305, + 0.0033056919928640127, + 0.0019496310269460082, + 0.0013193435734137893, + 0.0018158179009333253, + 0.0019623651169240475, + 0.0023642899468541145, + 0.0016775878611952066, + 0.0016903121722862124, + 0.0010538843926042318, + 0.0006813065847381949, + 0.001013515517115593, + 0.0009037042618729174, + 0.0006649423157796264, + 0.0010369751835241914, + 0.0008443450788035989, + 0.0005715992883779109, + 0.0008687511435709894, + 0.0005293506546877325, + 0.0009617092437110841, + 0.0009522218024358153, + 0.0006022118031978607, + 0.0009109414531849325, + 0.0006730559980496764, + 0.0007740004803054035, + 0.02380981482565403, + 0.009287422522902489, + 0.010338381864130497, + 0.008318941108882427, + 0.012805089354515076, + 0.01594230718910694, + 0.009327532723546028, + 0.007715318817645311, + 0.01832418702542782, + 0.005908110179007053, + 0.011406933888792992, + 0.008169252425432205, + 0.004786759614944458, + 0.01361503079533577, + 0.005455545149743557, + 0.005819755606353283, + 0.0025816652923822403, + 0.001224254141561687, + 0.005235456861555576, + 0.0015913982642814517, + 0.0016354355029761791, + 0.006948753725737333, + 0.0018886239267885685, + 0.0010869494872167706, + 0.023744774982333183, + 0.00997140072286129, + 0.08483243733644485, + 0.013207029551267624, + 0.012532059103250504, + 0.10934992879629135, + 0.010818317532539368, + 0.016995428130030632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.011083230376243591 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369521722197533 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.03369521722197533, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369521722197533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3816990852355957, + 0.10709047317504883, + 0.060044221580028534, + 0.03649009391665459, + 0.02649684064090252, + 0.020224696025252342, + 0.0636967197060585, + 0.03650874271988869, + 0.12057063728570938, + 0.14717847108840942 + ], + "credit_map": [ + 0.0993160754442215, + 0.13800421357154846, + 0.02044505439698696, + 0.01813170500099659, + 0.028255989775061607, + 0.04373124614357948, + 0.013313544914126396, + 0.02050125226378441, + 0.014376365579664707, + 0.01917453669011593, + 0.02415006048977375, + 0.008308791555464268, + 0.009239021688699722, + 0.013569148257374763, + 0.007388169411569834, + 0.01088437344878912, + 0.007162964437156916, + 0.006172834429889917, + 0.00662714010104537, + 0.006698427256196737, + 0.007009865716099739, + 0.006619945168495178, + 0.00947351474314928, + 0.010279533453285694, + 0.004632001277059317, + 0.0032480251975357533, + 0.003162041772156954, + 0.0026763775385916233, + 0.0046121967025101185, + 0.005325916223227978, + 0.005836558062583208, + 0.006996978539973497, + 0.003975288476794958, + 0.002798456931486726, + 0.005793839227408171, + 0.003019317053258419, + 0.002543242648243904, + 0.003082942683249712, + 0.00298328697681427, + 0.0023004645481705666, + 0.003406715579330921, + 0.0018561161123216152, + 0.0024704039096832275, + 0.001843449310399592, + 0.003299797186627984, + 0.0030033974908292294, + 0.0026327131781727076, + 0.0017121037235483527, + 0.010952956043183804, + 0.0052520208992064, + 0.008997523225843906, + 0.007840060628950596, + 0.006878886371850967, + 0.008778432384133339, + 0.009736094623804092, + 0.00526074506342411, + 0.007729326840490103, + 0.0030170974787324667, + 0.004339595790952444, + 0.0031854421831667423, + 0.0033741937950253487, + 0.009304237551987171, + 0.002785469638183713, + 0.002773379208520055, + 0.017128076404333115, + 0.009094098582863808, + 0.02646888978779316, + 0.012748192064464092, + 0.011895067989826202, + 0.021648947149515152, + 0.013480967842042446, + 0.008106405846774578, + 0.013339200988411903, + 0.023846805095672607, + 0.020617466419935226, + 0.01283667515963316, + 0.02071995474398136, + 0.02639906294643879, + 0.017144206911325455, + 0.012275097891688347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.03369521722197533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium close-up depicts a man with a neatly trimmed dark brown beard, wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap. He stands in the center of a sun-drenched meadow, the soft-focus background revealing tall green grass and scattered yellow wildflowers in the warm, golden afternoon light. His gaze is intently directed downward at a small metal ring delicately held between the thumb and index finger of each hand. He carefully wipes away specks of dark dirt from the surface, his lips moving as he speaks, narrating his actions with an air of focused discovery. He fluidly turns his head momentarily to the left, as if to greet someone off-screen, before quickly returning his gaze to the precious object in his hands, his expression a mixture of concentration and quiet excitement.", + "seed": 14244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_10/index_3321_seed_14244_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": -0.033023247262462974, + "vlm_logprob": -0.033023247262462974, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.033023247262462974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08188596367835999 + ], + "opposite_logprob_tokens": [ + -3.456886053085327 + ], + "logprob_reward": -0.08188596367835999, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.08188596367835999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05402586609125137, + 0.059016745537519455, + 0.044612787663936615, + 0.05125496909022331, + 0.24749048054218292, + 0.0659758672118187, + 0.04139178246259689, + 0.14864782989025116, + 0.02517828345298767, + 0.2624053955078125 + ], + "credit_map": [ + 0.010308443568646908, + 0.004459624644368887, + 0.0046108318492770195, + 0.0035343931522220373, + 0.0069710491225123405, + 0.012469541281461716, + 0.005791582632809877, + 0.005880400072783232, + 0.012261464260518551, + 0.0044828117825090885, + 0.004797573201358318, + 0.0022142392117530107, + 0.0063483393751084805, + 0.016513070091605186, + 0.006359774619340897, + 0.006039474159479141, + 0.012714049778878689, + 0.003036096226423979, + 0.0025792240630835295, + 0.0022221284452825785, + 0.003879404626786709, + 0.013678056187927723, + 0.003392592305317521, + 0.0031112348660826683, + 0.010454365983605385, + 0.00406617671251297, + 0.01038723811507225, + 0.00412949500605464, + 0.006450137123465538, + 0.006446380168199539, + 0.004921105224639177, + 0.004400072619318962, + 0.032334405928850174, + 0.015068340115249157, + 0.020790116861462593, + 0.045659139752388, + 0.026401318609714508, + 0.02420293167233467, + 0.04915745183825493, + 0.03387676924467087, + 0.0091497041285038, + 0.007882933132350445, + 0.006578577216714621, + 0.008572389371693134, + 0.009725421667098999, + 0.005033980589359999, + 0.012341054156422615, + 0.006691804621368647, + 0.00937820877879858, + 0.005616636015474796, + 0.007771361153572798, + 0.00377829116769135, + 0.004281487315893173, + 0.0031186328269541264, + 0.004323090426623821, + 0.003124077105894685, + 0.0507112592458725, + 0.02427072823047638, + 0.013124837540090084, + 0.00555051164701581, + 0.008554311469197273, + 0.009110325947403908, + 0.009051571600139141, + 0.028274279087781906, + 0.0029079639352858067, + 0.003191698342561722, + 0.002889114897698164, + 0.0023679432924836874, + 0.004055308178067207, + 0.003655713517218828, + 0.0034189338330179453, + 0.0026916097849607468, + 0.03282329440116882, + 0.03467029705643654, + 0.0537211075425148, + 0.08416717499494553, + 0.015286220237612724, + 0.009481804445385933, + 0.021001968532800674, + 0.011253518052399158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.08188596367835999 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016098516061902046 + ], + "opposite_logprob_tokens": [ + -4.141098499298096 + ], + "logprob_reward": -0.016098516061902046, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016098516061902046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11611199378967285, + 0.07034675776958466, + 0.1015971377491951, + 0.029343223199248314, + 0.09556117653846741, + 0.1140417754650116, + 0.02970857545733452, + 0.12393797188997269, + 0.036683160811662674, + 0.28266817331314087 + ], + "credit_map": [ + 0.020809369161725044, + 0.010485365986824036, + 0.005856834817677736, + 0.0066542173735797405, + 0.022151954472064972, + 0.023375321179628372, + 0.012532009743154049, + 0.014246922917664051, + 0.013885967433452606, + 0.0061978851445019245, + 0.004901929758489132, + 0.003192014992237091, + 0.012220501899719238, + 0.01580231823027134, + 0.006128844805061817, + 0.008017299696803093, + 0.03458937257528305, + 0.0039938041009008884, + 0.004548848606646061, + 0.0037874991539865732, + 0.004812157712876797, + 0.03887706249952316, + 0.003617891576141119, + 0.007370498962700367, + 0.0041086445562541485, + 0.0025652856566011906, + 0.002613931894302368, + 0.004139249213039875, + 0.003760807914659381, + 0.003260582685470581, + 0.005876690614968538, + 0.0030180327594280243, + 0.0194085743278265, + 0.006226787809282541, + 0.006676826160401106, + 0.017522169277071953, + 0.007001377176493406, + 0.013134145177900791, + 0.013783478178083897, + 0.011807833798229694, + 0.013782685622572899, + 0.011206164956092834, + 0.011758660897612572, + 0.01246341597288847, + 0.013971880078315735, + 0.011127404868602753, + 0.026603592559695244, + 0.013127979822456837, + 0.004273807164281607, + 0.00346653931774199, + 0.0044411467388272285, + 0.003353104693815112, + 0.0044649140909314156, + 0.0028244885616004467, + 0.004372677765786648, + 0.002511897822842002, + 0.04586182162165642, + 0.018300725147128105, + 0.012210183776915073, + 0.004647916182875633, + 0.00527578592300415, + 0.006574775092303753, + 0.007890010252594948, + 0.0231767687946558, + 0.005517677403986454, + 0.007103811018168926, + 0.004833319690078497, + 0.005683708004653454, + 0.003885655663907528, + 0.003220735117793083, + 0.0036231912672519684, + 0.002815064275637269, + 0.033617567270994186, + 0.04095902666449547, + 0.053175583481788635, + 0.08315322548151016, + 0.02023904211819172, + 0.013243350200355053, + 0.02231496386229992, + 0.015965433791279793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.016098516061902046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981683425605297 + ], + "opposite_logprob_tokens": [ + -3.5298168659210205 + ], + "logprob_reward": -0.02981683425605297, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981683425605297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10001267492771149, + 0.1343594789505005, + 0.056214865297079086, + 0.051339734345674515, + 0.1887422353029251, + 0.02714467979967594, + 0.027625106275081635, + 0.16576866805553436, + 0.01931925117969513, + 0.22947324812412262 + ], + "credit_map": [ + 0.01935795322060585, + 0.009438714012503624, + 0.007195830345153809, + 0.008481682278215885, + 0.011846456676721573, + 0.024063779041171074, + 0.008578948676586151, + 0.011049309745430946, + 0.022216597571969032, + 0.013409915380179882, + 0.01025648508220911, + 0.007648886181414127, + 0.01706126146018505, + 0.02720147743821144, + 0.020336385816335678, + 0.016228461638092995, + 0.010855254717171192, + 0.005967367440462112, + 0.0029559109825640917, + 0.006141437217593193, + 0.007382268086075783, + 0.012182358652353287, + 0.006624841596931219, + 0.00410542543977499, + 0.006955031771212816, + 0.004754239693284035, + 0.002910804469138384, + 0.006174690555781126, + 0.007019346114248037, + 0.007053029257804155, + 0.009167082607746124, + 0.007305511273443699, + 0.028623024001717567, + 0.012827410362660885, + 0.01570439711213112, + 0.03425534814596176, + 0.02502669021487236, + 0.01884978637099266, + 0.030983975157141685, + 0.022471603006124496, + 0.0037689097225666046, + 0.0018863347359001637, + 0.003245097352191806, + 0.0035194973461329937, + 0.002662740647792816, + 0.002745792968198657, + 0.006135016679763794, + 0.0031812896486371756, + 0.005550951696932316, + 0.0034875788260251284, + 0.0044461144134402275, + 0.0021509903017431498, + 0.0038811739068478346, + 0.0022283103317022324, + 0.0030469719786196947, + 0.002833017148077488, + 0.061831895262002945, + 0.025695204734802246, + 0.01323656365275383, + 0.005509366746991873, + 0.007791493088006973, + 0.009968209080398083, + 0.00986765418201685, + 0.03186829388141632, + 0.00251671951264143, + 0.0023670876398682594, + 0.003249961882829666, + 0.00387481483630836, + 0.0018576474394649267, + 0.001788665191270411, + 0.0018724184483289719, + 0.0017919369274750352, + 0.028469383716583252, + 0.0338313989341259, + 0.045102622359991074, + 0.06879858672618866, + 0.013053596019744873, + 0.009399832226336002, + 0.02057805471122265, + 0.010239767841994762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.02981683425605297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036208562087267637 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036208562087267637, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036208562087267637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05365042760968208, + 0.10538601130247116, + 0.1619204431772232, + 0.027244916185736656, + 0.0599149726331234, + 0.020865583792328835, + 0.05790989473462105, + 0.3176409602165222, + 0.0697091668844223, + 0.12575757503509521 + ], + "credit_map": [ + 0.008572865277528763, + 0.0046764640137553215, + 0.006028303410857916, + 0.006013165228068829, + 0.006623252294957638, + 0.00602721469476819, + 0.009057458490133286, + 0.006651708856225014, + 0.023381609469652176, + 0.00998441781848669, + 0.006780503783375025, + 0.0052694473415613174, + 0.01204246748238802, + 0.026954688131809235, + 0.011326778680086136, + 0.009646106511354446, + 0.05329784378409386, + 0.006176153663545847, + 0.011484947055578232, + 0.005354000721126795, + 0.008871355094015598, + 0.06041392311453819, + 0.006938943173736334, + 0.00938328169286251, + 0.007526425179094076, + 0.002191692590713501, + 0.005228244233876467, + 0.0024717114865779877, + 0.0022414703853428364, + 0.0035194025840610266, + 0.0020521124824881554, + 0.0020138579420745373, + 0.009487878531217575, + 0.003378084162250161, + 0.004225274082273245, + 0.011401830241084099, + 0.0059179821982979774, + 0.006451027933508158, + 0.010992135852575302, + 0.00806075893342495, + 0.003626918187364936, + 0.001858479343354702, + 0.0034921765327453613, + 0.0019488015677779913, + 0.0019456696463748813, + 0.0022920803166925907, + 0.002969897584989667, + 0.0027315604966133833, + 0.013065393082797527, + 0.010715814307332039, + 0.009318764321506023, + 0.007291456684470177, + 0.004581676330417395, + 0.004077035002410412, + 0.0048720198683440685, + 0.003987736068665981, + 0.11961679905653, + 0.048151928931474686, + 0.026270141825079918, + 0.009324651211500168, + 0.013410307466983795, + 0.017794184386730194, + 0.019546372815966606, + 0.06352659314870834, + 0.010410962626338005, + 0.010230711661279202, + 0.010678852908313274, + 0.015251344069838524, + 0.008784995414316654, + 0.005423082038760185, + 0.004690915811806917, + 0.004238302353769541, + 0.017079785466194153, + 0.013007193803787231, + 0.025994233787059784, + 0.04059237986803055, + 0.009126050397753716, + 0.0059742857702076435, + 0.008579452522099018, + 0.005404198542237282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.0036208562087267637 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, eye-level medium close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a heather-gray cotton short-sleeved shirt and a dark blue baseball cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man holding a small metal ring between the thumb and index finger of each hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sun-drenched meadow with tall green grass and scattered yellow wildflowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man wipe dirt from the ring, move his lips as if narrating, and turn his head to the left before returning his gaze?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336940661072731 + ], + "opposite_logprob_tokens": [ + -3.40869402885437 + ], + "logprob_reward": -0.0336940661072731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336940661072731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07707405090332031, + 0.09807924926280975, + 0.06720902025699615, + 0.1342928260564804, + 0.028811458498239517, + 0.03654326871037483, + 0.09496240317821503, + 0.09202374517917633, + 0.06918874382972717, + 0.3018152713775635 + ], + "credit_map": [ + 0.019992681220173836, + 0.006793701555579901, + 0.004833596758544445, + 0.0031630275771021843, + 0.00883631594479084, + 0.018544876947999, + 0.006171817891299725, + 0.008738038130104542, + 0.01655004918575287, + 0.009412254206836224, + 0.007013271562755108, + 0.005930112209171057, + 0.0107746422290802, + 0.021536583080887794, + 0.014486588537693024, + 0.01237575151026249, + 0.01823636330664158, + 0.003787511494010687, + 0.0023961409460753202, + 0.0038558675441890955, + 0.005762395448982716, + 0.01994817517697811, + 0.0069156065583229065, + 0.006306951865553856, + 0.029914868995547295, + 0.012117779813706875, + 0.02152811735868454, + 0.01516673993319273, + 0.013016666285693645, + 0.015408636070787907, + 0.013877536170184612, + 0.013262472115457058, + 0.0035161934792995453, + 0.0025791849475353956, + 0.0025073059368878603, + 0.00341569515876472, + 0.00464254803955555, + 0.002630463568493724, + 0.005319981835782528, + 0.004200085531920195, + 0.006109196227043867, + 0.003779970807954669, + 0.005002675577998161, + 0.004591234028339386, + 0.0038301669992506504, + 0.0029922949615865946, + 0.006722389720380306, + 0.003515340853482485, + 0.011453667655587196, + 0.009759502485394478, + 0.012491729110479355, + 0.010946350172162056, + 0.013525912538170815, + 0.00814537052065134, + 0.013896443881094456, + 0.01474342867732048, + 0.024565672501921654, + 0.014738913625478745, + 0.014821748249232769, + 0.006247417535632849, + 0.009658954106271267, + 0.007440853863954544, + 0.005473029799759388, + 0.009077153168618679, + 0.0093723488971591, + 0.013515000231564045, + 0.006968336179852486, + 0.007521067280322313, + 0.007655884604901075, + 0.009212197735905647, + 0.007699951063841581, + 0.0072439610958099365, + 0.038420435041189194, + 0.04355119168758392, + 0.06057632341980934, + 0.09395325928926468, + 0.01677452214062214, + 0.010785712860524654, + 0.025383993983268738, + 0.012369818985462189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.0336940661072731 + } + ], + "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/0000160/group_12/index_tensor([6744, 6744])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_tensor([6744, 6744])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..890ef06931ea6d6269ff8e03e42375135540f8f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_tensor([6744, 6744])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, moving slowly and gracefully in an arc around the subject, captures a classic, multi-faceted crystal bottle of Wuliangye, majestically placed upon a sheet of deep, light-absorbing black velvet. Dramatic, high-contrast Rembrandt light, cast with pinpoint precision from the side, carves sharp demarcations of light and dark into the bottle's crisp facets. Simultaneously, a strong rim light from behind outlines the bottle's crystalline rim, highlighting the bottle's iconic red label and the golden \"Wuliangye\" inscription. Crafted with extreme slow-motion particle effects, countless viscous droplets, shimmering like liquid gold, splash in from outside the frame. They gently collide and merge in mid-air, breaking into even smaller, shimmering beads before converging into elegant spirals, like a miniature galaxy, systematically encircling the still bottle. Each flying drop of wine reflects brilliant light, complementing the colorful scattered light and shadow reflected by the crystal bottle. The whole picture, under the warm tones, presents an extremely pure, luxurious and dynamic luxury advertising texture, creating a breathtaking atmosphere of refinement and nobility.", + "seed": 30154, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_6744_seed_30154_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": -0.03385135922580958, + "vlm_logprob": -0.03385135922580958, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03385135922580958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016411321237683296 + ], + "opposite_logprob_tokens": [ + -4.266411304473877 + ], + "logprob_reward": -0.016411321237683296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016411321237683296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21729667484760284, + 0.18346035480499268, + 0.041290078312158585, + 0.07449179142713547, + 0.038230571895837784, + 0.041259195655584335, + 0.03574644774198532, + 0.028105607256293297, + 0.05498180165886879, + 0.28513747453689575 + ], + "credit_map": [ + 0.025674540549516678, + 0.01779777742922306, + 0.03640442341566086, + 0.029148900881409645, + 0.022131938487291336, + 0.016986282542347908, + 0.026625879108905792, + 0.042526934295892715, + 0.013455916196107864, + 0.009932498447597027, + 0.04752444103360176, + 0.023179272189736366, + 0.014475524425506592, + 0.012591338716447353, + 0.014016290195286274, + 0.04828507825732231, + 0.006348762195557356, + 0.0031964925583451986, + 0.007139742374420166, + 0.0052386256866157055, + 0.0033415029756724834, + 0.005078968126326799, + 0.004729956854134798, + 0.006216031964868307, + 0.013541724532842636, + 0.00940038450062275, + 0.009676889516413212, + 0.007099783048033714, + 0.007786153815686703, + 0.009497907012701035, + 0.007099842187017202, + 0.010389115661382675, + 0.006848697550594807, + 0.0039863986894488335, + 0.0068898932076990604, + 0.0036317610647529364, + 0.002857414772734046, + 0.004732821602374315, + 0.004039921332150698, + 0.005243664141744375, + 0.010759411379694939, + 0.0030144457705318928, + 0.009288442321121693, + 0.00503345625475049, + 0.0023404720705002546, + 0.0021442528814077377, + 0.003205986460670829, + 0.005472727585583925, + 0.004228831734508276, + 0.00478773657232523, + 0.004091680981218815, + 0.00629840325564146, + 0.004569149576127529, + 0.0030285310931503773, + 0.004124297294765711, + 0.004617820493876934, + 0.006846569012850523, + 0.002994825365021825, + 0.005298929288983345, + 0.0022346549667418003, + 0.002021463355049491, + 0.0012989419046789408, + 0.0029610146302729845, + 0.004449211526662111, + 0.0077156308107078075, + 0.0039727455005049706, + 0.016053581610322, + 0.0027508786879479885, + 0.0026029725559055805, + 0.0025830829981714487, + 0.003012001281604171, + 0.01629091612994671, + 0.046148527413606644, + 0.06449256092309952, + 0.1082400232553482, + 0.013019108213484287, + 0.013475101441144943, + 0.005399931222200394, + 0.01466934010386467, + 0.019692890346050262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.016411321237683296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.10061502456665039, + 0.07123753428459167, + 0.07710612565279007, + 0.21387138962745667, + 0.03806488588452339, + 0.03556642681360245, + 0.06271540373563766, + 0.054089583456516266, + 0.04351110756397247, + 0.30322250723838806 + ], + "credit_map": [ + 0.012202650308609009, + 0.0072754682041704655, + 0.0166595671325922, + 0.011670893058180809, + 0.007444949820637703, + 0.006856526713818312, + 0.018387310206890106, + 0.02011765167117119, + 0.007118371315300465, + 0.004830764140933752, + 0.013970158994197845, + 0.011295759119093418, + 0.006055406294763088, + 0.006179435644298792, + 0.006791535299271345, + 0.014996104873716831, + 0.014958977699279785, + 0.005919087678194046, + 0.014722227118909359, + 0.009373443201184273, + 0.005881890654563904, + 0.008585653267800808, + 0.006245558150112629, + 0.01141929067671299, + 0.03241022676229477, + 0.02403425984084606, + 0.03717987611889839, + 0.02310998924076557, + 0.018323279917240143, + 0.01977943442761898, + 0.026998909190297127, + 0.032035425305366516, + 0.006686181295663118, + 0.002879314124584198, + 0.011582973413169384, + 0.0029720370657742023, + 0.002848758827894926, + 0.003010866232216358, + 0.00307563622482121, + 0.005009115673601627, + 0.007862708531320095, + 0.002833947539329529, + 0.009557240642607212, + 0.0036761295050382614, + 0.001599131035618484, + 0.0022146315313875675, + 0.0028469334356486797, + 0.004975705407559872, + 0.011276522651314735, + 0.006844666320830584, + 0.009601639583706856, + 0.0074623688124120235, + 0.004317782819271088, + 0.006073961034417152, + 0.006427284330129623, + 0.010711180977523327, + 0.008644386194646358, + 0.0051881722174584866, + 0.010859942063689232, + 0.004632460884749889, + 0.004721035249531269, + 0.00323083926923573, + 0.005125011783093214, + 0.011687733232975006, + 0.006441137287765741, + 0.003981824964284897, + 0.011148763820528984, + 0.003203636035323143, + 0.0021801681723445654, + 0.002493202220648527, + 0.004626653157174587, + 0.009435719810426235, + 0.053153254091739655, + 0.06692966818809509, + 0.11445192247629166, + 0.015130167827010155, + 0.013894499279558659, + 0.0065200659446418285, + 0.014463493600487709, + 0.018679415807127953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "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 main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.11283829063177109, + 0.12616971135139465, + 0.09270283579826355, + 0.09874770790338516, + 0.05299542099237442, + 0.022634845227003098, + 0.069687120616436, + 0.02501717023551464, + 0.13914057612419128, + 0.26006627082824707 + ], + "credit_map": [ + 0.014127243310213089, + 0.009088033810257912, + 0.019968882203102112, + 0.011622258462011814, + 0.009198615327477455, + 0.008131014183163643, + 0.016700860112905502, + 0.024001391604542732, + 0.0076124039478600025, + 0.005290328990668058, + 0.037300992757081985, + 0.01631041243672371, + 0.006153958849608898, + 0.007930940948426723, + 0.009255774319171906, + 0.036314915865659714, + 0.013681622222065926, + 0.007559873163700104, + 0.013963468372821808, + 0.007967377081513405, + 0.009935062378644943, + 0.012999389320611954, + 0.014250447042286396, + 0.012345600873231888, + 0.011274563148617744, + 0.006123527884483337, + 0.016696585342288017, + 0.010872993618249893, + 0.008277278393507004, + 0.007244122214615345, + 0.023030927404761314, + 0.015227721072733402, + 0.007283132988959551, + 0.0046017952263355255, + 0.00965313334017992, + 0.005039276089519262, + 0.0037774695083498955, + 0.009386538527905941, + 0.006182848941534758, + 0.007071234285831451, + 0.0031428849324584007, + 0.00189490825869143, + 0.00482481112703681, + 0.002703632228076458, + 0.0017122859135270119, + 0.002415246330201626, + 0.0030429549515247345, + 0.002898124046623707, + 0.013952397741377354, + 0.007251448929309845, + 0.009907390922307968, + 0.006218064110726118, + 0.00391816021874547, + 0.0070366389118134975, + 0.0080412020906806, + 0.013361820951104164, + 0.0052145421504974365, + 0.002384266583248973, + 0.004985405132174492, + 0.0019076189491897821, + 0.001441866043023765, + 0.0019173277541995049, + 0.0017160739516839385, + 0.005450074095278978, + 0.016516119241714478, + 0.0061929901130497456, + 0.048148155212402344, + 0.005694075021892786, + 0.005506986752152443, + 0.0046603442169725895, + 0.004628228023648262, + 0.047793690115213394, + 0.0418621189892292, + 0.05721506476402283, + 0.10053129494190216, + 0.012663987465202808, + 0.012719574384391308, + 0.005295109003782272, + 0.013588626869022846, + 0.01619051583111286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "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 main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18598905205726624, + 0.13706308603286743, + 0.03649001568555832, + 0.1674272119998932, + 0.05046284943819046, + 0.0490611307322979, + 0.06314510107040405, + 0.13981452584266663, + 0.10647854208946228, + 0.06406844407320023 + ], + "credit_map": [ + 0.026096299290657043, + 0.012417690828442574, + 0.03363810479640961, + 0.019346853718161583, + 0.01356532797217369, + 0.01421264000236988, + 0.027660643681883812, + 0.03905148431658745, + 0.010192903690040112, + 0.00581526942551136, + 0.043864067643880844, + 0.009476110339164734, + 0.007828595116734505, + 0.00782275851815939, + 0.008289344608783722, + 0.04377402737736702, + 0.0048768226988613605, + 0.0031589593272656202, + 0.006463142111897469, + 0.004909734707325697, + 0.00474111083894968, + 0.0035078974906355143, + 0.003717833897098899, + 0.005114514846354723, + 0.0218175258487463, + 0.011577021330595016, + 0.026295436546206474, + 0.023668630048632622, + 0.014455868862569332, + 0.01447274163365364, + 0.028601739555597305, + 0.02653825655579567, + 0.009583313949406147, + 0.004437986295670271, + 0.012764601036906242, + 0.00422267708927393, + 0.0039411140605807304, + 0.002974672010168433, + 0.004102939274162054, + 0.008435548283159733, + 0.010369597002863884, + 0.00382395857013762, + 0.01426282525062561, + 0.003375187050551176, + 0.0020878587383776903, + 0.0025353911332786083, + 0.0036882872227579355, + 0.00891802553087473, + 0.01304292120039463, + 0.008509963750839233, + 0.007160807028412819, + 0.008278454653918743, + 0.0045119342394173145, + 0.004445837344974279, + 0.006895934231579304, + 0.010299250483512878, + 0.029706869274377823, + 0.017748139798641205, + 0.03130229562520981, + 0.008197552524507046, + 0.005546614993363619, + 0.005934526212513447, + 0.01199454627931118, + 0.02938397414982319, + 0.019638819620013237, + 0.01265997439622879, + 0.020302120596170425, + 0.006228780839592218, + 0.00880595576018095, + 0.007082939613610506, + 0.009175874292850494, + 0.02258407697081566, + 0.021239161491394043, + 0.007092433050274849, + 0.01148555614054203, + 0.004676948767155409, + 0.003695909632369876, + 0.003879016498103738, + 0.005620643030852079, + 0.0063787768594920635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0232852790504694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900968730449677 + ], + "opposite_logprob_tokens": [ + -2.4640097618103027 + ], + "logprob_reward": -0.08900968730449677, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900968730449677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1813480705022812, + 0.08836951851844788, + 0.06998980045318604, + 0.232771635055542, + 0.027125714346766472, + 0.05823177844285965, + 0.04102533310651779, + 0.03798682615160942, + 0.0698118731379509, + 0.1933395117521286 + ], + "credit_map": [ + 0.019551005214452744, + 0.012803169898688793, + 0.0392485037446022, + 0.022145435214042664, + 0.01851656660437584, + 0.014053668826818466, + 0.020465344190597534, + 0.03456435725092888, + 0.010925792157649994, + 0.008846753276884556, + 0.015269283205270767, + 0.013281459920108318, + 0.006855412386357784, + 0.007913157343864441, + 0.011170521378517151, + 0.014107132330536842, + 0.009617934934794903, + 0.007071354426443577, + 0.0118752745911479, + 0.008317442610859871, + 0.00840026792138815, + 0.0059851668775081635, + 0.008673868142068386, + 0.010048484429717064, + 0.03402821719646454, + 0.015292582102119923, + 0.03972585126757622, + 0.028957247734069824, + 0.018437333405017853, + 0.021506251767277718, + 0.03810465708374977, + 0.03671947866678238, + 0.003817906603217125, + 0.003328124526888132, + 0.006803765427321196, + 0.002764591248705983, + 0.0017645840998739004, + 0.0032368788961321115, + 0.0024485145695507526, + 0.002961346646770835, + 0.01301344484090805, + 0.003995287697762251, + 0.017994537949562073, + 0.004529855214059353, + 0.0023142299614846706, + 0.00301670772023499, + 0.004134993068873882, + 0.00923271756619215, + 0.008451297879219055, + 0.005731157958507538, + 0.004724632017314434, + 0.003999694250524044, + 0.0037267636507749557, + 0.0035806617233902216, + 0.004225447773933411, + 0.0065856799483299255, + 0.004760794807225466, + 0.003516438417136669, + 0.010049819946289062, + 0.002849149750545621, + 0.0016912792343646288, + 0.0018275614129379392, + 0.003148497547954321, + 0.010143283754587173, + 0.00881053414195776, + 0.003852879162877798, + 0.018744604662060738, + 0.003530592191964388, + 0.004917827434837818, + 0.004163818433880806, + 0.004116581287235022, + 0.021675029769539833, + 0.03082803264260292, + 0.04382568597793579, + 0.07522039115428925, + 0.0079200379550457, + 0.008512817323207855, + 0.0039040560368448496, + 0.009618922136723995, + 0.013509562239050865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.08900968730449677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, moving slowly and gracefully in an arc around the subject, captures a classic, multi-faceted crystal bottle of Wuliangye, majestically placed upon a sheet of deep, light-absorbing black velvet. Dramatic, high-contrast Rembrandt light, cast with pinpoint precision from the side, carves sharp demarcations of light and dark into the bottle's crisp facets. Simultaneously, a strong rim light from behind outlines the bottle's crystalline rim, highlighting the bottle's iconic red label and the golden \"Wuliangye\" inscription. Crafted with extreme slow-motion particle effects, countless viscous droplets, shimmering like liquid gold, splash in from outside the frame. They gently collide and merge in mid-air, breaking into even smaller, shimmering beads before converging into elegant spirals, like a miniature galaxy, systematically encircling the still bottle. Each flying drop of wine reflects brilliant light, complementing the colorful scattered light and shadow reflected by the crystal bottle. The whole picture, under the warm tones, presents an extremely pure, luxurious and dynamic luxury advertising texture, creating a breathtaking atmosphere of refinement and nobility.", + "seed": 30155, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_6744_seed_30155_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": -0.024902858398854734, + "vlm_logprob": -0.024902858398854734, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.024902858398854734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015542098321020603 + ], + "opposite_logprob_tokens": [ + -4.265542030334473 + ], + "logprob_reward": -0.015542098321020603, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015542098321020603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08569131791591644, + 0.040049340575933456, + 0.05022906884551048, + 0.08734620362520218, + 0.15901459753513336, + 0.09766925871372223, + 0.057730596512556076, + 0.09964755922555923, + 0.15226629376411438, + 0.17035578191280365 + ], + "credit_map": [ + 0.0067804353311657906, + 0.008694646880030632, + 0.008466463536024094, + 0.01072680950164795, + 0.01458559837192297, + 0.01238041091710329, + 0.011411236599087715, + 0.012645711190998554, + 0.004815441090613604, + 0.0051896036602556705, + 0.006609627977013588, + 0.0046698865480721, + 0.004266055300831795, + 0.0034372771624475718, + 0.005427153781056404, + 0.0056342920288443565, + 0.005101984366774559, + 0.0057612028904259205, + 0.008524245582520962, + 0.005240689963102341, + 0.006184522528201342, + 0.004405834246426821, + 0.007322895806282759, + 0.00768769346177578, + 0.006816196721047163, + 0.006741762161254883, + 0.01660582795739174, + 0.013548949733376503, + 0.007351592183113098, + 0.006671634502708912, + 0.012322734110057354, + 0.01728750579059124, + 0.013215853832662106, + 0.020773256197571754, + 0.024138284847140312, + 0.018908610567450523, + 0.022061234340071678, + 0.014416676945984364, + 0.025129409506917, + 0.02037126012146473, + 0.010512948967516422, + 0.00939041469246149, + 0.014027873985469341, + 0.01144928578287363, + 0.013981473632156849, + 0.010248403996229172, + 0.008421698585152626, + 0.019637148827314377, + 0.0060137417167425156, + 0.005269426852464676, + 0.014204700477421284, + 0.0067123351618647575, + 0.007985163480043411, + 0.004449266474694014, + 0.003510122187435627, + 0.009585840627551079, + 0.012888711877167225, + 0.007388732861727476, + 0.013095430098474026, + 0.015815606340765953, + 0.01525779440999031, + 0.006913277320563793, + 0.009698229841887951, + 0.018589774146676064, + 0.021765222772955894, + 0.012981653213500977, + 0.01618019863963127, + 0.020385006442666054, + 0.023146647959947586, + 0.022073155269026756, + 0.016637425869703293, + 0.019096964970231056, + 0.019050652161240578, + 0.009622784331440926, + 0.040209002792835236, + 0.009782413952052593, + 0.014002092182636261, + 0.010235139168798923, + 0.007114320993423462, + 0.06033936142921448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.015542098321020603 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17406556010246277, + 0.06634273380041122, + 0.07208855450153351, + 0.08189312368631363, + 0.18013493716716766, + 0.11287862807512283, + 0.04259295389056206, + 0.05855034291744232, + 0.1648188978433609, + 0.0466342493891716 + ], + "credit_map": [ + 0.017104966565966606, + 0.018218757584691048, + 0.02956719696521759, + 0.016129285097122192, + 0.017496487125754356, + 0.0268569253385067, + 0.02616891637444496, + 0.022523021325469017, + 0.00897969026118517, + 0.007292908616364002, + 0.01458010170608759, + 0.006193573586642742, + 0.006199887953698635, + 0.005637230817228556, + 0.0071202111430466175, + 0.010339133441448212, + 0.007263218984007835, + 0.0072845639660954475, + 0.011377692222595215, + 0.0070495340041816235, + 0.008815542794764042, + 0.006112462375313044, + 0.013325748965144157, + 0.010859793052077293, + 0.012822093442082405, + 0.007987158372998238, + 0.013744167983531952, + 0.007325830403715372, + 0.007672714535146952, + 0.00846434198319912, + 0.009671645238995552, + 0.01420516986399889, + 0.014996405690908432, + 0.029998382553458214, + 0.023398641496896744, + 0.020673640072345734, + 0.029414070770144463, + 0.017032062634825706, + 0.02202083170413971, + 0.02260090410709381, + 0.014102659188210964, + 0.009134930558502674, + 0.019864706322550774, + 0.013792566955089569, + 0.010582080110907555, + 0.01675698533654213, + 0.011794851161539555, + 0.01684984751045704, + 0.0037215505726635456, + 0.003072944702580571, + 0.01149293128401041, + 0.005871408618986607, + 0.0037173901218920946, + 0.003464110428467393, + 0.0032866669353097677, + 0.00796594936400652, + 0.008907811716198921, + 0.0035719003062695265, + 0.007743895053863525, + 0.008110705763101578, + 0.005534524563699961, + 0.005465052556246519, + 0.005074252840131521, + 0.014142203144729137, + 0.02315172739326954, + 0.011547413654625416, + 0.025983722880482674, + 0.019797049462795258, + 0.023168424144387245, + 0.017820890992879868, + 0.017040781676769257, + 0.02630888670682907, + 0.006057353690266609, + 0.0038902361411601305, + 0.0033590123057365417, + 0.00484516890719533, + 0.006996657233685255, + 0.006752277258783579, + 0.006609100382775068, + 0.008124448359012604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350608095526695 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350608095526695, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350608095526695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10049383342266083, + 0.08995203673839569, + 0.12374338507652283, + 0.08191798627376556, + 0.09005335718393326, + 0.053237173706293106, + 0.08726539462804794, + 0.09037452936172485, + 0.0662265419960022, + 0.2167358249425888 + ], + "credit_map": [ + 0.006331378128379583, + 0.00796311255544424, + 0.017070017755031586, + 0.01321365311741829, + 0.010866165161132812, + 0.014395840466022491, + 0.017535250633955002, + 0.013118403032422066, + 0.01023335661739111, + 0.010438673198223114, + 0.01102275587618351, + 0.008306711912155151, + 0.013590142130851746, + 0.012434162199497223, + 0.011862924322485924, + 0.012063300237059593, + 0.007931673899292946, + 0.00985340029001236, + 0.017376013100147247, + 0.012198039330542088, + 0.01789197139441967, + 0.009800133295357227, + 0.028541840612888336, + 0.020150311291217804, + 0.011054501868784428, + 0.007186056114733219, + 0.015532602556049824, + 0.009520363993942738, + 0.006892810575664043, + 0.006552507635205984, + 0.009495874866843224, + 0.015683263540267944, + 0.00786187406629324, + 0.013966234400868416, + 0.011771532706916332, + 0.01000678539276123, + 0.016307320445775986, + 0.00830578338354826, + 0.007665904704481363, + 0.01416791696101427, + 0.00603820662945509, + 0.004764279816299677, + 0.007969308644533157, + 0.0060836016200482845, + 0.008779553696513176, + 0.005174103192985058, + 0.005935393273830414, + 0.008492721244692802, + 0.006874497048556805, + 0.00406288355588913, + 0.02655360847711563, + 0.010152360424399376, + 0.005105082411319017, + 0.006641417741775513, + 0.006102973595261574, + 0.021772563457489014, + 0.011539715342223644, + 0.005347436759620905, + 0.011715438216924667, + 0.015749461948871613, + 0.008037295192480087, + 0.007181259803473949, + 0.007701126858592033, + 0.023102791979908943, + 0.00949655007570982, + 0.004982477519661188, + 0.010866404511034489, + 0.008359664119780064, + 0.010239681228995323, + 0.0052073621191084385, + 0.00630061374977231, + 0.010773781687021255, + 0.01986604556441307, + 0.013428444042801857, + 0.06078468635678291, + 0.01151700597256422, + 0.012281308881938457, + 0.010473724454641342, + 0.008524350821971893, + 0.07986024767160416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.026350608095526695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14842656254768372, + 0.03861849755048752, + 0.061583057045936584, + 0.07890712469816208, + 0.10990149527788162, + 0.12275151163339615, + 0.040931135416030884, + 0.1489349603652954, + 0.12616196274757385, + 0.1237836703658104 + ], + "credit_map": [ + 0.014571095816791058, + 0.013779306784272194, + 0.029906675219535828, + 0.019386088475584984, + 0.014122743159532547, + 0.016819540411233902, + 0.020208677276968956, + 0.019632449373602867, + 0.005541936960071325, + 0.004896980244666338, + 0.006026361603289843, + 0.0032839777413755655, + 0.004353504627943039, + 0.003980553708970547, + 0.0052160839550197124, + 0.0053190975449979305, + 0.00496846716850996, + 0.0043016760610044, + 0.009525303728878498, + 0.006853634491562843, + 0.00979616679251194, + 0.00587432598695159, + 0.010695472359657288, + 0.009568008594214916, + 0.006159430369734764, + 0.0059293461963534355, + 0.01365877315402031, + 0.009421023540198803, + 0.007859543897211552, + 0.008987067267298698, + 0.011818507686257362, + 0.015073427930474281, + 0.00946660153567791, + 0.011433755978941917, + 0.018124327063560486, + 0.012346605770289898, + 0.012557349167764187, + 0.011456443928182125, + 0.01848364993929863, + 0.01603275164961815, + 0.013574982061982155, + 0.010050030425190926, + 0.0193083006888628, + 0.012564392760396004, + 0.01972809247672558, + 0.013797297142446041, + 0.013344569131731987, + 0.020383846014738083, + 0.0032544343266636133, + 0.0028378558345139027, + 0.00955580547451973, + 0.007519829086959362, + 0.004180057905614376, + 0.0033309420105069876, + 0.002306111855432391, + 0.007946097292006016, + 0.022480575367808342, + 0.0113685242831707, + 0.022647816687822342, + 0.022615937516093254, + 0.01382304634898901, + 0.014855909161269665, + 0.013811950571835041, + 0.027331199496984482, + 0.01763143576681614, + 0.00907690729945898, + 0.01643526740372181, + 0.015904894098639488, + 0.019817927852272987, + 0.01862848922610283, + 0.011990780010819435, + 0.016676252707839012, + 0.01274143997579813, + 0.010789458639919758, + 0.009559426456689835, + 0.022078359499573708, + 0.014815541915595531, + 0.0099173653870821, + 0.011227463372051716, + 0.032654616981744766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808284178376198 + ], + "opposite_logprob_tokens": [ + -3.2880828380584717 + ], + "logprob_reward": -0.03808284178376198, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808284178376198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14188700914382935, + 0.14437933266162872, + 0.07678760588169098, + 0.11613491922616959, + 0.09147045016288757, + 0.07553967833518982, + 0.06440135091543198, + 0.05744262412190437, + 0.03738750144839287, + 0.19456960260868073 + ], + "credit_map": [ + 0.010641727596521378, + 0.010550182312726974, + 0.0240677110850811, + 0.02066872827708721, + 0.01237866748124361, + 0.019143210723996162, + 0.023654745891690254, + 0.020782025530934334, + 0.011503429152071476, + 0.012232327833771706, + 0.021246114745736122, + 0.01478855311870575, + 0.026101481169462204, + 0.01190321333706379, + 0.024716412648558617, + 0.02188778482377529, + 0.005973991937935352, + 0.0060536861419677734, + 0.011142249219119549, + 0.008724641054868698, + 0.008983919396996498, + 0.005820240825414658, + 0.01690276525914669, + 0.013186103664338589, + 0.011753314174711704, + 0.008403056301176548, + 0.02123195491731167, + 0.013854211196303368, + 0.008805408142507076, + 0.010169044137001038, + 0.01739887334406376, + 0.0245190542191267, + 0.007810675073415041, + 0.012753169052302837, + 0.015406938269734383, + 0.010532635264098644, + 0.011743787676095963, + 0.0084851598367095, + 0.012064144015312195, + 0.01267392747104168, + 0.010333760641515255, + 0.006093909032642841, + 0.013128583319485188, + 0.008241023868322372, + 0.009533757343888283, + 0.006138816941529512, + 0.006285109091550112, + 0.01578471064567566, + 0.007005398161709309, + 0.004236398730427027, + 0.015126110054552555, + 0.006678012199699879, + 0.0056746648624539375, + 0.00666820490732789, + 0.004117181990295649, + 0.014895372092723846, + 0.0085560642182827, + 0.004064574372023344, + 0.008275792002677917, + 0.007221783511340618, + 0.006101695820689201, + 0.0042930226773023605, + 0.006175910588353872, + 0.012753769755363464, + 0.004185089375823736, + 0.004091392736881971, + 0.004728876985609531, + 0.004920388571918011, + 0.005646490957587957, + 0.003880595089867711, + 0.0029736561700701714, + 0.006961006671190262, + 0.017323317006230354, + 0.010589835233986378, + 0.062160227447748184, + 0.012784571386873722, + 0.008197562769055367, + 0.008221347816288471, + 0.004783432465046644, + 0.07050929218530655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.03808284178376198 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, moving slowly and gracefully in an arc around the subject, captures a classic, multi-faceted crystal bottle of Wuliangye, majestically placed upon a sheet of deep, light-absorbing black velvet. Dramatic, high-contrast Rembrandt light, cast with pinpoint precision from the side, carves sharp demarcations of light and dark into the bottle's crisp facets. Simultaneously, a strong rim light from behind outlines the bottle's crystalline rim, highlighting the bottle's iconic red label and the golden \"Wuliangye\" inscription. Crafted with extreme slow-motion particle effects, countless viscous droplets, shimmering like liquid gold, splash in from outside the frame. They gently collide and merge in mid-air, breaking into even smaller, shimmering beads before converging into elegant spirals, like a miniature galaxy, systematically encircling the still bottle. Each flying drop of wine reflects brilliant light, complementing the colorful scattered light and shadow reflected by the crystal bottle. The whole picture, under the warm tones, presents an extremely pure, luxurious and dynamic luxury advertising texture, creating a breathtaking atmosphere of refinement and nobility.", + "seed": 30156, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_6744_seed_30156_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": -0.04013593960553408, + "vlm_logprob": -0.04013593960553408, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.04013593960553408, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006825583986938 + ], + "opposite_logprob_tokens": [ + -5.6318254470825195 + ], + "logprob_reward": -0.006825583986938, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006825583986938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10128605365753174, + 0.0660889521241188, + 0.028145521879196167, + 0.06091194972395897, + 0.06707988679409027, + 0.3803060054779053, + 0.07056979089975357, + 0.14379729330539703, + 0.03336498513817787, + 0.0484495609998703 + ], + "credit_map": [ + 0.012272033840417862, + 0.012323607690632343, + 0.01588786579668522, + 0.013288161717355251, + 0.013938762247562408, + 0.010396646335721016, + 0.010317232459783554, + 0.012861744500696659, + 0.011314134113490582, + 0.009627459570765495, + 0.005306052044034004, + 0.009974864311516285, + 0.00872467365115881, + 0.00662051560357213, + 0.007079218979924917, + 0.007442038506269455, + 0.005516835488379002, + 0.0023553678765892982, + 0.005069193430244923, + 0.0021195346489548683, + 0.0022369916550815105, + 0.0060629588551819324, + 0.0012659530621021986, + 0.0035186861641705036, + 0.010933292098343372, + 0.003400571644306183, + 0.009412878192961216, + 0.003795359982177615, + 0.004268412943929434, + 0.012518053874373436, + 0.003877087030559778, + 0.012706292793154716, + 0.005554905626922846, + 0.009256360121071339, + 0.006452929228544235, + 0.007202822715044022, + 0.01571517623960972, + 0.010320349596440792, + 0.006919551640748978, + 0.005657791625708342, + 0.06281333416700363, + 0.024486001580953598, + 0.06865812093019485, + 0.034558799117803574, + 0.026064401492476463, + 0.05865615978837013, + 0.02495402842760086, + 0.08011516183614731, + 0.00889682024717331, + 0.005267924629151821, + 0.007160342298448086, + 0.008876051753759384, + 0.007008499000221491, + 0.010906201787292957, + 0.01031317375600338, + 0.012140778824687004, + 0.030424097552895546, + 0.009237511083483696, + 0.01796688325703144, + 0.011656727641820908, + 0.01710362732410431, + 0.020349806174635887, + 0.01390625350177288, + 0.023152384907007217, + 0.004188999067991972, + 0.0039489902555942535, + 0.0038712588138878345, + 0.0046969507820904255, + 0.005736088380217552, + 0.003973609302192926, + 0.002871287753805518, + 0.004077799152582884, + 0.006149211432784796, + 0.004937644116580486, + 0.004489158745855093, + 0.010300622321665287, + 0.006114092655479908, + 0.006418792996555567, + 0.0061296215280890465, + 0.003910416271537542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.006825583986938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.08784955739974976, + 0.0822589322924614, + 0.07723836600780487, + 0.12016963213682175, + 0.08570986241102219, + 0.1191084086894989, + 0.06622444838285446, + 0.17112645506858826, + 0.08283445984125137, + 0.10747986286878586 + ], + "credit_map": [ + 0.0128727275878191, + 0.009140157140791416, + 0.015090574510395527, + 0.01084583718329668, + 0.009980596601963043, + 0.011247736401855946, + 0.008068455383181572, + 0.010603473521769047, + 0.013860628008842468, + 0.0070548891089856625, + 0.01320628821849823, + 0.007478374522179365, + 0.0074735297821462154, + 0.01546746026724577, + 0.005414232145994902, + 0.012303536757826805, + 0.015304743312299252, + 0.00626636715605855, + 0.012645400129258633, + 0.006098995916545391, + 0.007484287954866886, + 0.013377109542489052, + 0.0048953634686768055, + 0.011166100390255451, + 0.020759111270308495, + 0.009924829006195068, + 0.015504397451877594, + 0.008050169795751572, + 0.011571732349693775, + 0.023665713146328926, + 0.007267212960869074, + 0.023426469415426254, + 0.010632327757775784, + 0.010914206504821777, + 0.010475720278918743, + 0.007668022997677326, + 0.015376567840576172, + 0.013680377043783665, + 0.00668112188577652, + 0.010281512513756752, + 0.016694294288754463, + 0.008831692859530449, + 0.02832937240600586, + 0.007365673780441284, + 0.009156165644526482, + 0.019637377932667732, + 0.005585885141044855, + 0.02350795641541481, + 0.00699078431352973, + 0.0033752473536878824, + 0.005920799449086189, + 0.004274051170796156, + 0.0030482274014502764, + 0.015240643173456192, + 0.004176294431090355, + 0.023198403418064117, + 0.03863026574254036, + 0.009276261553168297, + 0.032388173043727875, + 0.010725579224526882, + 0.012198037467896938, + 0.02661300078034401, + 0.013961541466414928, + 0.02733360230922699, + 0.011552698910236359, + 0.00650458550080657, + 0.010043559595942497, + 0.012341964058578014, + 0.009716685861349106, + 0.010608254000544548, + 0.008859489113092422, + 0.013207216747105122, + 0.020765548571944237, + 0.009735969826579094, + 0.012795072980225086, + 0.006968543399125338, + 0.009039531461894512, + 0.017437295988202095, + 0.009930831380188465, + 0.020807063207030296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.020587211474776268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370801359415054 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.03370801359415054, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370801359415054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1582852452993393, + 0.07093396037817001, + 0.04463419318199158, + 0.05925322324037552, + 0.03961486369371414, + 0.2572575807571411, + 0.10153640061616898, + 0.10592135041952133, + 0.08844122290611267, + 0.07412195950746536 + ], + "credit_map": [ + 0.02734370529651642, + 0.013520326465368271, + 0.030786026269197464, + 0.014325498603284359, + 0.011339863762259483, + 0.025589976459741592, + 0.01047183945775032, + 0.024908028542995453, + 0.012445642612874508, + 0.010736916214227676, + 0.0058299340307712555, + 0.009807509370148182, + 0.010867999866604805, + 0.006124497391283512, + 0.007615355309098959, + 0.007506109774112701, + 0.007685323711484671, + 0.00409043999388814, + 0.008103576488792896, + 0.0032849563285708427, + 0.005079585127532482, + 0.006555289961397648, + 0.003248022636398673, + 0.006587001960724592, + 0.011440211907029152, + 0.003590349107980728, + 0.008732819929718971, + 0.004051874857395887, + 0.004580419510602951, + 0.011890602298080921, + 0.0040706368163228035, + 0.010896312072873116, + 0.006671967916190624, + 0.004756568931043148, + 0.0050334990955889225, + 0.0027503797318786383, + 0.005556427873671055, + 0.00573463412001729, + 0.0033216869924217463, + 0.005789700895547867, + 0.046417269855737686, + 0.01691429875791073, + 0.044232796877622604, + 0.02154023014008999, + 0.015738599002361298, + 0.040238156914711, + 0.017716791480779648, + 0.05445944890379906, + 0.011026175692677498, + 0.00738571910187602, + 0.01108583528548479, + 0.011768420226871967, + 0.012174157425761223, + 0.016911741346120834, + 0.01215361151844263, + 0.019030744209885597, + 0.023895148187875748, + 0.0054784975945949554, + 0.016014335677027702, + 0.008723311126232147, + 0.008152459748089314, + 0.017154376953840256, + 0.011377223767340183, + 0.0151260020211339, + 0.016051026061177254, + 0.0052930437959730625, + 0.014632903970777988, + 0.007901357486844063, + 0.005175415426492691, + 0.013202995993196964, + 0.006114153191447258, + 0.020070327445864677, + 0.009473386220633984, + 0.010069087147712708, + 0.007098379544913769, + 0.005983104929327965, + 0.010196744464337826, + 0.009986904449760914, + 0.0075418842025101185, + 0.01377247367054224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.03370801359415054 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538649141788483 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.012538649141788483, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538649141788483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26406314969062805, + 0.061837077140808105, + 0.030375542119145393, + 0.036659032106399536, + 0.07782844454050064, + 0.20871268212795258, + 0.11504928767681122, + 0.07441714406013489, + 0.05014292150735855, + 0.08091465383768082 + ], + "credit_map": [ + 0.04258374124765396, + 0.01974034495651722, + 0.04977928102016449, + 0.028839103877544403, + 0.021793244406580925, + 0.04230814427137375, + 0.018709756433963776, + 0.04030952230095863, + 0.010774143040180206, + 0.006444298196583986, + 0.009213232435286045, + 0.0057568917982280254, + 0.0069370390847325325, + 0.009399894624948502, + 0.00485315453261137, + 0.008458422496914864, + 0.005052551627159119, + 0.0035486631095409393, + 0.0037136413156986237, + 0.0025494517758488655, + 0.0037888388615101576, + 0.005135881714522839, + 0.002660072408616543, + 0.003926441073417664, + 0.005221411120146513, + 0.0038548754528164864, + 0.004636423662304878, + 0.0046116383746266365, + 0.0035172400530427694, + 0.004914070013910532, + 0.0028535707388073206, + 0.0070498003624379635, + 0.009221471846103668, + 0.005529654212296009, + 0.013018517754971981, + 0.006452769972383976, + 0.007185884285718203, + 0.01499117910861969, + 0.004216493107378483, + 0.01721246913075447, + 0.038477133959531784, + 0.01582111231982708, + 0.030706999823451042, + 0.02139570750296116, + 0.015245496295392513, + 0.03107423149049282, + 0.01637745089828968, + 0.039614543318748474, + 0.011539879254996777, + 0.008985421620309353, + 0.013659839518368244, + 0.015230745077133179, + 0.014461461454629898, + 0.019631436094641685, + 0.012958677485585213, + 0.018581826239824295, + 0.01149848010390997, + 0.0038138951640576124, + 0.010557413101196289, + 0.00577483419328928, + 0.007312454748898745, + 0.011331398971378803, + 0.006150940433144569, + 0.017977727577090263, + 0.007371231913566589, + 0.008160033263266087, + 0.005189361982047558, + 0.004093033727258444, + 0.007702849805355072, + 0.006545701529830694, + 0.005119142588227987, + 0.005961568560451269, + 0.011226828210055828, + 0.008705712854862213, + 0.008775523863732815, + 0.0064118122681975365, + 0.010104049928486347, + 0.012206262908875942, + 0.01082488615065813, + 0.012659572064876556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.012538649141788483 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270202398300171 + ], + "opposite_logprob_tokens": [ + -2.1270203590393066 + ], + "logprob_reward": -0.1270202398300171, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270202398300171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06777633726596832, + 0.08405573666095734, + 0.050174422562122345, + 0.05626238137483597, + 0.11792702227830887, + 0.07910767197608948, + 0.215378999710083, + 0.16063149273395538, + 0.06692606955766678, + 0.10175985842943192 + ], + "credit_map": [ + 0.011133627966046333, + 0.0070436010137200356, + 0.009730519726872444, + 0.010790286585688591, + 0.006631289143115282, + 0.00796652864664793, + 0.004944946151226759, + 0.00953554268926382, + 0.01514016930013895, + 0.011616075411438942, + 0.0069452207535505295, + 0.012850373983383179, + 0.012285224162042141, + 0.007343975827097893, + 0.009897538460791111, + 0.007977153174579144, + 0.00878969021141529, + 0.007123093120753765, + 0.005238566547632217, + 0.004098199307918549, + 0.00707728648558259, + 0.007491246331483126, + 0.004048715345561504, + 0.006307628471404314, + 0.008965362794697285, + 0.004885081201791763, + 0.008286985568702221, + 0.004976441618055105, + 0.005457150749862194, + 0.010214343667030334, + 0.0035967431031167507, + 0.009880274534225464, + 0.010733882896602154, + 0.011810589581727982, + 0.01611664891242981, + 0.011284754611551762, + 0.017284894362092018, + 0.02143779955804348, + 0.00846656784415245, + 0.02079189382493496, + 0.011303840205073357, + 0.007366510108113289, + 0.01463431864976883, + 0.006296464707702398, + 0.009096035733819008, + 0.01158720813691616, + 0.006695954594761133, + 0.012127331458032131, + 0.023488862439990044, + 0.018121963366866112, + 0.025576621294021606, + 0.028376059606671333, + 0.029970107600092888, + 0.03312632441520691, + 0.027509290724992752, + 0.029209773987531662, + 0.03026358224451542, + 0.0072282301262021065, + 0.023636547848582268, + 0.013057993724942207, + 0.009826943278312683, + 0.027106767520308495, + 0.01580137573182583, + 0.0337100550532341, + 0.0097119165584445, + 0.006394189782440662, + 0.008925184607505798, + 0.007441582623869181, + 0.008534388616681099, + 0.009225461632013321, + 0.005730864591896534, + 0.010962487198412418, + 0.013493338599801064, + 0.014088338240981102, + 0.010322176851332188, + 0.007774874102324247, + 0.013919858261942863, + 0.013421525247395039, + 0.013287557289004326, + 0.015452196821570396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 4 + ], + "selected_reward": -0.1270202398300171 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, moving slowly and gracefully in an arc around the subject, captures a classic, multi-faceted crystal bottle of Wuliangye, majestically placed upon a sheet of deep, light-absorbing black velvet. Dramatic, high-contrast Rembrandt light, cast with pinpoint precision from the side, carves sharp demarcations of light and dark into the bottle's crisp facets. Simultaneously, a strong rim light from behind outlines the bottle's crystalline rim, highlighting the bottle's iconic red label and the golden \"Wuliangye\" inscription. Crafted with extreme slow-motion particle effects, countless viscous droplets, shimmering like liquid gold, splash in from outside the frame. They gently collide and merge in mid-air, breaking into even smaller, shimmering beads before converging into elegant spirals, like a miniature galaxy, systematically encircling the still bottle. Each flying drop of wine reflects brilliant light, complementing the colorful scattered light and shadow reflected by the crystal bottle. The whole picture, under the warm tones, presents an extremely pure, luxurious and dynamic luxury advertising texture, creating a breathtaking atmosphere of refinement and nobility.", + "seed": 30157, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_12/index_6744_seed_30157_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": -0.029497401043772697, + "vlm_logprob": -0.029497401043772697, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.029497401043772697, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014761041849851608 + ], + "opposite_logprob_tokens": [ + -4.389760971069336 + ], + "logprob_reward": -0.014761041849851608, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014761041849851608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30071043968200684, + 0.13596054911613464, + 0.04885948821902275, + 0.09759590029716492, + 0.21558786928653717, + 0.03976529464125633, + 0.07358352094888687, + 0.03308555856347084, + 0.0349334180355072, + 0.019917979836463928 + ], + "credit_map": [ + 0.03175657242536545, + 0.01804412715137005, + 0.07153485715389252, + 0.0180347952991724, + 0.018609721213579178, + 0.07347601652145386, + 0.014550763182342052, + 0.05470358207821846, + 0.007688106037676334, + 0.0033646670635789633, + 0.008606002666056156, + 0.004118545912206173, + 0.004679241683334112, + 0.05190948396921158, + 0.0064428118057549, + 0.04915168508887291, + 0.00781295821070671, + 0.003941812552511692, + 0.005261162295937538, + 0.003984661307185888, + 0.004465991631150246, + 0.009482610039412975, + 0.004421386867761612, + 0.009488905780017376, + 0.019653750583529472, + 0.009740431793034077, + 0.0164270531386137, + 0.006405947729945183, + 0.00830529723316431, + 0.017038144171237946, + 0.005534110590815544, + 0.014491172507405281, + 0.04302679002285004, + 0.018217233940958977, + 0.054830778390169144, + 0.014214637689292431, + 0.01923530548810959, + 0.02712942659854889, + 0.011039589531719685, + 0.02789410948753357, + 0.008971137925982475, + 0.004544353578239679, + 0.006164207123219967, + 0.003257656702771783, + 0.0034931644331663847, + 0.004384399391710758, + 0.0031176854390650988, + 0.005832691676914692, + 0.0073015810921788216, + 0.00397670641541481, + 0.006937860976904631, + 0.004322620574384928, + 0.007691618055105209, + 0.023243507370352745, + 0.0038066513370722532, + 0.01630297303199768, + 0.004231925588101149, + 0.003695568535476923, + 0.004100367426872253, + 0.004063863772898912, + 0.004179417621344328, + 0.003604704514145851, + 0.004860508255660534, + 0.004349201917648315, + 0.00655267434194684, + 0.00359263620339334, + 0.005463399924337864, + 0.0027567583601921797, + 0.00249862065538764, + 0.006457315292209387, + 0.002519309287890792, + 0.005092706996947527, + 0.0026164837181568146, + 0.0023927162401378155, + 0.0021810957696288824, + 0.0015340783866122365, + 0.0023595436941832304, + 0.004016964230686426, + 0.0017280095489695668, + 0.003089088248088956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014761041849851608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980283461511135 + ], + "opposite_logprob_tokens": [ + -3.5298027992248535 + ], + "logprob_reward": -0.02980283461511135, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980283461511135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10820526629686356, + 0.18993598222732544, + 0.06815337389707565, + 0.03063254989683628, + 0.2505979835987091, + 0.05705950781702995, + 0.05757099390029907, + 0.02527102641761303, + 0.17551706731319427, + 0.03705620765686035 + ], + "credit_map": [ + 0.01325253676623106, + 0.005537714343518019, + 0.03333912789821625, + 0.0076233381405472755, + 0.003957835491746664, + 0.02127840556204319, + 0.0051459576934576035, + 0.018070362508296967, + 0.015346446074545383, + 0.005781780928373337, + 0.012339659035205841, + 0.0059442538768053055, + 0.00851434376090765, + 0.0719008520245552, + 0.005825892090797424, + 0.06428276002407074, + 0.011712056584656239, + 0.006190601736307144, + 0.012652140110731125, + 0.006311978679150343, + 0.005422912072390318, + 0.011415556073188782, + 0.004949529189616442, + 0.009498609229922295, + 0.004653989337384701, + 0.0019462479976937175, + 0.006120417732745409, + 0.0018283325480297208, + 0.0024378278758376837, + 0.0065235719084739685, + 0.0019981807563453913, + 0.005123984068632126, + 0.05958005040884018, + 0.012341328896582127, + 0.07169333845376968, + 0.016914905980229378, + 0.010491489432752132, + 0.03231044113636017, + 0.009576704353094101, + 0.03768973797559738, + 0.00927363708615303, + 0.004840945824980736, + 0.0152970552444458, + 0.006199153605848551, + 0.0037332214415073395, + 0.006117368116974831, + 0.004206414334475994, + 0.007391719147562981, + 0.004771736916154623, + 0.0024604452773928642, + 0.0040702358819544315, + 0.0018935416592285037, + 0.00490705668926239, + 0.022278107702732086, + 0.002643743995577097, + 0.01454613171517849, + 0.003672168590128422, + 0.006038685794919729, + 0.0026323790661990643, + 0.0023236635606735945, + 0.0029463996179401875, + 0.0022767374757677317, + 0.002415126422420144, + 0.002965868916362524, + 0.028893642127513885, + 0.01266456302255392, + 0.019359393045306206, + 0.013446054421365261, + 0.00958467647433281, + 0.05582838132977486, + 0.010621176101267338, + 0.025119207799434662, + 0.007282746955752373, + 0.0024229446426033974, + 0.005992165766656399, + 0.0018129724776372313, + 0.003121191868558526, + 0.006208065431565046, + 0.0032463509123772383, + 0.006969770882278681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 8 + ], + "selected_reward": -0.02980283461511135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370651230216026 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370651230216026, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370651230216026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09657156467437744, + 0.10297881066799164, + 0.08066678792238235, + 0.045348457992076874, + 0.4444343149662018, + 0.05631326511502266, + 0.014387187547981739, + 0.02027158997952938, + 0.0584065318107605, + 0.08062155544757843 + ], + "credit_map": [ + 0.013406587764620781, + 0.008823996409773827, + 0.026422720402479172, + 0.006779648829251528, + 0.007761070504784584, + 0.013484873808920383, + 0.00853397324681282, + 0.011358692310750484, + 0.01527811586856842, + 0.005532114300876856, + 0.018646815791726112, + 0.007207498885691166, + 0.0059788464568555355, + 0.022038426250219345, + 0.006113995797932148, + 0.02218298241496086, + 0.012057620100677013, + 0.005838156212121248, + 0.014113356359302998, + 0.005232977215200663, + 0.005390731617808342, + 0.015932392328977585, + 0.006712133996188641, + 0.015389415435492992, + 0.005957425571978092, + 0.004587527364492416, + 0.01044411864131689, + 0.0028419673908501863, + 0.004864160902798176, + 0.007583038881421089, + 0.003111403901129961, + 0.00595881137996912, + 0.1029326319694519, + 0.02040315419435501, + 0.12222323566675186, + 0.029467428103089333, + 0.016521286219358444, + 0.06488151103258133, + 0.017125118523836136, + 0.07087992131710052, + 0.01044542621821165, + 0.005864753387868404, + 0.008543863892555237, + 0.004675849340856075, + 0.005227573215961456, + 0.007144700735807419, + 0.004849228076636791, + 0.009561868384480476, + 0.0020776656456291676, + 0.0010504492092877626, + 0.0019074375741183758, + 0.0011647404171526432, + 0.0014207252534106374, + 0.0034184374380856752, + 0.0011348960688337684, + 0.0022128354758024216, + 0.0034460409078747034, + 0.002955234842374921, + 0.0029201279394328594, + 0.0017293986165896058, + 0.0020040597300976515, + 0.0030253983568400145, + 0.0016166152199730277, + 0.0025747132021933794, + 0.007742270361632109, + 0.003486817702651024, + 0.00999003741890192, + 0.004236155655235052, + 0.003154337638989091, + 0.01508314535021782, + 0.004502803552895784, + 0.010210960172116756, + 0.012017392553389072, + 0.006691108923405409, + 0.008141639642417431, + 0.0043508983217179775, + 0.0062509384006261826, + 0.02567105181515217, + 0.005070483777672052, + 0.012428034096956253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.03370651230216026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581256598234177 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581256598234177, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581256598234177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08820607513189316, + 0.08419158309698105, + 0.03769592568278313, + 0.05991921201348305, + 0.1941317915916443, + 0.07739272713661194, + 0.22775764763355255, + 0.08039893954992294, + 0.13293689489364624, + 0.01736922562122345 + ], + "credit_map": [ + 0.014509807340800762, + 0.004608886782079935, + 0.02334921434521675, + 0.008326386101543903, + 0.0059396508149802685, + 0.014579751528799534, + 0.005286330357193947, + 0.011606050655245781, + 0.011637105606496334, + 0.003202069317921996, + 0.008349575102329254, + 0.004114771261811256, + 0.004112818744033575, + 0.024354448541998863, + 0.0035993333440274, + 0.024821460247039795, + 0.006750612054020166, + 0.001880934927612543, + 0.0032504815608263016, + 0.002659559017047286, + 0.0022970132995396852, + 0.009408670477569103, + 0.0028604487888514996, + 0.008588207885622978, + 0.01513031218200922, + 0.0043542226776480675, + 0.010632822290062904, + 0.0028436703141778708, + 0.0033634027931839228, + 0.011109800077974796, + 0.002612947253510356, + 0.009872033260762691, + 0.04504102095961571, + 0.009000455029308796, + 0.055139005184173584, + 0.013586492277681828, + 0.007624991238117218, + 0.025901393964886665, + 0.007691771723330021, + 0.030146658420562744, + 0.014162193983793259, + 0.00828479416668415, + 0.013166632503271103, + 0.007013232912868261, + 0.006557227578014135, + 0.008075771853327751, + 0.006706221029162407, + 0.013426657766103745, + 0.0286041758954525, + 0.009615660645067692, + 0.010899992659687996, + 0.006017944775521755, + 0.01968604326248169, + 0.09566710889339447, + 0.007659506052732468, + 0.04960722476243973, + 0.01445415522903204, + 0.00867803581058979, + 0.013637969270348549, + 0.006899557076394558, + 0.006228328216820955, + 0.011684594675898552, + 0.006511776242405176, + 0.012304522097110748, + 0.017912695184350014, + 0.013718914240598679, + 0.028856104239821434, + 0.011161301285028458, + 0.009189070202410221, + 0.017399216070771217, + 0.011995222419500351, + 0.02270437218248844, + 0.0031349575147032738, + 0.0019250576151534915, + 0.0021563179325312376, + 0.0014826005790382624, + 0.0025535861495882273, + 0.0024709394201636314, + 0.001773342490196228, + 0.0018724229885265231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.020581256598234177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a multi-faceted crystal bottle with a red label and golden 'Wuliangye' inscription?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the bottle placed upon a sheet of deep, light-absorbing black velvet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene lit with high-contrast Rembrandt lighting from the side and a strong rim light from behind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless viscous droplets resembling liquid gold splash in from outside the frame and encircle the bottle in spirals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera moving slowly and gracefully in an arc around the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863535985350609 + ], + "opposite_logprob_tokens": [ + -3.048635244369507 + ], + "logprob_reward": -0.04863535985350609, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863535985350609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2045995146036148, + 0.13625302910804749, + 0.06294532865285873, + 0.08893997967243195, + 0.13697947561740875, + 0.036456335335969925, + 0.1034531369805336, + 0.029876500368118286, + 0.13141021132469177, + 0.0690864771604538 + ], + "credit_map": [ + 0.026559393852949142, + 0.009755690582096577, + 0.07081384211778641, + 0.016188062727451324, + 0.011785914190113544, + 0.030503086745738983, + 0.015480202622711658, + 0.023513320833444595, + 0.016298042610287666, + 0.005083919502794743, + 0.015663469210267067, + 0.007644321769475937, + 0.0053298259153962135, + 0.0382106713950634, + 0.0066253505647182465, + 0.04139743372797966, + 0.014573727734386921, + 0.005729516502469778, + 0.007889185100793839, + 0.004548725206404924, + 0.005507117137312889, + 0.00912003219127655, + 0.0058958460576832294, + 0.009681176394224167, + 0.01549606304615736, + 0.007650131359696388, + 0.015933990478515625, + 0.004688338376581669, + 0.009631064720451832, + 0.016902852803468704, + 0.004554117098450661, + 0.01408341433852911, + 0.026036452502012253, + 0.010446154512465, + 0.03213721141219139, + 0.008302642963826656, + 0.010574082843959332, + 0.023092275485396385, + 0.006930130999535322, + 0.01946052722632885, + 0.005068606697022915, + 0.004041577689349651, + 0.004991205874830484, + 0.0033100147265940905, + 0.004470316227525473, + 0.00478186784312129, + 0.0040976800955832005, + 0.005695066414773464, + 0.009768554009497166, + 0.004636948462575674, + 0.008844450116157532, + 0.004373840056359768, + 0.009057942777872086, + 0.03856151923537254, + 0.0044610039331018925, + 0.023748883977532387, + 0.005486902315169573, + 0.00413683382794261, + 0.003934692591428757, + 0.002385543892160058, + 0.003123318077996373, + 0.0045080543495714664, + 0.0021604837384074926, + 0.004140674602240324, + 0.015197969041764736, + 0.01460576243698597, + 0.02310626208782196, + 0.013636826537549496, + 0.010101119987666607, + 0.02320070005953312, + 0.01110825128853321, + 0.020453311502933502, + 0.007483916822820902, + 0.00505403708666563, + 0.006654433440417051, + 0.0037031627725809813, + 0.005480940453708172, + 0.021028798073530197, + 0.0047976733185350895, + 0.014883510768413544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04863535985350609 + } + ], + "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/0000160/group_13/index_tensor([3882, 3882])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_tensor([3882, 3882])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9613810d8362f819e7e0e2f64781dd8f54746ad2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_tensor([3882, 3882])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic aerial tracking shot begins from an upward perspective and smoothly transitions parallel to the subject. Atop a steep mountaintop shrouded in thick, swirling clouds, a massive eagle stands at the center of the frame. Its plumage is a deep chocolate brown, each feather edged with a metallic sheen. A halo of golden-brown down curls trembles gently in the strong wind. Its piercing amber eyes gaze into the distance, its sturdy, dark-gray beak tightly closed, a look of solemn concentration. With a powerful yet deliberate movement, its massive wings unfurl, their span stretching nearly the entire width of the frame, each feather distinctly visible in the air currents. It embraces the invisible wind, not fighting it, but with remarkable grace, skillfully harnessing its force, tilting its body slightly to glide steadily through the thin air. The camera follows the eagle, revealing an endless, ocean-like sea of clouds below. Occasionally, rays of golden sunlight penetrate the clouds, casting sacred beams of light that momentarily illuminate the intricate patterns on its wings. The entire scene is dominated by a cool color scheme, dominated by the gray-black of the rocks and the pure white of the clouds. The warm brown tones of the giant eagle become the visual focus, creating a tranquil atmosphere of loneliness, majesty, and vitality.", + "seed": 20181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20181_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": -0.4332005402073264, + "vlm_logprob": -0.4332005402073264, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.4332005402073264, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -2.0188307762145996 + ], + "opposite_logprob_tokens": [ + -0.1438307911157608 + ], + "logprob_reward": -2.0188307762145996, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.0188307762145996, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.30296650528907776, + 0.23629479110240936, + 0.06206958740949631, + 0.019702717661857605, + 0.0494866818189621, + 0.017190415412187576, + 0.03359399363398552, + 0.16670604050159454, + 0.026035958901047707, + 0.08595333248376846 + ], + "credit_map": [ + 0.021862788125872612, + 0.01703144796192646, + 0.06832574307918549, + 0.020359713584184647, + 0.01962059922516346, + 0.04665297269821167, + 0.029803255572915077, + 0.07930997014045715, + 0.022439928725361824, + 0.015607100911438465, + 0.06881683319807053, + 0.021927913650870323, + 0.011394191533327103, + 0.0171548742800951, + 0.043045900762081146, + 0.0359080471098423, + 0.016200028359889984, + 0.004280726891011, + 0.009054935537278652, + 0.006931029725819826, + 0.0039032348431646824, + 0.009666230529546738, + 0.005538535304367542, + 0.006494864821434021, + 0.0026231359224766493, + 0.0013189503224566579, + 0.002571363467723131, + 0.001736183650791645, + 0.002224719151854515, + 0.0043778130784630775, + 0.002784525044262409, + 0.002066028770059347, + 0.007743476424366236, + 0.010114165022969246, + 0.0031423892360180616, + 0.004841030575335026, + 0.008289018645882607, + 0.007186936214566231, + 0.004733060486614704, + 0.0034366052132099867, + 0.002340054139494896, + 0.0018870667554438114, + 0.0016236831434071064, + 0.0022666004952043295, + 0.002674150513485074, + 0.001874963054433465, + 0.0024357798974961042, + 0.0020881162490695715, + 0.0031607160344719887, + 0.004852065350860357, + 0.0034086061641573906, + 0.003735603764653206, + 0.004635922145098448, + 0.004919097758829594, + 0.0041825780645012856, + 0.004699403420090675, + 0.02048533596098423, + 0.023021522909402847, + 0.021223893389105797, + 0.02233586274087429, + 0.021806888282299042, + 0.014928566291928291, + 0.02176995947957039, + 0.021134013310074806, + 0.0035867972765117884, + 0.0030594582203775644, + 0.004585742950439453, + 0.0035057622008025646, + 0.0027568647637963295, + 0.0029498455114662647, + 0.0032088670413941145, + 0.0023826223332434893, + 0.01217996422201395, + 0.0073927296325564384, + 0.010948841460049152, + 0.006004151422530413, + 0.016858667135238647, + 0.01556075643748045, + 0.006329847499728203, + 0.010678376071155071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -2.0188307762145996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012571376748383045 + ], + "opposite_logprob_tokens": [ + -4.387571334838867 + ], + "logprob_reward": -0.012571376748383045, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012571376748383045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2784201204776764, + 0.17618206143379211, + 0.050457973033189774, + 0.028021905571222305, + 0.04545225575566292, + 0.03453376144170761, + 0.06391281634569168, + 0.24370458722114563, + 0.024770421907305717, + 0.05454406142234802 + ], + "credit_map": [ + 0.018094349652528763, + 0.015535867772996426, + 0.06124201789498329, + 0.020154781639575958, + 0.018418073654174805, + 0.04435645788908005, + 0.02945745922625065, + 0.07116113603115082, + 0.01903257705271244, + 0.012226688675582409, + 0.04980088025331497, + 0.017059847712516785, + 0.008073889650404453, + 0.013652423396706581, + 0.030184252187609673, + 0.026151517406105995, + 0.010794296860694885, + 0.0032886192202568054, + 0.011017659679055214, + 0.0052032251842319965, + 0.0037646230775862932, + 0.0053008003160357475, + 0.005095073953270912, + 0.005993672180920839, + 0.004933437332510948, + 0.0021446319296956062, + 0.004434213042259216, + 0.0027697631157934666, + 0.002429456217214465, + 0.00423412537202239, + 0.004373942036181688, + 0.002702336059883237, + 0.006806269288063049, + 0.007936331443488598, + 0.0034506351221352816, + 0.004802809562534094, + 0.007449957076460123, + 0.006752242799848318, + 0.00499330647289753, + 0.0032607009634375572, + 0.003966339863836765, + 0.004334781784564257, + 0.0028317554388195276, + 0.004185006488114595, + 0.005049139726907015, + 0.004825191106647253, + 0.005119301378726959, + 0.004222244955599308, + 0.005900903604924679, + 0.008549042046070099, + 0.0086018992587924, + 0.009632773697376251, + 0.008216452784836292, + 0.007955971173942089, + 0.006652913987636566, + 0.008402863517403603, + 0.030355535447597504, + 0.03415939211845398, + 0.030164150521159172, + 0.03137390315532684, + 0.033476587384939194, + 0.020465224981307983, + 0.031462542712688446, + 0.032247260212898254, + 0.003463233355432749, + 0.002729465253651142, + 0.0036749320570379496, + 0.002703475533053279, + 0.002940228208899498, + 0.0040695443749427795, + 0.003126509254798293, + 0.002063034102320671, + 0.007510364521294832, + 0.004668070934712887, + 0.007001788355410099, + 0.003803109982982278, + 0.010323203168809414, + 0.00964188203215599, + 0.004263536538928747, + 0.0073321047239005566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012571376748383045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012558897957205772 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.012558897957205772, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012558897957205772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23584996163845062, + 0.13227176666259766, + 0.10365290939807892, + 0.020856089890003204, + 0.08110114187002182, + 0.029200106859207153, + 0.040091708302497864, + 0.22747477889060974, + 0.06070249155163765, + 0.06879900395870209 + ], + "credit_map": [ + 0.01772378571331501, + 0.013088538311421871, + 0.05025710538029671, + 0.015731023624539375, + 0.016648661345243454, + 0.039279062300920486, + 0.023734545335173607, + 0.05938724800944328, + 0.014823568984866142, + 0.009821864776313305, + 0.03759564831852913, + 0.012016071937978268, + 0.0056001232005655766, + 0.010191569104790688, + 0.022917119786143303, + 0.019305799156427383, + 0.03167194873094559, + 0.0067626843228936195, + 0.014177700504660606, + 0.008811610750854015, + 0.005347236525267363, + 0.01640911214053631, + 0.008238425478339195, + 0.01223419327288866, + 0.0029952996410429478, + 0.0019881075713783503, + 0.002512618899345398, + 0.00272095063701272, + 0.002289785072207451, + 0.003949832636862993, + 0.0024673771113157272, + 0.0019321197178214788, + 0.012487577274441719, + 0.014916626736521721, + 0.005289074499160051, + 0.0076020946726202965, + 0.016368359327316284, + 0.01122214924544096, + 0.007606783416122198, + 0.005608480889350176, + 0.003270629094913602, + 0.004033958539366722, + 0.0027288927230983973, + 0.003588748862966895, + 0.0039539635181427, + 0.003964147064834833, + 0.004153155721724033, + 0.0035066106356680393, + 0.004168655723333359, + 0.006083190906792879, + 0.004711498040705919, + 0.005051093176007271, + 0.005139774177223444, + 0.004960194230079651, + 0.004573172423988581, + 0.005404126364737749, + 0.0269858930259943, + 0.032447997480630875, + 0.030706871300935745, + 0.029849441722035408, + 0.030581722036004066, + 0.01915721967816353, + 0.028406323865056038, + 0.029339302331209183, + 0.008574089035391808, + 0.008354543708264828, + 0.009190289303660393, + 0.009040961042046547, + 0.0059087201952934265, + 0.007039004936814308, + 0.0077995238825678825, + 0.00479535898193717, + 0.009498519822955132, + 0.006293852347880602, + 0.009295349009335041, + 0.005422086454927921, + 0.012718789279460907, + 0.01220744103193283, + 0.005380704998970032, + 0.007982255890965462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.012558897957205772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789739117026329 + ], + "opposite_logprob_tokens": [ + -2.5789740085601807 + ], + "logprob_reward": -0.0789739117026329, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789739117026329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1478520631790161, + 0.07419797778129578, + 0.06691179424524307, + 0.06963058561086655, + 0.055298324674367905, + 0.021218201145529747, + 0.056246317923069, + 0.4375181496143341, + 0.030073143541812897, + 0.041053406894207 + ], + "credit_map": [ + 0.015158350579440594, + 0.011709385551512241, + 0.03269446641206741, + 0.01088310219347477, + 0.009212660603225231, + 0.017959104850888252, + 0.013032671995460987, + 0.03720232844352722, + 0.008869192562997341, + 0.007671328727155924, + 0.017381152138113976, + 0.006903560366481543, + 0.004969579633325338, + 0.00830748863518238, + 0.009656532667577267, + 0.010439151898026466, + 0.009472183883190155, + 0.005502181127667427, + 0.01120575238019228, + 0.009161695837974548, + 0.005854676011949778, + 0.008314637467265129, + 0.008630852214992046, + 0.008769821375608444, + 0.014516454190015793, + 0.005019132513552904, + 0.00880973506718874, + 0.006824126932770014, + 0.007794518023729324, + 0.014392495155334473, + 0.006460282951593399, + 0.005813846830278635, + 0.011723160743713379, + 0.0063118780963122845, + 0.004528425168246031, + 0.004767531529068947, + 0.005323915276676416, + 0.011211996898055077, + 0.005838447716087103, + 0.005592971108853817, + 0.00247015617787838, + 0.002479230985045433, + 0.0017020748928189278, + 0.0030953658279031515, + 0.0024019600823521614, + 0.0035249346401542425, + 0.0028251034673303366, + 0.0027193755377084017, + 0.006203570403158665, + 0.008649016730487347, + 0.006271730177104473, + 0.007082951255142689, + 0.006773096974939108, + 0.007345312740653753, + 0.006505161989480257, + 0.0074154832400381565, + 0.052252162247896194, + 0.06256509572267532, + 0.06400939077138901, + 0.055460814386606216, + 0.05703660473227501, + 0.039070650935173035, + 0.0524512454867363, + 0.05467221513390541, + 0.003933348227292299, + 0.004039993043988943, + 0.004933896008878946, + 0.004627111367881298, + 0.003918024245649576, + 0.002763425698503852, + 0.0032745248172432184, + 0.0025828215293586254, + 0.005916047841310501, + 0.0034915118012577295, + 0.005800395738333464, + 0.0027526835910975933, + 0.007908971980214119, + 0.007248974870890379, + 0.002813909901306033, + 0.005120918154716492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0789739117026329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306773841381073 + ], + "opposite_logprob_tokens": [ + -3.168067693710327 + ], + "logprob_reward": -0.04306773841381073, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306773841381073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21846584975719452, + 0.1376579850912094, + 0.2649409770965576, + 0.02536923997104168, + 0.03592955321073532, + 0.013859456405043602, + 0.046435464173555374, + 0.09279810637235641, + 0.043738044798374176, + 0.1208052709698677 + ], + "credit_map": [ + 0.01708614081144333, + 0.012725531123578548, + 0.04271917790174484, + 0.014119111001491547, + 0.017227325588464737, + 0.04158832132816315, + 0.023954883217811584, + 0.0490453876554966, + 0.015609387308359146, + 0.009726295247673988, + 0.038018859922885895, + 0.012664402835071087, + 0.006163071841001511, + 0.011261719278991222, + 0.02335868775844574, + 0.020855555310845375, + 0.08333244919776917, + 0.01750558614730835, + 0.03520244359970093, + 0.016976337879896164, + 0.012708794325590134, + 0.0430726483464241, + 0.02174447663128376, + 0.034398261457681656, + 0.004282978363335133, + 0.0013063353253528476, + 0.00232239393517375, + 0.0015283024404197931, + 0.0018305348930880427, + 0.0029965152498334646, + 0.007155990228056908, + 0.003946192562580109, + 0.005969200283288956, + 0.0085838558152318, + 0.0020568836480379105, + 0.0038069142028689384, + 0.004805630072951317, + 0.0037943413481116295, + 0.0042924401350319386, + 0.0026202930603176355, + 0.0018699115607887506, + 0.0014985240995883942, + 0.001351533574052155, + 0.0018954266561195254, + 0.0024886750616133213, + 0.0016599632799625397, + 0.0016056874301284552, + 0.0014897348592057824, + 0.004423306789249182, + 0.006375257391482592, + 0.0059082480147480965, + 0.006562019698321819, + 0.005410803947597742, + 0.006038855295628309, + 0.005336889065802097, + 0.0063800872303545475, + 0.011192313395440578, + 0.012784801423549652, + 0.010723927989602089, + 0.011833546683192253, + 0.01477828063070774, + 0.008043777197599411, + 0.01121414452791214, + 0.012227321974933147, + 0.005469403229653835, + 0.005042998120188713, + 0.005874343682080507, + 0.00601769657805562, + 0.007172321900725365, + 0.0046815709210932255, + 0.005707903765141964, + 0.0037718117237091064, + 0.016884543001651764, + 0.01085522212088108, + 0.015249774791300297, + 0.009272669441998005, + 0.02282673679292202, + 0.022160246968269348, + 0.008608951233327389, + 0.014947137795388699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04306773841381073 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic aerial tracking shot begins from an upward perspective and smoothly transitions parallel to the subject. Atop a steep mountaintop shrouded in thick, swirling clouds, a massive eagle stands at the center of the frame. Its plumage is a deep chocolate brown, each feather edged with a metallic sheen. A halo of golden-brown down curls trembles gently in the strong wind. Its piercing amber eyes gaze into the distance, its sturdy, dark-gray beak tightly closed, a look of solemn concentration. With a powerful yet deliberate movement, its massive wings unfurl, their span stretching nearly the entire width of the frame, each feather distinctly visible in the air currents. It embraces the invisible wind, not fighting it, but with remarkable grace, skillfully harnessing its force, tilting its body slightly to glide steadily through the thin air. The camera follows the eagle, revealing an endless, ocean-like sea of clouds below. Occasionally, rays of golden sunlight penetrate the clouds, casting sacred beams of light that momentarily illuminate the intricate patterns on its wings. The entire scene is dominated by a cool color scheme, dominated by the gray-black of the rocks and the pure white of the clouds. The warm brown tones of the giant eagle become the visual focus, creating a tranquil atmosphere of loneliness, majesty, and vitality.", + "seed": 20182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20182_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": -0.021586757339537143, + "vlm_logprob": -0.021586757339537143, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.021586757339537143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05223081633448601 + ], + "opposite_logprob_tokens": [ + -3.0522308349609375 + ], + "logprob_reward": -0.05223081633448601, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05223081633448601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12988726794719696, + 0.3268128037452698, + 0.05295766144990921, + 0.09625083953142166, + 0.03940402716398239, + 0.07832121104001999, + 0.030113399028778076, + 0.06585322320461273, + 0.1281825751066208, + 0.05221698805689812 + ], + "credit_map": [ + 0.01481044851243496, + 0.009393996559083462, + 0.012445568107068539, + 0.01872873120009899, + 0.022570835426449776, + 0.01201592106372118, + 0.021178802475333214, + 0.018742971122264862, + 0.03593937307596207, + 0.03424246981739998, + 0.035497527569532394, + 0.034603051841259, + 0.02961435168981552, + 0.05147142335772514, + 0.05632217600941658, + 0.04912241920828819, + 0.011218325234949589, + 0.006694275885820389, + 0.006053077522665262, + 0.0046621100045740604, + 0.005182058084756136, + 0.004184527322649956, + 0.006746282801032066, + 0.008217005990445614, + 0.010985856875777245, + 0.011451003141701221, + 0.009365830570459366, + 0.00906537938863039, + 0.009679875336587429, + 0.012163394130766392, + 0.015737777575850487, + 0.017801735550165176, + 0.005530988331884146, + 0.0049884989857673645, + 0.0032820268534123898, + 0.005142548121511936, + 0.006152468733489513, + 0.004293857607990503, + 0.005433352198451757, + 0.0045802900567650795, + 0.011504230089485645, + 0.012385644018650055, + 0.014486712403595448, + 0.008568650111556053, + 0.007097357418388128, + 0.0054057929664850235, + 0.008725233376026154, + 0.010147592052817345, + 0.002613997785374522, + 0.002863258821889758, + 0.0033331294544041157, + 0.004199430346488953, + 0.0059626055881381035, + 0.0036997278220951557, + 0.004372293129563332, + 0.0030689584091305733, + 0.005767386872321367, + 0.00455824751406908, + 0.00970142986625433, + 0.014396808110177517, + 0.010283258743584156, + 0.0037318780086934566, + 0.006589676719158888, + 0.010824539698660374, + 0.023623069748282433, + 0.030471140518784523, + 0.01879161037504673, + 0.010863233357667923, + 0.009673875756561756, + 0.01457730121910572, + 0.009494263678789139, + 0.010688090696930885, + 0.01076860073953867, + 0.006086304783821106, + 0.012622689828276634, + 0.004220244940370321, + 0.004563507158309221, + 0.0040590157732367516, + 0.005219846498221159, + 0.004676785785704851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.05223081633448601 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979415699839592 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.00979415699839592, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979415699839592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05853298678994179, + 0.3798048794269562, + 0.033952996134757996, + 0.06252048909664154, + 0.10513174533843994, + 0.07247751206159592, + 0.15340162813663483, + 0.04186603054404259, + 0.03547957167029381, + 0.05683222785592079 + ], + "credit_map": [ + 0.0077159469947218895, + 0.003691837191581726, + 0.004503785166889429, + 0.004994397517293692, + 0.00859130173921585, + 0.009872582741081715, + 0.010061031207442284, + 0.009102107025682926, + 0.042903538793325424, + 0.03542957827448845, + 0.040849220007658005, + 0.0412927120923996, + 0.0371566042304039, + 0.05870358273386955, + 0.06426432728767395, + 0.0592053085565567, + 0.005593603476881981, + 0.0031802828889340162, + 0.0040003517642617226, + 0.0033053297083824873, + 0.004027112387120724, + 0.002942187013104558, + 0.004837960470467806, + 0.006066168658435345, + 0.013016846030950546, + 0.005395183339715004, + 0.007454765494912863, + 0.006397255230695009, + 0.006354178301990032, + 0.007869298569858074, + 0.006598686799407005, + 0.00943427998572588, + 0.01346630696207285, + 0.010417260229587555, + 0.010232318192720413, + 0.015065865591168404, + 0.016353804618120193, + 0.01267064455896616, + 0.014937249943614006, + 0.011988294310867786, + 0.012719281017780304, + 0.00958744715899229, + 0.008330781012773514, + 0.01045023649930954, + 0.010388990864157677, + 0.0048998259007930756, + 0.007781557738780975, + 0.00831939559429884, + 0.018715761601924896, + 0.017726389691233635, + 0.020618194714188576, + 0.018423626199364662, + 0.025388142094016075, + 0.016449185088276863, + 0.019580159336328506, + 0.016500167548656464, + 0.005538196302950382, + 0.006719617638736963, + 0.005256851203739643, + 0.005070780403912067, + 0.005261481273919344, + 0.004762075375765562, + 0.003981133457273245, + 0.005275895353406668, + 0.005897060036659241, + 0.00698917405679822, + 0.004738707095384598, + 0.004808349534869194, + 0.00342621305026114, + 0.00393622275441885, + 0.0029081148095428944, + 0.002775729401037097, + 0.010820643976330757, + 0.006589943543076515, + 0.015088917687535286, + 0.0042917365208268166, + 0.004647520370781422, + 0.003938151523470879, + 0.005661866161972284, + 0.005793446209281683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 4 + ], + "selected_reward": -0.00979415699839592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792031720280647 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.009792031720280647, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792031720280647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0472259521484375, + 0.4398490786552429, + 0.03424963355064392, + 0.10519266128540039, + 0.08382091671228409, + 0.16523446142673492, + 0.022759223356842995, + 0.025875207036733627, + 0.035111457109451294, + 0.040681447833776474 + ], + "credit_map": [ + 0.006759985815733671, + 0.0032881044317036867, + 0.003436862723901868, + 0.005159649532288313, + 0.007057118695229292, + 0.007024924736469984, + 0.008465164341032505, + 0.006034138612449169, + 0.055783309042453766, + 0.04406706616282463, + 0.044209789484739304, + 0.049128927290439606, + 0.04806840419769287, + 0.06575063616037369, + 0.06559312343597412, + 0.06724779307842255, + 0.005957590416073799, + 0.0037692475598305464, + 0.0036921370774507523, + 0.0032971363980323076, + 0.0025260921102017164, + 0.0034081079065799713, + 0.00472825113683939, + 0.006871069315820932, + 0.020031359046697617, + 0.010406563989818096, + 0.014255763031542301, + 0.009838677942752838, + 0.008768167346715927, + 0.012603243812918663, + 0.01126048807054758, + 0.018028389662504196, + 0.010171874426305294, + 0.007979067042469978, + 0.008999938145279884, + 0.014152639545500278, + 0.013620189391076565, + 0.00894157588481903, + 0.010503632016479969, + 0.00945199839770794, + 0.018934523686766624, + 0.022616447880864143, + 0.015879588201642036, + 0.023670276626944542, + 0.019673366099596024, + 0.020306779071688652, + 0.02270367555320263, + 0.021449802443385124, + 0.002856185659766197, + 0.002543763956055045, + 0.00253855693154037, + 0.003727992996573448, + 0.0028966101817786694, + 0.002945834072306752, + 0.0029596039094030857, + 0.0022906723897904158, + 0.003686180803924799, + 0.003971248399466276, + 0.0039044504519551992, + 0.002712339162826538, + 0.0033240728080272675, + 0.0022932307329028845, + 0.002808117074891925, + 0.003175564343109727, + 0.004777128342539072, + 0.006659387145191431, + 0.004779393319040537, + 0.00545229809358716, + 0.003939968999475241, + 0.003043888369575143, + 0.0032945165876299143, + 0.0031648732256144285, + 0.007341970223933458, + 0.0058802892453968525, + 0.007251569535583258, + 0.004135739058256149, + 0.004250398371368647, + 0.003674185834825039, + 0.004483241122215986, + 0.003664052579551935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.009792031720280647 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0263367909938097 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.0263367909938097, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0263367909938097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09038620442152023, + 0.1992592066526413, + 0.0722634494304657, + 0.07396295666694641, + 0.13308940827846527, + 0.1147436797618866, + 0.06199939548969269, + 0.09780124574899673, + 0.07011374831199646, + 0.08638070523738861 + ], + "credit_map": [ + 0.008747987449169159, + 0.006108206696808338, + 0.006908406503498554, + 0.01165427453815937, + 0.019871985539793968, + 0.011179760098457336, + 0.012976258061826229, + 0.012939317151904106, + 0.02653312496840954, + 0.021459298208355904, + 0.02086329087615013, + 0.020426232367753983, + 0.01886110007762909, + 0.03204319626092911, + 0.030435429885983467, + 0.02863752655684948, + 0.011583815328776836, + 0.00664651719853282, + 0.009818880818784237, + 0.008031017147004604, + 0.007328782230615616, + 0.007641003932803869, + 0.009852561168372631, + 0.01136085856705904, + 0.016538353636860847, + 0.008901047520339489, + 0.007177648600190878, + 0.005260305013507605, + 0.006620343774557114, + 0.009261904284358025, + 0.008395581506192684, + 0.011807762086391449, + 0.015971621498465538, + 0.011969194747507572, + 0.016908518970012665, + 0.022193731740117073, + 0.019243720918893814, + 0.01459655724465847, + 0.017374511808156967, + 0.014831543900072575, + 0.012351338751614094, + 0.012466034851968288, + 0.021622974425554276, + 0.015563461929559708, + 0.014117114245891571, + 0.01280065905302763, + 0.01118155475705862, + 0.014640526846051216, + 0.00587306497618556, + 0.006877455394715071, + 0.007702173665165901, + 0.012089964933693409, + 0.0059488629922270775, + 0.0057083917781710625, + 0.011708688922226429, + 0.006090785842388868, + 0.012318256311118603, + 0.01078207790851593, + 0.0172665286809206, + 0.010742600075900555, + 0.013873785734176636, + 0.007938612252473831, + 0.008668395690619946, + 0.016210980713367462, + 0.011479067616164684, + 0.01670217514038086, + 0.009314842522144318, + 0.007911013439297676, + 0.007334833964705467, + 0.0072812288999557495, + 0.005405262112617493, + 0.00468531996011734, + 0.01692373864352703, + 0.009293156675994396, + 0.02407393418252468, + 0.006813894025981426, + 0.005818977952003479, + 0.007138099987059832, + 0.008778102695941925, + 0.007540796883404255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.0263367909938097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09059139341115952, + 0.24706339836120605, + 0.1805720180273056, + 0.033272095024585724, + 0.25302258133888245, + 0.030641425400972366, + 0.049566321074962616, + 0.03750661015510559, + 0.05703470855951309, + 0.02072948031127453 + ], + "credit_map": [ + 0.010180639103055, + 0.006669558584690094, + 0.00548713281750679, + 0.008303910493850708, + 0.017851226031780243, + 0.01042858138680458, + 0.017227882519364357, + 0.014442461542785168, + 0.03135749697685242, + 0.026709752157330513, + 0.024919062852859497, + 0.026378341019153595, + 0.025390591472387314, + 0.03694389760494232, + 0.03856922313570976, + 0.03679502382874489, + 0.023439854383468628, + 0.03586619719862938, + 0.018828559666872025, + 0.016447626054286957, + 0.017666669562458992, + 0.01731199212372303, + 0.025433165952563286, + 0.02557794749736786, + 0.004272709600627422, + 0.003865226171910763, + 0.003313050838187337, + 0.0032274636905640364, + 0.002725559286773205, + 0.004328861366957426, + 0.005296642426401377, + 0.006242579780519009, + 0.028501396998763084, + 0.02328309789299965, + 0.028232386335730553, + 0.04439971596002579, + 0.03765898197889328, + 0.026819657534360886, + 0.034387048333883286, + 0.029740292578935623, + 0.004214834421873093, + 0.004614246543496847, + 0.004596625920385122, + 0.0025976772885769606, + 0.003708459436893463, + 0.003543123137205839, + 0.0036749294959008694, + 0.0036915293894708157, + 0.0051942262798547745, + 0.004403975326567888, + 0.006072909105569124, + 0.006685146130621433, + 0.005675289314240217, + 0.005210502538830042, + 0.010601513087749481, + 0.005722758825868368, + 0.004833547864109278, + 0.004951790906488895, + 0.005573746748268604, + 0.0037379954010248184, + 0.004341901745647192, + 0.0033762119710445404, + 0.003727323841303587, + 0.006964089348912239, + 0.00927070900797844, + 0.013025278225541115, + 0.009451357647776604, + 0.005983582232147455, + 0.004464594181627035, + 0.005088986363261938, + 0.005524442531168461, + 0.0042257546447217464, + 0.003518963698297739, + 0.003262197133153677, + 0.004135425202548504, + 0.0016876503359526396, + 0.002202557632699609, + 0.0016841135220602155, + 0.002007370349019766, + 0.002231201622635126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.009779990650713444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic aerial tracking shot begins from an upward perspective and smoothly transitions parallel to the subject. Atop a steep mountaintop shrouded in thick, swirling clouds, a massive eagle stands at the center of the frame. Its plumage is a deep chocolate brown, each feather edged with a metallic sheen. A halo of golden-brown down curls trembles gently in the strong wind. Its piercing amber eyes gaze into the distance, its sturdy, dark-gray beak tightly closed, a look of solemn concentration. With a powerful yet deliberate movement, its massive wings unfurl, their span stretching nearly the entire width of the frame, each feather distinctly visible in the air currents. It embraces the invisible wind, not fighting it, but with remarkable grace, skillfully harnessing its force, tilting its body slightly to glide steadily through the thin air. The camera follows the eagle, revealing an endless, ocean-like sea of clouds below. Occasionally, rays of golden sunlight penetrate the clouds, casting sacred beams of light that momentarily illuminate the intricate patterns on its wings. The entire scene is dominated by a cool color scheme, dominated by the gray-black of the rocks and the pure white of the clouds. The warm brown tones of the giant eagle become the visual focus, creating a tranquil atmosphere of loneliness, majesty, and vitality.", + "seed": 20183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20183_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": -0.05434612147510052, + "vlm_logprob": -0.05434612147510052, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.05434612147510052, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22892941534519196 + ], + "opposite_logprob_tokens": [ + -1.6039294004440308 + ], + "logprob_reward": -0.22892941534519196, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22892941534519196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04498721659183502, + 0.034406695514917374, + 0.12618961930274963, + 0.20151060819625854, + 0.16011404991149902, + 0.06512699276208878, + 0.0461210235953331, + 0.0641215592622757, + 0.1001986563205719, + 0.15722361207008362 + ], + "credit_map": [ + 0.003633868647739291, + 0.003938252106308937, + 0.003885244484990835, + 0.007519528269767761, + 0.004112806636840105, + 0.004234865307807922, + 0.010192935355007648, + 0.007469713222235441, + 0.004310992080718279, + 0.004528759978711605, + 0.0030722301453351974, + 0.003313899738714099, + 0.00321906222961843, + 0.005338242277503014, + 0.006147697567939758, + 0.004475809168070555, + 0.013989773578941822, + 0.01430604513734579, + 0.014414596371352673, + 0.02251107431948185, + 0.01568688452243805, + 0.013288980349898338, + 0.014620316214859486, + 0.017371943220496178, + 0.027725955471396446, + 0.018558181822299957, + 0.013350092805922031, + 0.022623226046562195, + 0.019686922430992126, + 0.04690086096525192, + 0.026291003450751305, + 0.02637437917292118, + 0.018484842032194138, + 0.010694636963307858, + 0.0379803329706192, + 0.009781640022993088, + 0.012754127383232117, + 0.01643163338303566, + 0.008178962394595146, + 0.045807868242263794, + 0.006006339099258184, + 0.010259541682898998, + 0.006076892372220755, + 0.006445521023124456, + 0.014374174177646637, + 0.006584846414625645, + 0.005846179556101561, + 0.009533499367535114, + 0.006373105105012655, + 0.004218585789203644, + 0.004754289053380489, + 0.004022085573524237, + 0.003177632112056017, + 0.006987735163420439, + 0.007085569202899933, + 0.009502019733190536, + 0.009869550354778767, + 0.005548315122723579, + 0.00873491633683443, + 0.00899017695337534, + 0.004061222542077303, + 0.005674270447343588, + 0.005080618429929018, + 0.016162486746907234, + 0.007371784653514624, + 0.008128059096634388, + 0.019518297165632248, + 0.016807811334729195, + 0.0069381315261125565, + 0.008287572301924229, + 0.0190491434186697, + 0.014097854495048523, + 0.006932786665856838, + 0.010539071634411812, + 0.010580930858850479, + 0.020505379885435104, + 0.010503433644771576, + 0.013757506385445595, + 0.048644889146089554, + 0.03575962036848068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.22892941534519196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002498839981853962 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.002498839981853962, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002498839981853962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1496303230524063, + 0.039739251136779785, + 0.1602325588464737, + 0.17719757556915283, + 0.22159917652606964, + 0.09271710366010666, + 0.04372662678360939, + 0.02386845275759697, + 0.04805337265133858, + 0.04323554039001465 + ], + "credit_map": [ + 0.01484998781234026, + 0.015655767172574997, + 0.011351815424859524, + 0.02369450405240059, + 0.018985029309988022, + 0.018365168944001198, + 0.022823989391326904, + 0.023904064670205116, + 0.004263292066752911, + 0.006418648641556501, + 0.003661821596324444, + 0.005395500920712948, + 0.0041082133539021015, + 0.006272927857935429, + 0.004805436823517084, + 0.004813413601368666, + 0.019354969263076782, + 0.01630883291363716, + 0.018930377438664436, + 0.029247837141156197, + 0.01577698066830635, + 0.01687534898519516, + 0.01762363687157631, + 0.026114575564861298, + 0.029492860659956932, + 0.013094686903059483, + 0.014808272011578083, + 0.02125495672225952, + 0.016322748735547066, + 0.04149429872632027, + 0.01575993001461029, + 0.02496982552111149, + 0.028496574610471725, + 0.018708838149905205, + 0.04847073182463646, + 0.015407436527311802, + 0.019675174728035927, + 0.02213340438902378, + 0.011499563232064247, + 0.05720743536949158, + 0.01017497293651104, + 0.010360614396631718, + 0.009349176660180092, + 0.01577320136129856, + 0.012298718094825745, + 0.01039868127554655, + 0.010255460627377033, + 0.014106286689639091, + 0.006294930353760719, + 0.0033012155909091234, + 0.004140172619372606, + 0.003176947357133031, + 0.0037776180543005466, + 0.006153539288789034, + 0.007270818576216698, + 0.009611386805772781, + 0.0032599156256765127, + 0.0021938285790383816, + 0.002946278313174844, + 0.0035427066031843424, + 0.002114681527018547, + 0.0030758678913116455, + 0.0023746283259242773, + 0.004360548220574856, + 0.003558010095730424, + 0.0034288482274860144, + 0.008105827495455742, + 0.007818981073796749, + 0.0037638491485267878, + 0.004913726355880499, + 0.007780123967677355, + 0.00868400651961565, + 0.0028117820620536804, + 0.004061687737703323, + 0.003853520844131708, + 0.0049631050787866116, + 0.0036444233264774084, + 0.0036777746863663197, + 0.011121757328510284, + 0.009101491421461105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.002498839981853962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953080952167511 + ], + "opposite_logprob_tokens": [ + -5.130953311920166 + ], + "logprob_reward": -0.005953080952167511, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953080952167511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17255456745624542, + 0.06028837338089943, + 0.06568795442581177, + 0.13407959043979645, + 0.21933698654174805, + 0.08906348049640656, + 0.040344152599573135, + 0.03550593927502632, + 0.11844543367624283, + 0.06469352543354034 + ], + "credit_map": [ + 0.024067554622888565, + 0.016764892265200615, + 0.029229527339339256, + 0.022320980206131935, + 0.020539266988635063, + 0.01637154258787632, + 0.018600430339574814, + 0.024660378694534302, + 0.006400125101208687, + 0.005966892000287771, + 0.005277675110846758, + 0.006808757316321135, + 0.005454984027892351, + 0.011918190866708755, + 0.010328684002161026, + 0.008133066818118095, + 0.009794490411877632, + 0.006499157287180424, + 0.007455719169229269, + 0.010128618218004704, + 0.006885921116918325, + 0.0070625110529363155, + 0.008804854936897755, + 0.00905668456107378, + 0.02056560479104519, + 0.01665990985929966, + 0.008814709261059761, + 0.016484485939145088, + 0.016501648351550102, + 0.026882631704211235, + 0.012145484797656536, + 0.016025124117732048, + 0.019941167905926704, + 0.011880825273692608, + 0.057303592562675476, + 0.01282348670065403, + 0.014947780407965183, + 0.022821806371212006, + 0.010193178430199623, + 0.06942515820264816, + 0.009169746190309525, + 0.00833040475845337, + 0.009898554533720016, + 0.016476042568683624, + 0.010273916646838188, + 0.009689069353044033, + 0.011093181557953358, + 0.014132567681372166, + 0.007259674835950136, + 0.003427444724366069, + 0.0033781423699110746, + 0.002382818143814802, + 0.003781235544010997, + 0.006046146620064974, + 0.005544126965105534, + 0.008524563163518906, + 0.00494186207652092, + 0.002702987752854824, + 0.005128171294927597, + 0.00516748009249568, + 0.0028383887838572264, + 0.003398578381165862, + 0.004579664673656225, + 0.0067488057538867, + 0.00964827835559845, + 0.008772389031946659, + 0.017649130895733833, + 0.020343290641903877, + 0.008382563479244709, + 0.01665579341351986, + 0.01786593347787857, + 0.019128063693642616, + 0.002787014003843069, + 0.005301560740917921, + 0.0036475316155701876, + 0.006168946623802185, + 0.003719308180734515, + 0.006758952047675848, + 0.018308134749531746, + 0.01800207793712616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005953080952167511 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327735908329487 + ], + "opposite_logprob_tokens": [ + -3.7732772827148438 + ], + "logprob_reward": -0.02327735908329487, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327735908329487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08468980342149734, + 0.07080791890621185, + 0.1073712557554245, + 0.1362946480512619, + 0.09565618634223938, + 0.2781662344932556, + 0.03859628364443779, + 0.055890463292598724, + 0.10193073749542236, + 0.03059641644358635 + ], + "credit_map": [ + 0.011813728138804436, + 0.007659768220037222, + 0.011469795368611813, + 0.012800772674381733, + 0.008042306639254093, + 0.007747126743197441, + 0.011857433244585991, + 0.013298881240189075, + 0.0069926041178405285, + 0.006574464961886406, + 0.005575189832597971, + 0.0070366645231842995, + 0.007601233199238777, + 0.012794814072549343, + 0.013659995049238205, + 0.010572957806289196, + 0.010406655259430408, + 0.011995693668723106, + 0.012242263182997704, + 0.01901140809059143, + 0.012509978376328945, + 0.010112361051142216, + 0.013856998644769192, + 0.017235897481441498, + 0.01960562914609909, + 0.017847491428256035, + 0.009786759503185749, + 0.01567370630800724, + 0.014254552312195301, + 0.02849450893700123, + 0.01429393794387579, + 0.016338061541318893, + 0.015663878992199898, + 0.009874442592263222, + 0.012990761548280716, + 0.009076365269720554, + 0.012687715701758862, + 0.011922367848455906, + 0.0071658967062830925, + 0.01627475582063198, + 0.021484877914190292, + 0.03875366598367691, + 0.02840118482708931, + 0.03491036966443062, + 0.06848937273025513, + 0.027654532343149185, + 0.023678062483668327, + 0.034794170409440994, + 0.005234908312559128, + 0.006423243321478367, + 0.0055843559093773365, + 0.0031033663544803858, + 0.002872633282095194, + 0.005546745378524065, + 0.003999408334493637, + 0.005831623449921608, + 0.006894794758409262, + 0.005070130340754986, + 0.00736768264323473, + 0.00755098694935441, + 0.00493974145501852, + 0.007137948181480169, + 0.00681953364983201, + 0.010109642520546913, + 0.008480384014546871, + 0.007280379068106413, + 0.01766427978873253, + 0.014567656442523003, + 0.005397695582360029, + 0.010835336521267891, + 0.019013751298189163, + 0.01869126781821251, + 0.0028605558909475803, + 0.0030694715678691864, + 0.0028570261783897877, + 0.00468455208465457, + 0.002628224901854992, + 0.003270443296059966, + 0.00555795943364501, + 0.005668183788657188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.02327735908329487 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.2516513466835022, + 0.053598858416080475, + 0.18710578978061676, + 0.058619461953639984, + 0.13341307640075684, + 0.09693978726863861, + 0.060534991323947906, + 0.06342439353466034, + 0.03355180844664574, + 0.06116050109267235 + ], + "credit_map": [ + 0.03318282589316368, + 0.024282844737172127, + 0.04238440841436386, + 0.033219143748283386, + 0.02924986183643341, + 0.021466530859470367, + 0.03190712630748749, + 0.035958606749773026, + 0.006378037855029106, + 0.009138235822319984, + 0.004235466942191124, + 0.005419821012765169, + 0.006479951553046703, + 0.009354955516755581, + 0.00603816332295537, + 0.006554222200065851, + 0.016992980614304543, + 0.021998316049575806, + 0.019105523824691772, + 0.03244341164827347, + 0.02286396361887455, + 0.019721679389476776, + 0.02427046373486519, + 0.02970943972468376, + 0.007366764359176159, + 0.0048408107832074165, + 0.005307260900735855, + 0.008076943457126617, + 0.006097897887229919, + 0.011985828168690205, + 0.006310685072094202, + 0.00863326620310545, + 0.017862753942608833, + 0.01384643092751503, + 0.023048032075166702, + 0.012156794779002666, + 0.01738881878554821, + 0.013655826449394226, + 0.007754337973892689, + 0.02770007587969303, + 0.011587422341108322, + 0.01133687887340784, + 0.010199894197285175, + 0.016370436176657677, + 0.013907068409025669, + 0.010412810370326042, + 0.010772163048386574, + 0.012353110127151012, + 0.007660843897610903, + 0.008909318596124649, + 0.008012252859771252, + 0.004915151745080948, + 0.004933358635753393, + 0.007095154840499163, + 0.008501704782247543, + 0.010507204569876194, + 0.013489961624145508, + 0.0049574291333556175, + 0.008491582237184048, + 0.008420773781836033, + 0.003043749602511525, + 0.006916430313140154, + 0.005356836598366499, + 0.012747636996209621, + 0.002284028334543109, + 0.0023383048828691244, + 0.0068296282552182674, + 0.005812035873532295, + 0.002809511963278055, + 0.0028277761302888393, + 0.006038336083292961, + 0.004612188786268234, + 0.0028276601806282997, + 0.004039260558784008, + 0.004531719256192446, + 0.007205243222415447, + 0.006072135642170906, + 0.007158976513892412, + 0.013903392478823662, + 0.015422115102410316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.01107191201299429 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic aerial tracking shot begins from an upward perspective and smoothly transitions parallel to the subject. Atop a steep mountaintop shrouded in thick, swirling clouds, a massive eagle stands at the center of the frame. Its plumage is a deep chocolate brown, each feather edged with a metallic sheen. A halo of golden-brown down curls trembles gently in the strong wind. Its piercing amber eyes gaze into the distance, its sturdy, dark-gray beak tightly closed, a look of solemn concentration. With a powerful yet deliberate movement, its massive wings unfurl, their span stretching nearly the entire width of the frame, each feather distinctly visible in the air currents. It embraces the invisible wind, not fighting it, but with remarkable grace, skillfully harnessing its force, tilting its body slightly to glide steadily through the thin air. The camera follows the eagle, revealing an endless, ocean-like sea of clouds below. Occasionally, rays of golden sunlight penetrate the clouds, casting sacred beams of light that momentarily illuminate the intricate patterns on its wings. The entire scene is dominated by a cool color scheme, dominated by the gray-black of the rocks and the pure white of the clouds. The warm brown tones of the giant eagle become the visual focus, creating a tranquil atmosphere of loneliness, majesty, and vitality.", + "seed": 20184, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_13/index_3882_seed_20184_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": -0.02651598919183016, + "vlm_logprob": -0.02651598919183016, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02651598919183016, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.091462641954422 + ], + "opposite_logprob_tokens": [ + -2.4664626121520996 + ], + "logprob_reward": -0.091462641954422, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.091462641954422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10436420887708664, + 0.1669081151485443, + 0.03422267362475395, + 0.10378406196832657, + 0.10279785841703415, + 0.14438529312610626, + 0.12350243330001831, + 0.0734608992934227, + 0.03485606238245964, + 0.11171840876340866 + ], + "credit_map": [ + 0.010354920290410519, + 0.016362499445676804, + 0.01242165919393301, + 0.01273839920759201, + 0.011676035821437836, + 0.009091711603105068, + 0.016597900539636612, + 0.01512107253074646, + 0.013606616295874119, + 0.026451321318745613, + 0.021099861711263657, + 0.014565210789442062, + 0.02439621463418007, + 0.025288719683885574, + 0.015087437815964222, + 0.026412716135382652, + 0.003744086716324091, + 0.00402468116953969, + 0.004578355234116316, + 0.00320352497510612, + 0.0056763277389109135, + 0.004624971654266119, + 0.00340221612714231, + 0.004968509078025818, + 0.013260683976113796, + 0.0234120674431324, + 0.01905665174126625, + 0.005417652428150177, + 0.019592994824051857, + 0.0076854960061609745, + 0.006420440040528774, + 0.008938077837228775, + 0.01119200699031353, + 0.02252545952796936, + 0.008630034513771534, + 0.010719336569309235, + 0.014716758392751217, + 0.013369246385991573, + 0.009971640072762966, + 0.011673364788293839, + 0.03659147024154663, + 0.02060823328793049, + 0.009384904988110065, + 0.007277247495949268, + 0.011463358998298645, + 0.012078829109668732, + 0.009509831666946411, + 0.03747141361236572, + 0.01917276717722416, + 0.014743753708899021, + 0.01497148908674717, + 0.011131067760288715, + 0.020061470568180084, + 0.014809289015829563, + 0.01184654887765646, + 0.01676604524254799, + 0.023048464208841324, + 0.006988511420786381, + 0.005699804984033108, + 0.0038751671090722084, + 0.007171690464019775, + 0.005903324577957392, + 0.0041647544130682945, + 0.016609180718660355, + 0.003392991842702031, + 0.00498068006709218, + 0.004285276401787996, + 0.0030347038991749287, + 0.005136657040566206, + 0.0052777789533138275, + 0.004995726980268955, + 0.0037522437050938606, + 0.007076862268149853, + 0.011093661189079285, + 0.007530733477324247, + 0.02846592850983143, + 0.008207892067730427, + 0.009794123470783234, + 0.010545605793595314, + 0.02900359034538269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.091462641954422 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639478426426649 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639478426426649, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639478426426649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07397531718015671, + 0.05004327371716499, + 0.020776277408003807, + 0.033034633845090866, + 0.019355155527591705, + 0.1825600117444992, + 0.02916606329381466, + 0.21607939898967743, + 0.042556747794151306, + 0.33245307207107544 + ], + "credit_map": [ + 0.005685626994818449, + 0.01548809465020895, + 0.0067641856148839, + 0.008336720988154411, + 0.013763838447630405, + 0.005727515555918217, + 0.009478666819632053, + 0.008730673231184483, + 0.004386075772345066, + 0.008724387735128403, + 0.005804239306598902, + 0.003910101018846035, + 0.009672935120761395, + 0.006732704117894173, + 0.0034559560008347034, + 0.00735687930136919, + 0.0019159556832164526, + 0.0026321210898458958, + 0.0028498347382992506, + 0.0022843668702989817, + 0.003169337986037135, + 0.0023959630634635687, + 0.002300308318808675, + 0.0032283919863402843, + 0.0034092660062015057, + 0.006128096487373114, + 0.005420001223683357, + 0.0033397290389984846, + 0.00460723415017128, + 0.002640994032844901, + 0.003355717286467552, + 0.004133597481995821, + 0.0025075487792491913, + 0.0036468375474214554, + 0.001615683431737125, + 0.0015913688112050295, + 0.003744183573871851, + 0.002156320493668318, + 0.001636989414691925, + 0.0024562240578234196, + 0.04075663536787033, + 0.025330178439617157, + 0.016126565635204315, + 0.011618143878877163, + 0.014583395794034004, + 0.01554834470152855, + 0.015832850709557533, + 0.04276391863822937, + 0.00431603379547596, + 0.0037604300305247307, + 0.003421924775466323, + 0.002939930185675621, + 0.0052915834821760654, + 0.003322775475680828, + 0.0025923834182322025, + 0.0035210056230425835, + 0.07026199251413345, + 0.02111273631453514, + 0.018882738426327705, + 0.008150472305715084, + 0.01941400021314621, + 0.01572427526116371, + 0.011120963841676712, + 0.05141223222017288, + 0.005979587789624929, + 0.006163180805742741, + 0.004985922947525978, + 0.003854237962514162, + 0.004164127632975578, + 0.005813982337713242, + 0.00695638544857502, + 0.004639324266463518, + 0.018287107348442078, + 0.0376342236995697, + 0.010834414511919022, + 0.11225789785385132, + 0.009826230816543102, + 0.026104126125574112, + 0.01684446446597576, + 0.1006646454334259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.007639478426426649 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789789095520973 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009789789095520973, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789789095520973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09303548187017441, + 0.060132596641778946, + 0.01477812696248293, + 0.01615702360868454, + 0.037532202899456024, + 0.14344492554664612, + 0.01733793318271637, + 0.0345684215426445, + 0.04441102221608162, + 0.5386022329330444 + ], + "credit_map": [ + 0.0109008913859725, + 0.013169260695576668, + 0.011095115914940834, + 0.009924205020070076, + 0.010727002285420895, + 0.011271631345152855, + 0.012130769900977612, + 0.01381660345941782, + 0.011968999169766903, + 0.018330102786421776, + 0.004984344355762005, + 0.0036512527149170637, + 0.005090225022286177, + 0.006346383597701788, + 0.0035020445939153433, + 0.006259247660636902, + 0.0016430231044068933, + 0.002495071617886424, + 0.0018088883953168988, + 0.0013843976194038987, + 0.0026041697710752487, + 0.00189640570897609, + 0.001256519346497953, + 0.0016896513989195228, + 0.0015776113141328096, + 0.002496012020856142, + 0.0022914803121238947, + 0.0014712823322042823, + 0.0025345792528241873, + 0.0015489219222217798, + 0.0019660121761262417, + 0.002271125093102455, + 0.003834894159808755, + 0.008072304539382458, + 0.0031674003694206476, + 0.003191093448549509, + 0.006169172935187817, + 0.004869673866778612, + 0.0024797506630420685, + 0.005747913382947445, + 0.0392313115298748, + 0.014110958203673363, + 0.008894252590835094, + 0.007534125354140997, + 0.010562128387391567, + 0.010758454911410809, + 0.00965159758925438, + 0.042702097445726395, + 0.0024628741666674614, + 0.0018260824726894498, + 0.0027913160156458616, + 0.0021165404468774796, + 0.0021234226878732443, + 0.002632776042446494, + 0.0015761491376906633, + 0.0018087724456563592, + 0.012202171608805656, + 0.0028864785563200712, + 0.0015063541941344738, + 0.001428580959327519, + 0.002454910660162568, + 0.0019936540629714727, + 0.0019218629458919168, + 0.010174407623708248, + 0.0056807901710271835, + 0.005418241955339909, + 0.005911144893616438, + 0.0034435030538588762, + 0.005203165113925934, + 0.004946116823703051, + 0.008747694082558155, + 0.005060367286205292, + 0.02933555096387863, + 0.060763128101825714, + 0.016949130222201347, + 0.18910974264144897, + 0.009343281388282776, + 0.04105842486023903, + 0.024323318153619766, + 0.16771966218948364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.009789789095520973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606074348092079 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.01606074348092079, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606074348092079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07432656735181808, + 0.0894530713558197, + 0.019576044753193855, + 0.05170241370797157, + 0.10331178456544876, + 0.18864800035953522, + 0.028647614642977715, + 0.0334080271422863, + 0.024864859879016876, + 0.3860616087913513 + ], + "credit_map": [ + 0.006835091393440962, + 0.009964440949261189, + 0.007743139285594225, + 0.007530517876148224, + 0.01107989251613617, + 0.009017512202262878, + 0.010732846334576607, + 0.011423122137784958, + 0.009341918863356113, + 0.0175226591527462, + 0.011906844563782215, + 0.009154853411018848, + 0.007803019136190414, + 0.0074273645877838135, + 0.011088067665696144, + 0.015208343043923378, + 0.00242925388738513, + 0.0022522304207086563, + 0.0027895227540284395, + 0.0017964409198611975, + 0.0031237886287271976, + 0.0030673216097056866, + 0.0018388122553005815, + 0.0022786722984164953, + 0.006895052269101143, + 0.009102080017328262, + 0.006829795427620411, + 0.00385678862221539, + 0.010339220985770226, + 0.005146755836904049, + 0.0038138157688081264, + 0.005718905478715897, + 0.010009702295064926, + 0.029186096042394638, + 0.008465616963803768, + 0.008873652666807175, + 0.013445793651044369, + 0.013350001536309719, + 0.006752420216798782, + 0.013228506781160831, + 0.04918774217367172, + 0.020025432109832764, + 0.012414548546075821, + 0.01086363010108471, + 0.01284539233893156, + 0.014771779999136925, + 0.014779431745409966, + 0.05376003310084343, + 0.003910161089152098, + 0.0035345987416803837, + 0.004026160575449467, + 0.0032874441239982843, + 0.005226125940680504, + 0.00334334559738636, + 0.0022269829642027617, + 0.0030927970074117184, + 0.010017849504947662, + 0.004323020577430725, + 0.0021470352075994015, + 0.0012809254694730043, + 0.004053432494401932, + 0.0018711939919739962, + 0.0019650934264063835, + 0.0077494787983596325, + 0.003404024988412857, + 0.003890739055350423, + 0.0034161540679633617, + 0.0026649197097867727, + 0.0031804670579731464, + 0.002836001105606556, + 0.003029040526598692, + 0.0024435108061879873, + 0.020692387595772743, + 0.043582022190093994, + 0.012470095418393612, + 0.13391472399234772, + 0.010175509378314018, + 0.028604835271835327, + 0.019017789512872696, + 0.11760422587394714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.01606074348092079 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 steep mountaintop shrouded in thick, swirling clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive eagle with deep chocolate brown plumage edged with a metallic sheen visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the eagle unfurl its massive wings to a span stretching nearly the entire width of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a dynamic aerial tracking shot that begins from an upward perspective and transitions parallel to the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are rays of golden sunlight visible penetrating the clouds to illuminate the eagle's wings?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101786583662033, + 0.10493553429841995, + 0.017341256141662598, + 0.042544279247522354, + 0.030462823808193207, + 0.07729825377464294, + 0.11996486783027649, + 0.05639321729540825, + 0.018113214522600174, + 0.322767972946167 + ], + "credit_map": [ + 0.02388957142829895, + 0.027816664427518845, + 0.02529730834066868, + 0.03195014223456383, + 0.021731644868850708, + 0.02261405810713768, + 0.025297023355960846, + 0.03158225119113922, + 0.02029963582754135, + 0.02199985645711422, + 0.01143401488661766, + 0.00794308539479971, + 0.012114194221794605, + 0.013487347401678562, + 0.010090037249028683, + 0.00756736658513546, + 0.0013956738403066993, + 0.0018140007741749287, + 0.0022328468039631844, + 0.002822942566126585, + 0.002837865147739649, + 0.002209315076470375, + 0.0016222544945776463, + 0.0024063587188720703, + 0.006075629033148289, + 0.0073856934905052185, + 0.006672049406915903, + 0.0027153892442584038, + 0.008604775182902813, + 0.00391667103394866, + 0.0025219928938895464, + 0.004652079660445452, + 0.003950981888920069, + 0.0046898378059268, + 0.002508302452042699, + 0.0033853661734610796, + 0.0050248862244188786, + 0.004539565183222294, + 0.00292854942381382, + 0.003435334889218211, + 0.022223899140954018, + 0.0073846434243023396, + 0.004725306760519743, + 0.0033228376414626837, + 0.004575966391712427, + 0.005789973773062229, + 0.004379458725452423, + 0.02489616349339485, + 0.017289744690060616, + 0.014850443229079247, + 0.01329604722559452, + 0.014281034469604492, + 0.02105194330215454, + 0.01626366190612316, + 0.011389225721359253, + 0.011542771942913532, + 0.019623493775725365, + 0.0056009129621088505, + 0.00430670939385891, + 0.002409952459856868, + 0.004187578801065683, + 0.0034704068675637245, + 0.0028626418206840754, + 0.013931522145867348, + 0.002232184400781989, + 0.0021418905816972256, + 0.0030028808396309614, + 0.001736604725010693, + 0.002531321719288826, + 0.0019300637068226933, + 0.002740664640441537, + 0.001797603676095605, + 0.017405489459633827, + 0.03607531636953354, + 0.011220244690775871, + 0.10990741103887558, + 0.007855071686208248, + 0.02534313127398491, + 0.01540222018957138, + 0.09955908358097076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.00762729300186038 + } + ], + "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/0000160/group_14/index_tensor([7820, 7820])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_tensor([7820, 7820])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..84d4095352c0a352ff32756fdf1788e7f225f321 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_tensor([7820, 7820])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle close-up gazes across the intricate eaves of a traditional Chinese building, where dark, rain-soaked wooden beams, their texture deepened by centuries of weather, reveal microscopic cracks. Black rainwater cautiously seeps through these ancient cracks, slowly seeping through the overlapping, dark gray clay tiles, saturated with moisture and covered in a velvety, deep-green moss. The camera then executes a subtle, slow downward tilt, following a single, viscous stream of water, tracing its path across the curved surface of the tiles. At the lowest edge of the tile, the water gathers into a heavy drop that swells and trembles for a moment, its perfectly spherical surface capturing a distorted, miniature reflection of the tranquil, damp stone courtyard below, before it separates and falls. The entire scene is bathed in the soft, diffuse light of a perpetually gray, overcast sky, creating a cool, subdued color palette of earthy browns, slate grays, and damp greens, evoking an atmosphere of profound tranquility and contemplation.", + "seed": 12130, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_7820_seed_12130_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": -0.01224759933538735, + "vlm_logprob": -0.01224759933538735, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.01224759933538735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02013622224330902 + ], + "opposite_logprob_tokens": [ + -4.27013635635376 + ], + "logprob_reward": -0.02013622224330902, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02013622224330902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039419665932655334, + 0.03571765497326851, + 0.09105384349822998, + 0.02343999221920967, + 0.015980154275894165, + 0.04179337993264198, + 0.6383723020553589, + 0.038652386516332626, + 0.05459100380539894, + 0.02097962610423565 + ], + "credit_map": [ + 0.006225552409887314, + 0.006991250440478325, + 0.005985458381474018, + 0.005083012394607067, + 0.0036873870994895697, + 0.003163037821650505, + 0.004047101829200983, + 0.004236868117004633, + 0.004121270030736923, + 0.0041455659084022045, + 0.006665822118520737, + 0.007396253291517496, + 0.0037749873008579016, + 0.00251714326441288, + 0.0036145879421383142, + 0.0034820293076336384, + 0.025449546054005623, + 0.0035127042792737484, + 0.03885454311966896, + 0.01352675911039114, + 0.001980970613658428, + 0.002690547611564398, + 0.0022570330183953047, + 0.0027817366644740105, + 0.0026038161013275385, + 0.003395930165424943, + 0.0028229975141584873, + 0.003404770977795124, + 0.0022611208260059357, + 0.002300752094015479, + 0.0024996227584779263, + 0.004150984343141317, + 0.0023822167422622442, + 0.002253815298900008, + 0.0022903704084455967, + 0.0013725337339565158, + 0.002717785770073533, + 0.002128578955307603, + 0.0015273105818778276, + 0.0013075437163934112, + 0.00621707271784544, + 0.0057853893376886845, + 0.0031915241852402687, + 0.002767507918179035, + 0.004524998366832733, + 0.004143233876675367, + 0.00893003772944212, + 0.006233619526028633, + 0.02869901992380619, + 0.04922957345843315, + 0.04206432029604912, + 0.027192963287234306, + 0.0295563917607069, + 0.21451619267463684, + 0.2138502299785614, + 0.03326360881328583, + 0.004097934812307358, + 0.005781707353889942, + 0.005228875670582056, + 0.004496544133871794, + 0.005070745944976807, + 0.006201158277690411, + 0.0033631413243710995, + 0.004412281326949596, + 0.004563503433018923, + 0.005024279467761517, + 0.006412845104932785, + 0.015114163979887962, + 0.006000767461955547, + 0.003648133249953389, + 0.009308340959250927, + 0.0045189750380814075, + 0.0022161041852086782, + 0.002120411256328225, + 0.002976133720949292, + 0.0040595270693302155, + 0.0018712998135015368, + 0.0020025307312607765, + 0.003264711471274495, + 0.0024689079727977514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.02013622224330902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007650007028132677 + ], + "opposite_logprob_tokens": [ + -4.882649898529053 + ], + "logprob_reward": -0.007650007028132677, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007650007028132677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03676762804389, + 0.020977171137928963, + 0.017361043021082878, + 0.020803919062018394, + 0.01955794170498848, + 0.06595373153686523, + 0.7442567348480225, + 0.03610721230506897, + 0.014124375768005848, + 0.024090198799967766 + ], + "credit_map": [ + 0.0030808034352958202, + 0.004774400033056736, + 0.0061536310240626335, + 0.0050982399843633175, + 0.0038774539716541767, + 0.003318964736536145, + 0.007177772466093302, + 0.003286364022642374, + 0.0020742223132401705, + 0.0019000128377228975, + 0.0029136415105313063, + 0.002807243959978223, + 0.004148811101913452, + 0.0015149092068895698, + 0.0021131611429154873, + 0.0035051710437983274, + 0.002959355479106307, + 0.0021334087941795588, + 0.0032507767900824547, + 0.0033730000723153353, + 0.0012298536021262407, + 0.0015541717875748873, + 0.001681570429354906, + 0.0011789080454036593, + 0.0024937400594353676, + 0.0037355960812419653, + 0.0021256734617054462, + 0.003571024863049388, + 0.002791144186630845, + 0.002180693903937936, + 0.0015281356172636151, + 0.002377912634983659, + 0.0027710688300430775, + 0.002979795914143324, + 0.0026313206180930138, + 0.002101222286000848, + 0.0027552687097340822, + 0.0019175648922100663, + 0.0018475912511348724, + 0.0025541100185364485, + 0.008086592890322208, + 0.006724100094288588, + 0.004394620191305876, + 0.005616279784590006, + 0.006183120887726545, + 0.005684823729097843, + 0.014616426080465317, + 0.014647776260972023, + 0.034159786999225616, + 0.05323437973856926, + 0.05876743420958519, + 0.03361878916621208, + 0.03338715061545372, + 0.2362976223230362, + 0.2580111026763916, + 0.03678056225180626, + 0.0030688547994941473, + 0.0043855165131390095, + 0.005125467665493488, + 0.0042349109426140785, + 0.004236022476106882, + 0.005734853912144899, + 0.00292771914973855, + 0.006393868941813707, + 0.0013738577254116535, + 0.0014668136136606336, + 0.0016532291192561388, + 0.0029913820326328278, + 0.0019457810558378696, + 0.0013296243268996477, + 0.0017077653901651502, + 0.0016559234354645014, + 0.0026943106204271317, + 0.004380630794912577, + 0.003005757462233305, + 0.004328105598688126, + 0.0018626514356583357, + 0.0024538522120565176, + 0.002849588403478265, + 0.0025153043679893017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.007650007028132677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607036218047142 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607036218047142, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607036218047142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027939362451434135, + 0.02468886785209179, + 0.02768855169415474, + 0.02732572890818119, + 0.018325604498386383, + 0.02736169658601284, + 0.7704403400421143, + 0.02039969339966774, + 0.012287722900509834, + 0.043542470782995224 + ], + "credit_map": [ + 0.004928003065288067, + 0.004027749411761761, + 0.0049248081631958485, + 0.003619344439357519, + 0.002812271937727928, + 0.0024150952231138945, + 0.0027855548541992903, + 0.0024265367537736893, + 0.0025004809722304344, + 0.003434224985539913, + 0.004879890475422144, + 0.0050082080997526646, + 0.0027765047270804644, + 0.0017965398728847504, + 0.0020865504629909992, + 0.0022064680233597755, + 0.004725635517388582, + 0.003335320157930255, + 0.0052694352343678474, + 0.005430682562291622, + 0.0017794073792174459, + 0.0024539525620639324, + 0.002809891477227211, + 0.0018842255230993032, + 0.002819797722622752, + 0.004495661705732346, + 0.004073882009834051, + 0.004904087632894516, + 0.002404287923127413, + 0.002264503389596939, + 0.0019472220446914434, + 0.004416286945343018, + 0.0033052796497941017, + 0.003170429263263941, + 0.001941678230650723, + 0.0019171544117853045, + 0.0027036203537136316, + 0.0022619289811700583, + 0.0012930394150316715, + 0.001732475240714848, + 0.003752073273062706, + 0.004587012343108654, + 0.0022954109590500593, + 0.002451283624395728, + 0.0026413220912218094, + 0.0033032468054443598, + 0.004680747631937265, + 0.0036506005562841892, + 0.0359649658203125, + 0.05182332545518875, + 0.07239402830600739, + 0.03534143418073654, + 0.03398263081908226, + 0.22771063446998596, + 0.2756821811199188, + 0.037541091442108154, + 0.0022996882908046246, + 0.002349608577787876, + 0.002178965834900737, + 0.002329956041648984, + 0.0025862662587314844, + 0.0028402511961758137, + 0.0019311571959406137, + 0.003883799770846963, + 0.0011413736501708627, + 0.001296769711188972, + 0.001285478938370943, + 0.0020692625548690557, + 0.002425741171464324, + 0.0008469587774015963, + 0.0013658441603183746, + 0.0018562932964414358, + 0.004504547454416752, + 0.007535180542618036, + 0.006254099775105715, + 0.006638660095632076, + 0.0034942945931106806, + 0.005191879812628031, + 0.004830203950405121, + 0.0050936052575707436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.01607036218047142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188718050718307 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188718050718307, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188718050718307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02345420978963375, + 0.01791997253894806, + 0.063783198595047, + 0.023669222369790077, + 0.03033480979502201, + 0.021109256893396378, + 0.7499403357505798, + 0.024210063740611076, + 0.02766907960176468, + 0.017909834161400795 + ], + "credit_map": [ + 0.002519781468436122, + 0.004123589489609003, + 0.0038681223522871733, + 0.0033704950474202633, + 0.0028360295109450817, + 0.0019305278547108173, + 0.002603273605927825, + 0.0022023902274668217, + 0.0019226837903261185, + 0.0018422362627461553, + 0.0026888397987931967, + 0.004133838694542646, + 0.003086216514930129, + 0.0011026243446394801, + 0.0016628265148028731, + 0.0014807056868448853, + 0.019141480326652527, + 0.0039433869533240795, + 0.02297557331621647, + 0.006987662985920906, + 0.0018853320507332683, + 0.0028658013325184584, + 0.003774403128772974, + 0.002209556056186557, + 0.002491733292117715, + 0.004408020526170731, + 0.004400031175464392, + 0.0038892237935215235, + 0.0024911912623792887, + 0.0018368249293416739, + 0.0015592329436913133, + 0.0025929640978574753, + 0.0047947559505701065, + 0.004148783162236214, + 0.0026102117262780666, + 0.002410448156297207, + 0.005177672021090984, + 0.006911954376846552, + 0.001782632782123983, + 0.0024983533658087254, + 0.00282280333340168, + 0.002690736437216401, + 0.0015698069473728538, + 0.002067007590085268, + 0.001965123228728771, + 0.002628016984090209, + 0.0037552821449935436, + 0.0036104803439229727, + 0.03709159046411514, + 0.04444810003042221, + 0.08236449956893921, + 0.03671169653534889, + 0.03251674771308899, + 0.1926795244216919, + 0.28210270404815674, + 0.04202550649642944, + 0.0031687768641859293, + 0.00335034541785717, + 0.0027249290142208338, + 0.0022226772271096706, + 0.004331329371780157, + 0.002592565258964896, + 0.0030782807152718306, + 0.0027411612682044506, + 0.0023667511995881796, + 0.003741453867405653, + 0.0032462230883538723, + 0.006845821626484394, + 0.003943303134292364, + 0.0018055365653708577, + 0.0029598786495625973, + 0.002760109957307577, + 0.002126369858160615, + 0.002949159126728773, + 0.002616653684526682, + 0.0024223204236477613, + 0.0021663857623934746, + 0.0013433475978672504, + 0.002449583727866411, + 0.0018360137473791838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.014188718050718307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.03468373045325279, + 0.01467480231076479, + 0.020856881514191628, + 0.0196946170181036, + 0.01705792173743248, + 0.02176007628440857, + 0.8107911944389343, + 0.020224586129188538, + 0.02486780658364296, + 0.015388359315693378 + ], + "credit_map": [ + 0.0038008494302630424, + 0.004648725502192974, + 0.007521708961576223, + 0.005558176897466183, + 0.0029202101286500692, + 0.003037037095054984, + 0.004208249971270561, + 0.00298877339810133, + 0.0017901334213092923, + 0.0021118614822626114, + 0.002755762543529272, + 0.0030730892904102802, + 0.0011875524651259184, + 0.0009612920111976564, + 0.0016332969535142183, + 0.0011618138523772359, + 0.0036162883043289185, + 0.001693781465291977, + 0.006394437048584223, + 0.003998138941824436, + 0.001140135107561946, + 0.0010057049803435802, + 0.0016443614149466157, + 0.0013640327379107475, + 0.0019297379767522216, + 0.002831817604601383, + 0.0025480245240032673, + 0.002220264868810773, + 0.0013112382730469108, + 0.0019479827024042606, + 0.0014454403426498175, + 0.005460112821310759, + 0.0023576414678245783, + 0.0026741281617432833, + 0.0019701679702848196, + 0.0015956128481775522, + 0.0030047984328120947, + 0.0031033155974000692, + 0.0010278631234541535, + 0.0013243942521512508, + 0.003708325559273362, + 0.00439471798017621, + 0.0018669393612071872, + 0.0014073135098442435, + 0.0025612092576920986, + 0.0025299456901848316, + 0.003571777604520321, + 0.0017198475543409586, + 0.03678711876273155, + 0.05925820395350456, + 0.05873628333210945, + 0.03631020709872246, + 0.03687245398759842, + 0.26124683022499084, + 0.281013548374176, + 0.04056659713387489, + 0.0016900026239454746, + 0.0030125784687697887, + 0.0024276492185890675, + 0.0022208960726857185, + 0.00243090046569705, + 0.004554287064820528, + 0.0015081845922395587, + 0.0023800889030098915, + 0.0022560276556760073, + 0.0028771075885742903, + 0.003026347141712904, + 0.006847924552857876, + 0.002846305025741458, + 0.002118597971275449, + 0.002782120369374752, + 0.0021133776754140854, + 0.0020343924406915903, + 0.0015232492005452514, + 0.0023420818615704775, + 0.003187828930094838, + 0.001563470927067101, + 0.0013308190973475575, + 0.001982315443456173, + 0.0014242019969969988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.00319268717430532 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up gazes across the intricate eaves of a traditional Chinese building, where dark, rain-soaked wooden beams, their texture deepened by centuries of weather, reveal microscopic cracks. Black rainwater cautiously seeps through these ancient cracks, slowly seeping through the overlapping, dark gray clay tiles, saturated with moisture and covered in a velvety, deep-green moss. The camera then executes a subtle, slow downward tilt, following a single, viscous stream of water, tracing its path across the curved surface of the tiles. At the lowest edge of the tile, the water gathers into a heavy drop that swells and trembles for a moment, its perfectly spherical surface capturing a distorted, miniature reflection of the tranquil, damp stone courtyard below, before it separates and falls. The entire scene is bathed in the soft, diffuse light of a perpetually gray, overcast sky, creating a cool, subdued color palette of earthy browns, slate grays, and damp greens, evoking an atmosphere of profound tranquility and contemplation.", + "seed": 12131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_7820_seed_12131_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": -0.01945654759183526, + "vlm_logprob": -0.01945654759183526, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.01945654759183526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035721730440855026 + ], + "opposite_logprob_tokens": [ + -3.535721778869629 + ], + "logprob_reward": -0.035721730440855026, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.035721730440855026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08306793123483658, + 0.03334111347794533, + 0.1478656679391861, + 0.2396518588066101, + 0.1540834903717041, + 0.03605026379227638, + 0.057777442038059235, + 0.16301953792572021, + 0.05054926127195358, + 0.03459344431757927 + ], + "credit_map": [ + 0.006776174996048212, + 0.010204419493675232, + 0.018948577344417572, + 0.00939131062477827, + 0.012004644609987736, + 0.011212876997888088, + 0.005157117266207933, + 0.009372802451252937, + 0.0025185986887663603, + 0.003495227312669158, + 0.005809481255710125, + 0.004857788328081369, + 0.0044380915351212025, + 0.005641931667923927, + 0.0026472373865544796, + 0.0039327568374574184, + 0.013073389418423176, + 0.026378465816378593, + 0.028764894232153893, + 0.013752402737736702, + 0.022694429382681847, + 0.02375234290957451, + 0.0097670191898942, + 0.009682726114988327, + 0.04223985597491264, + 0.01894259639084339, + 0.02816709876060486, + 0.01066026370972395, + 0.09722678363323212, + 0.025404976680874825, + 0.007486513815820217, + 0.009523767046630383, + 0.024580752477049828, + 0.019096843898296356, + 0.020399847999215126, + 0.017329752445220947, + 0.015141863375902176, + 0.02847597934305668, + 0.01563887484371662, + 0.013419579714536667, + 0.006186053156852722, + 0.0045298864133656025, + 0.005679272580891848, + 0.005951855797320604, + 0.0030538197606801987, + 0.002978031523525715, + 0.0038872426375746727, + 0.003784102387726307, + 0.009818381629884243, + 0.005722177680581808, + 0.007984038442373276, + 0.008110594935715199, + 0.005203545093536377, + 0.006686918903142214, + 0.008109788410365582, + 0.006141995545476675, + 0.014901009388267994, + 0.005608179606497288, + 0.05190829932689667, + 0.01177894789725542, + 0.005211640149354935, + 0.008104627951979637, + 0.013112151995301247, + 0.05239468067884445, + 0.005031106062233448, + 0.005678567104041576, + 0.006458508316427469, + 0.006370522081851959, + 0.008373988792300224, + 0.006918633822351694, + 0.005788588896393776, + 0.0059293475933372974, + 0.004474211484193802, + 0.004469689447432756, + 0.004613345954567194, + 0.0040983837097883224, + 0.005729306023567915, + 0.0051342798396945, + 0.0026059867814183235, + 0.0034682401455938816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.035721730440855026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02638300135731697 + ], + "opposite_logprob_tokens": [ + -3.6513829231262207 + ], + "logprob_reward": -0.02638300135731697, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02638300135731697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19719356298446655, + 0.12025509774684906, + 0.06972169131040573, + 0.08746886998414993, + 0.09400409460067749, + 0.033611249178647995, + 0.08165019750595093, + 0.11913570016622543, + 0.07557716965675354, + 0.12138233333826065 + ], + "credit_map": [ + 0.01670103333890438, + 0.023227669298648834, + 0.03774912655353546, + 0.023349450901150703, + 0.02630496770143509, + 0.02376355603337288, + 0.01979018934071064, + 0.026307567954063416, + 0.00838424265384674, + 0.010049612261354923, + 0.028751973062753677, + 0.01120083313435316, + 0.012836803682148457, + 0.033105287700891495, + 0.008862986229360104, + 0.0070633552968502045, + 0.0060923281125724316, + 0.012703455053269863, + 0.01384134590625763, + 0.005918454844504595, + 0.011520830914378166, + 0.011397100053727627, + 0.003794885938987136, + 0.004453288856893778, + 0.009738810360431671, + 0.010979586280882359, + 0.015459184534847736, + 0.008856422267854214, + 0.01814763993024826, + 0.010179934091866016, + 0.008027157746255398, + 0.006080135703086853, + 0.015406705439090729, + 0.011358450166881084, + 0.012407183647155762, + 0.012319090776145458, + 0.008456723764538765, + 0.014704462140798569, + 0.010026081465184689, + 0.009325399063527584, + 0.006845514755696058, + 0.004120039287954569, + 0.004520631395280361, + 0.005190529394894838, + 0.0031106420792639256, + 0.0025928786490112543, + 0.0030809412710368633, + 0.004150074440985918, + 0.013862202875316143, + 0.00866974238306284, + 0.01121860183775425, + 0.012366387993097305, + 0.005133326631039381, + 0.007664614822715521, + 0.012681522406637669, + 0.010053794831037521, + 0.012038071639835835, + 0.011686735786497593, + 0.011234195902943611, + 0.017656926065683365, + 0.01233881339430809, + 0.017215967178344727, + 0.021694622933864594, + 0.015270370058715343, + 0.0077799418941140175, + 0.007607464678585529, + 0.006269766949117184, + 0.008877048268914223, + 0.012835630215704441, + 0.013679358176887035, + 0.010002740658819675, + 0.008525222539901733, + 0.01554886344820261, + 0.014281661249697208, + 0.01634041778743267, + 0.013949093408882618, + 0.023706622421741486, + 0.016795972362160683, + 0.009021409787237644, + 0.011738298460841179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02638300135731697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058546058833599 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.02058546058833599, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058546058833599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06941570341587067, + 0.040580812841653824, + 0.04424053058028221, + 0.11164630204439163, + 0.059103935956954956, + 0.05597797408699989, + 0.10786478966474533, + 0.4390565752983093, + 0.02374909445643425, + 0.048364270478487015 + ], + "credit_map": [ + 0.006010880693793297, + 0.011220046319067478, + 0.011728232726454735, + 0.008652918040752411, + 0.008404822088778019, + 0.00821884535253048, + 0.009206648916006088, + 0.005973305553197861, + 0.0026360866613686085, + 0.0037712224293500185, + 0.0090658999979496, + 0.004102821461856365, + 0.004917673300951719, + 0.010068057104945183, + 0.003057976020500064, + 0.0029610740020871162, + 0.003241957165300846, + 0.007954754866659641, + 0.009890236891806126, + 0.0038669072091579437, + 0.0065027866512537, + 0.007438361179083586, + 0.0026359211187809706, + 0.0027096071280539036, + 0.017277846112847328, + 0.009740308858454227, + 0.015098116360604763, + 0.006765426602214575, + 0.04063490033149719, + 0.012445617467164993, + 0.00454690819606185, + 0.005137174390256405, + 0.009809606708586216, + 0.0069501991383731365, + 0.008622825145721436, + 0.006455393973737955, + 0.006165689323097467, + 0.008288227953016758, + 0.00608857162296772, + 0.006723423022776842, + 0.01129559800028801, + 0.0063506606966257095, + 0.0094221830368042, + 0.008661476895213127, + 0.004638724960386753, + 0.0041863322257995605, + 0.00440090661868453, + 0.007022094447165728, + 0.020669803023338318, + 0.0096674133092165, + 0.015419443137943745, + 0.01812804862856865, + 0.008387654088437557, + 0.0087629659101367, + 0.014899492263793945, + 0.011929966509342194, + 0.03819604963064194, + 0.01556870061904192, + 0.12101251631975174, + 0.04059077054262161, + 0.015659848228096962, + 0.018085643649101257, + 0.056827157735824585, + 0.13311588764190674, + 0.0027015493251383305, + 0.002290191128849983, + 0.002886276924982667, + 0.0056688180193305016, + 0.0018963278271257877, + 0.002559512387961149, + 0.003088366938754916, + 0.0026580537669360638, + 0.006295269355177879, + 0.0061180670745670795, + 0.007676302921026945, + 0.005015833303332329, + 0.008641226217150688, + 0.006482674274593592, + 0.003534402698278427, + 0.004600496031343937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.02058546058833599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637331426143646 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637331426143646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637331426143646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04837305098772049, + 0.023571353405714035, + 0.03697720542550087, + 0.15606288611888885, + 0.08335283398628235, + 0.052009254693984985, + 0.1797604113817215, + 0.3497118055820465, + 0.03930426761507988, + 0.030876852571964264 + ], + "credit_map": [ + 0.0044334144331514835, + 0.008978776633739471, + 0.009132529608905315, + 0.005032951012253761, + 0.007267746143043041, + 0.006897118408232927, + 0.0030407130252569914, + 0.0035898073110729456, + 0.0018662132788449526, + 0.003086147364228964, + 0.00435690488666296, + 0.002411992521956563, + 0.003404491813853383, + 0.004286264535039663, + 0.0022242648992687464, + 0.0019350767834112048, + 0.00372342043556273, + 0.005161804147064686, + 0.008606549352407455, + 0.0036023862194269896, + 0.004470185376703739, + 0.006031944416463375, + 0.002313669305294752, + 0.003067251993343234, + 0.022829681634902954, + 0.016059214249253273, + 0.023945925757288933, + 0.009944473393261433, + 0.05381082370877266, + 0.016827240586280823, + 0.00657591363415122, + 0.006069631781429052, + 0.015167618170380592, + 0.0111340107396245, + 0.012027954682707787, + 0.009850556030869484, + 0.007714064326137304, + 0.010907423682510853, + 0.0076983324252069, + 0.008852881379425526, + 0.010064677335321903, + 0.005717470310628414, + 0.011176002211868763, + 0.007261033169925213, + 0.0036002304404973984, + 0.003739710198715329, + 0.0047878301702439785, + 0.005662302486598492, + 0.035735633224248886, + 0.017926275730133057, + 0.027981700375676155, + 0.033683232963085175, + 0.012601113878190517, + 0.01359433401376009, + 0.018037715926766396, + 0.020200422033667564, + 0.031142834573984146, + 0.021189222112298012, + 0.05368443578481674, + 0.0542147196829319, + 0.0204602237790823, + 0.02387244626879692, + 0.07699086517095566, + 0.06815709918737411, + 0.004482886288315058, + 0.00482306070625782, + 0.003833575639873743, + 0.004385222680866718, + 0.006484344135969877, + 0.005976108368486166, + 0.005094875115901232, + 0.004224199336022139, + 0.00417996384203434, + 0.0035896620247513056, + 0.004934719298034906, + 0.003245699219405651, + 0.005618997383862734, + 0.004197685979306698, + 0.0023745624348521233, + 0.0027355654165148735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.008637331426143646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.16501769423484802, + 0.06617914140224457, + 0.12702393531799316, + 0.08160828053951263, + 0.08751381933689117, + 0.024575572460889816, + 0.1206865981221199, + 0.15312930941581726, + 0.0824558436870575, + 0.09180974215269089 + ], + "credit_map": [ + 0.017445527017116547, + 0.028776636347174644, + 0.02628616988658905, + 0.01874726079404354, + 0.01982247829437256, + 0.018902411684393883, + 0.016001317650079727, + 0.01903589628636837, + 0.004556260537356138, + 0.006776439491659403, + 0.014359857887029648, + 0.00958325807005167, + 0.0062688058242201805, + 0.016147905960679054, + 0.004317601211369038, + 0.004169008228927851, + 0.011695374734699726, + 0.022931842133402824, + 0.025374270975589752, + 0.011192019097507, + 0.020312996581196785, + 0.021048638969659805, + 0.007085738237947226, + 0.007383069489151239, + 0.012521514669060707, + 0.009273319505155087, + 0.00954009871929884, + 0.011914985254406929, + 0.010748912580311298, + 0.011725603602826595, + 0.007217565551400185, + 0.008666281588375568, + 0.014249568805098534, + 0.01080721989274025, + 0.011083630844950676, + 0.010281456634402275, + 0.009368066675961018, + 0.013900559395551682, + 0.010009014047682285, + 0.007814302109181881, + 0.004553407896310091, + 0.003053141525015235, + 0.003701008390635252, + 0.0028353845700621605, + 0.0024894478265196085, + 0.0023963917046785355, + 0.0028910376131534576, + 0.002655754331499338, + 0.02672123722732067, + 0.012453698553144932, + 0.020089086145162582, + 0.021863190457224846, + 0.0066921524703502655, + 0.008657945320010185, + 0.010099124163389206, + 0.01411016657948494, + 0.015104052610695362, + 0.0060580517165362835, + 0.04960326477885246, + 0.011967213824391365, + 0.0057646105997264385, + 0.0105153638869524, + 0.007746757939457893, + 0.0463699996471405, + 0.008353055454790592, + 0.007773910649120808, + 0.009302327409386635, + 0.01242086198180914, + 0.013506810180842876, + 0.011965530924499035, + 0.010018185712397099, + 0.009115159511566162, + 0.014808367937803268, + 0.010667894035577774, + 0.014220590703189373, + 0.009452637284994125, + 0.014420946128666401, + 0.012604784220457077, + 0.0063146501779556274, + 0.00931987352669239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.005955214146524668 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up gazes across the intricate eaves of a traditional Chinese building, where dark, rain-soaked wooden beams, their texture deepened by centuries of weather, reveal microscopic cracks. Black rainwater cautiously seeps through these ancient cracks, slowly seeping through the overlapping, dark gray clay tiles, saturated with moisture and covered in a velvety, deep-green moss. The camera then executes a subtle, slow downward tilt, following a single, viscous stream of water, tracing its path across the curved surface of the tiles. At the lowest edge of the tile, the water gathers into a heavy drop that swells and trembles for a moment, its perfectly spherical surface capturing a distorted, miniature reflection of the tranquil, damp stone courtyard below, before it separates and falls. The entire scene is bathed in the soft, diffuse light of a perpetually gray, overcast sky, creating a cool, subdued color palette of earthy browns, slate grays, and damp greens, evoking an atmosphere of profound tranquility and contemplation.", + "seed": 12132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_7820_seed_12132_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": -0.010021823411807417, + "vlm_logprob": -0.010021823411807417, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.010021823411807417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016279971227049828 + ], + "opposite_logprob_tokens": [ + -4.891280174255371 + ], + "logprob_reward": -0.016279971227049828, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.016279971227049828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1243874654173851, + 0.1578688770532608, + 0.05241794511675835, + 0.04169779643416405, + 0.231357142329216, + 0.0645851194858551, + 0.07983412593603134, + 0.05804506316781044, + 0.06294947862625122, + 0.1268569678068161 + ], + "credit_map": [ + 0.016007499769330025, + 0.018982701003551483, + 0.018695810809731483, + 0.018125034868717194, + 0.011162196286022663, + 0.012211741879582405, + 0.01311463676393032, + 0.016087844967842102, + 0.018248334527015686, + 0.01678699254989624, + 0.030863376334309578, + 0.018903015181422234, + 0.023343466222286224, + 0.014175180345773697, + 0.019101684913039207, + 0.016446832567453384, + 0.00716132577508688, + 0.005992974620312452, + 0.005497935693711042, + 0.00472872331738472, + 0.007827987894415855, + 0.004899940453469753, + 0.0071937572211027145, + 0.00911529827862978, + 0.005740225315093994, + 0.007247575558722019, + 0.0034729200415313244, + 0.005760109052062035, + 0.003263745689764619, + 0.003904360346496105, + 0.0056431652046740055, + 0.006665694061666727, + 0.009663431905210018, + 0.00819475669413805, + 0.060548681765794754, + 0.010943984612822533, + 0.005801480729132891, + 0.011141860857605934, + 0.027776645496487617, + 0.09728630632162094, + 0.004896636586636305, + 0.005540778860449791, + 0.00739152729511261, + 0.0084341736510396, + 0.013187741860747337, + 0.008441504091024399, + 0.007189217489212751, + 0.009503541514277458, + 0.006846341770142317, + 0.00884458888322115, + 0.007634310983121395, + 0.015688456594944, + 0.008367236703634262, + 0.006178843788802624, + 0.016222627833485603, + 0.010051717981696129, + 0.004779987037181854, + 0.006534462794661522, + 0.007513824850320816, + 0.008548933081328869, + 0.005418437533080578, + 0.005146670620888472, + 0.010166960768401623, + 0.00993578601628542, + 0.004715356044471264, + 0.0065664430148899555, + 0.007608396932482719, + 0.010189502499997616, + 0.0066598248668015, + 0.006759627256542444, + 0.009081719443202019, + 0.01136860903352499, + 0.007132702972739935, + 0.024829266592860222, + 0.03257115185260773, + 0.012040135450661182, + 0.015008415095508099, + 0.016180075705051422, + 0.011855627410113811, + 0.007239578291773796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 9 + ], + "selected_reward": -0.016279971227049828 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008674440905451775 + ], + "opposite_logprob_tokens": [ + -4.758674621582031 + ], + "logprob_reward": -0.008674440905451775, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008674440905451775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03663695231080055, + 0.11593512445688248, + 0.017465872690081596, + 0.04025771841406822, + 0.2543538510799408, + 0.02146780490875244, + 0.02167975716292858, + 0.0527893528342247, + 0.22947460412979126, + 0.20993900299072266 + ], + "credit_map": [ + 0.005033012013882399, + 0.004258567932993174, + 0.006950279697775841, + 0.004981453996151686, + 0.003250596346333623, + 0.0037288221064954996, + 0.004169214051216841, + 0.004265007562935352, + 0.011520392261445522, + 0.013723821379244328, + 0.021127089858055115, + 0.011683211661875248, + 0.02001068741083145, + 0.011793472804129124, + 0.013275804929435253, + 0.012800642289221287, + 0.0036473653744906187, + 0.0023750634863972664, + 0.0015098873991519213, + 0.0017814337043091655, + 0.001325554447248578, + 0.001694409642368555, + 0.002512178849428892, + 0.0026199803687632084, + 0.005541282705962658, + 0.0035378881730139256, + 0.004146608058363199, + 0.00569387711584568, + 0.004138745833188295, + 0.003398425877094269, + 0.006392162758857012, + 0.007408728823065758, + 0.01096814963966608, + 0.009059788659214973, + 0.06692022830247879, + 0.015128662809729576, + 0.006306204479187727, + 0.012989760376513004, + 0.03163452073931694, + 0.10134651511907578, + 0.003095863852649927, + 0.002151509514078498, + 0.0029983462300151587, + 0.002737998263910413, + 0.002654954558238387, + 0.0020892962347716093, + 0.002335075754672289, + 0.003404760966077447, + 0.0018230574205517769, + 0.0035916974302381277, + 0.0023956631775945425, + 0.0034123435616493225, + 0.002463045297190547, + 0.0021497318521142006, + 0.0030939418356865644, + 0.0027502768207341433, + 0.004350155126303434, + 0.004002899397164583, + 0.009695989079773426, + 0.006588367745280266, + 0.004336091224104166, + 0.004230680875480175, + 0.011313259601593018, + 0.008271913044154644, + 0.015612888149917126, + 0.024676145985722542, + 0.03003876283764839, + 0.03850128501653671, + 0.0258945282548666, + 0.024213282391428947, + 0.030690256506204605, + 0.03984745591878891, + 0.009687847457826138, + 0.05023907497525215, + 0.05783945322036743, + 0.019676178693771362, + 0.016565732657909393, + 0.029638489708304405, + 0.016044102609157562, + 0.01024811901152134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.008674440905451775 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790970012545586 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009790970012545586, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790970012545586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054524097591638565, + 0.06974189728498459, + 0.026369431987404823, + 0.04557018354535103, + 0.5139818787574768, + 0.02299191802740097, + 0.027220692485570908, + 0.023829255253076553, + 0.14878278970718384, + 0.06698789447546005 + ], + "credit_map": [ + 0.006920208223164082, + 0.00779846403747797, + 0.009323343634605408, + 0.0066274660639464855, + 0.00659693730995059, + 0.006416869815438986, + 0.004939054138958454, + 0.00590174924582243, + 0.006398847792297602, + 0.006680624559521675, + 0.011848924681544304, + 0.007794038392603397, + 0.008165296167135239, + 0.00600265571847558, + 0.010726481676101685, + 0.012125025503337383, + 0.004785780794918537, + 0.0034073200076818466, + 0.00266095413826406, + 0.0029377792961895466, + 0.0019193588523194194, + 0.0022986906114965677, + 0.003409413620829582, + 0.004950131755322218, + 0.006559626664966345, + 0.0043335165828466415, + 0.004865260794758797, + 0.006038106046617031, + 0.004819665104150772, + 0.0037971329875290394, + 0.007009647786617279, + 0.008147223852574825, + 0.021055394783616066, + 0.011694841086864471, + 0.14576080441474915, + 0.024747973307967186, + 0.007340215612202883, + 0.022928565740585327, + 0.062387969344854355, + 0.21806608140468597, + 0.003152711084112525, + 0.002226146636530757, + 0.0033745693508535624, + 0.0033339213114231825, + 0.0025473518762737513, + 0.0023835995234549046, + 0.002786058932542801, + 0.0031875576823949814, + 0.0024805497378110886, + 0.004352523945271969, + 0.002953016199171543, + 0.004491443280130625, + 0.0022419430315494537, + 0.0020103957504034042, + 0.005752321798354387, + 0.002938496647402644, + 0.0016914502484723926, + 0.0024680523201823235, + 0.0025165106635540724, + 0.0034758434630930424, + 0.002392357913777232, + 0.0027253972366452217, + 0.0036500876303762197, + 0.004909553099423647, + 0.009944926016032696, + 0.014993124641478062, + 0.02624443545937538, + 0.02126234583556652, + 0.010638438165187836, + 0.01266681868582964, + 0.0347183495759964, + 0.018314342945814133, + 0.004775810055434704, + 0.012550684623420238, + 0.008996902033686638, + 0.00574323209002614, + 0.010529433377087116, + 0.009483816102147102, + 0.007714547216892242, + 0.007193457335233688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.009790970012545586 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538178823888302 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538178823888302, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538178823888302, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06565811485052109, + 0.10120368003845215, + 0.025902679190039635, + 0.024203000590205193, + 0.3045574128627777, + 0.023553941398859024, + 0.04317881539463997, + 0.04175112769007683, + 0.04959521070122719, + 0.3203960657119751 + ], + "credit_map": [ + 0.009339950978755951, + 0.010303661227226257, + 0.009769456461071968, + 0.009262356907129288, + 0.006654228083789349, + 0.004983133636415005, + 0.0072802589274942875, + 0.008065064437687397, + 0.01246689073741436, + 0.011209516786038876, + 0.02460085041821003, + 0.009757241234183311, + 0.015690727159380913, + 0.00924098864197731, + 0.009352728724479675, + 0.00888473354279995, + 0.005381395574659109, + 0.002877255203202367, + 0.00282331183552742, + 0.0025715481024235487, + 0.0029077327344566584, + 0.0021999815944582224, + 0.0032184747979044914, + 0.003922977484762669, + 0.00395085709169507, + 0.0035741778556257486, + 0.002385158557444811, + 0.0031426595523953438, + 0.0019321561558172107, + 0.0018747719004750252, + 0.002466714009642601, + 0.004876505583524704, + 0.013428409583866596, + 0.010217895731329918, + 0.08421368151903152, + 0.01611277274787426, + 0.005521138198673725, + 0.014051665551960468, + 0.037757039070129395, + 0.12325478345155716, + 0.0022375169210135937, + 0.0022788478527218103, + 0.003200296312570572, + 0.0035109941381961107, + 0.0032421723008155823, + 0.002256667008623481, + 0.003007404273375869, + 0.003820043057203293, + 0.0028658320661634207, + 0.0044575948268175125, + 0.0036829523742198944, + 0.008856292814016342, + 0.0026049555744975805, + 0.002356762532144785, + 0.01409964356571436, + 0.004254783969372511, + 0.0039008555468171835, + 0.0045451391488313675, + 0.003913194872438908, + 0.005175530910491943, + 0.004764098674058914, + 0.0037744129076600075, + 0.00859974417835474, + 0.007078150287270546, + 0.0041304887272417545, + 0.005005742888897657, + 0.007050273008644581, + 0.009124375879764557, + 0.0038484763354063034, + 0.006251573096960783, + 0.009216791950166225, + 0.00496748648583889, + 0.014260068535804749, + 0.0901845321059227, + 0.08004618436098099, + 0.025898655876517296, + 0.022480569779872894, + 0.04656147584319115, + 0.02257406897842884, + 0.018390510231256485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.012538178823888302 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028255560901015997 + ], + "opposite_logprob_tokens": [ + -5.877825736999512 + ], + "logprob_reward": -0.0028255560901015997, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028255560901015997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04584464058279991, + 0.02536185085773468, + 0.022128162905573845, + 0.015384381636977196, + 0.4617350101470947, + 0.012309418991208076, + 0.017420664429664612, + 0.02855037897825241, + 0.07018903642892838, + 0.3010764718055725 + ], + "credit_map": [ + 0.006102842278778553, + 0.005781217012554407, + 0.008048545569181442, + 0.006279615219682455, + 0.006413604132831097, + 0.00487090926617384, + 0.004041685257107019, + 0.004306221846491098, + 0.0024508966598659754, + 0.0031504700891673565, + 0.0030514716636389494, + 0.002724212594330311, + 0.003520742990076542, + 0.002888425486162305, + 0.0036555586848407984, + 0.003920074552297592, + 0.0027679933700710535, + 0.0023583623114973307, + 0.0018532811664044857, + 0.002887249458581209, + 0.0026189698837697506, + 0.0023577010724693537, + 0.0032633491791784763, + 0.004021261818706989, + 0.0024360870011150837, + 0.00190567912068218, + 0.0015723125543445349, + 0.0017455719644203782, + 0.0015789818717166781, + 0.0016813422553241253, + 0.0020900708623230457, + 0.0023743368219584227, + 0.01835673861205578, + 0.009344120509922504, + 0.1283900886774063, + 0.022148480638861656, + 0.005570731591433287, + 0.021182898432016373, + 0.05588800460100174, + 0.20085397362709045, + 0.0013445387594401836, + 0.0012272371677681804, + 0.0014022353570908308, + 0.00197256775572896, + 0.001425304333679378, + 0.0013755481922999024, + 0.00220142281614244, + 0.001360565540380776, + 0.0012779418611899018, + 0.0017811954021453857, + 0.001699371263384819, + 0.003323460230603814, + 0.0012485216138884425, + 0.0010036514140665531, + 0.004809044301509857, + 0.0022774802055209875, + 0.0025205162819474936, + 0.0026252795942127705, + 0.004153575748205185, + 0.0038464253302663565, + 0.0026014831382781267, + 0.00209982437081635, + 0.006288621108978987, + 0.004414655733853579, + 0.004897699225693941, + 0.006827987730503082, + 0.009049918502569199, + 0.011709535494446754, + 0.008303048089146614, + 0.007251945324242115, + 0.009796904399991035, + 0.01235200185328722, + 0.014370047487318516, + 0.08784511685371399, + 0.07178803533315659, + 0.023005886003375053, + 0.018968157470226288, + 0.04587479680776596, + 0.02108457125723362, + 0.01813986897468567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.0028255560901015997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle close-up gazes across the intricate eaves of a traditional Chinese building, where dark, rain-soaked wooden beams, their texture deepened by centuries of weather, reveal microscopic cracks. Black rainwater cautiously seeps through these ancient cracks, slowly seeping through the overlapping, dark gray clay tiles, saturated with moisture and covered in a velvety, deep-green moss. The camera then executes a subtle, slow downward tilt, following a single, viscous stream of water, tracing its path across the curved surface of the tiles. At the lowest edge of the tile, the water gathers into a heavy drop that swells and trembles for a moment, its perfectly spherical surface capturing a distorted, miniature reflection of the tranquil, damp stone courtyard below, before it separates and falls. The entire scene is bathed in the soft, diffuse light of a perpetually gray, overcast sky, creating a cool, subdued color palette of earthy browns, slate grays, and damp greens, evoking an atmosphere of profound tranquility and contemplation.", + "seed": 12133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_14/index_7820_seed_12133_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": -0.016379180178046227, + "vlm_logprob": -0.016379180178046227, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.016379180178046227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028553903102874756 + ], + "opposite_logprob_tokens": [ + -3.7785539627075195 + ], + "logprob_reward": -0.028553903102874756, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028553903102874756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1570710837841034, + 0.10517486929893494, + 0.14835584163665771, + 0.07676779478788376, + 0.0775570347905159, + 0.10920744389295578, + 0.07515021413564682, + 0.05576488748192787, + 0.12918636202812195, + 0.06576444953680038 + ], + "credit_map": [ + 0.011978410184383392, + 0.022427858784794807, + 0.016288941726088524, + 0.021758191287517548, + 0.02141847088932991, + 0.017567003145813942, + 0.021764744073152542, + 0.02386748231947422, + 0.014127080328762531, + 0.020943623036146164, + 0.012831310741603374, + 0.006922287400811911, + 0.018326198682188988, + 0.011815343983471394, + 0.010323645547032356, + 0.00988538283854723, + 0.013157387264072895, + 0.01943392865359783, + 0.017681127414107323, + 0.015477396547794342, + 0.02279963716864586, + 0.0267501063644886, + 0.018832717090845108, + 0.014223547652363777, + 0.007439741864800453, + 0.007054401095956564, + 0.00853731483221054, + 0.009835374541580677, + 0.010725756175816059, + 0.016420351341366768, + 0.006927328184247017, + 0.00982753187417984, + 0.011245098896324635, + 0.011706803925335407, + 0.009326771833002567, + 0.008692281320691109, + 0.00598490796983242, + 0.011557006277143955, + 0.010595100931823254, + 0.008449072018265724, + 0.01036455575376749, + 0.010606433264911175, + 0.014633103273808956, + 0.019276827573776245, + 0.012339080683887005, + 0.00797192845493555, + 0.018126346170902252, + 0.015889175236225128, + 0.006477302871644497, + 0.010175555944442749, + 0.012513330206274986, + 0.015239602886140347, + 0.0066252113319933414, + 0.005329824518412352, + 0.012500174343585968, + 0.006289223674684763, + 0.005020215176045895, + 0.006939414422959089, + 0.00829723197966814, + 0.014734704978764057, + 0.004273504018783569, + 0.004625098314136267, + 0.005311836022883654, + 0.00656288955360651, + 0.011798816733062267, + 0.018384795635938644, + 0.015771185979247093, + 0.0349501334130764, + 0.012790914624929428, + 0.013960403390228748, + 0.010654964484274387, + 0.010875160805881023, + 0.005726183298975229, + 0.01614983007311821, + 0.00886601023375988, + 0.004487465135753155, + 0.011728369630873203, + 0.0034778958652168512, + 0.007963685318827629, + 0.007365011144429445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.028553903102874756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00867101363837719 + ], + "opposite_logprob_tokens": [ + -4.758670806884766 + ], + "logprob_reward": -0.00867101363837719, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00867101363837719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10962431877851486, + 0.08764303475618362, + 0.18579691648483276, + 0.12209451198577881, + 0.05150147154927254, + 0.03738692030310631, + 0.045942433178424835, + 0.07575573772192001, + 0.22423575818538666, + 0.06001894176006317 + ], + "credit_map": [ + 0.008227634243667126, + 0.018066374585032463, + 0.01358547993004322, + 0.011787718161940575, + 0.023441055789589882, + 0.011628159321844578, + 0.00884339865297079, + 0.014044494368135929, + 0.014053707011044025, + 0.014733078889548779, + 0.021113259717822075, + 0.005355721805244684, + 0.013106537982821465, + 0.008176344446837902, + 0.004852653481066227, + 0.006251727230846882, + 0.014646309427917004, + 0.04499903693795204, + 0.022846629843115807, + 0.010396325960755348, + 0.02228042669594288, + 0.03540332615375519, + 0.021189622581005096, + 0.014035246334969997, + 0.010159607045352459, + 0.00892212800681591, + 0.009905894286930561, + 0.027760814875364304, + 0.011630663648247719, + 0.026587732136249542, + 0.0074042812921106815, + 0.01972338929772377, + 0.007569102570414543, + 0.006651247385889292, + 0.00687845703214407, + 0.0059554195031523705, + 0.004928077105432749, + 0.0070211803540587425, + 0.007327298633754253, + 0.005170692689716816, + 0.0022118480410426855, + 0.003588311839848757, + 0.006754077970981598, + 0.006085166707634926, + 0.0034959912300109863, + 0.004998340271413326, + 0.00609184755012393, + 0.0041613345965743065, + 0.003746594535186887, + 0.006019864697009325, + 0.006570720113813877, + 0.009724371135234833, + 0.0037987874820828438, + 0.0034011402167379856, + 0.008481382392346859, + 0.00419957097619772, + 0.005613074637949467, + 0.00877930037677288, + 0.011620095930993557, + 0.023440737277269363, + 0.004489132668823004, + 0.0044814241118729115, + 0.009167402051389217, + 0.008164573460817337, + 0.016569357365369797, + 0.04188359156250954, + 0.037255264818668365, + 0.034099142998456955, + 0.024376442655920982, + 0.018540063872933388, + 0.026798684149980545, + 0.024713190272450447, + 0.0066428417339921, + 0.014869977720081806, + 0.008342508226633072, + 0.0047873081639409065, + 0.008197301998734474, + 0.004877328407019377, + 0.0071336654946208, + 0.005168010946363211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.00867101363837719 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023294363170862198 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.023294363170862198, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023294363170862198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15947675704956055, + 0.12224972248077393, + 0.11114270240068436, + 0.14493420720100403, + 0.08816003054380417, + 0.052067436277866364, + 0.020167523995041847, + 0.031137734651565552, + 0.17736051976680756, + 0.0933033749461174 + ], + "credit_map": [ + 0.011433704756200314, + 0.025325698778033257, + 0.017249858006834984, + 0.017365917563438416, + 0.026174934580922127, + 0.018134713172912598, + 0.03064919449388981, + 0.01314273290336132, + 0.012168368324637413, + 0.01737126149237156, + 0.022413847967982292, + 0.007841043174266815, + 0.025357577949762344, + 0.013448896817862988, + 0.012428058311343193, + 0.011220667511224747, + 0.007974613457918167, + 0.016630802303552628, + 0.01221874076873064, + 0.010318659245967865, + 0.01193834189325571, + 0.02279582992196083, + 0.019483478739857674, + 0.009782244451344013, + 0.018955253064632416, + 0.015463097020983696, + 0.01413060538470745, + 0.030951015651226044, + 0.011307592503726482, + 0.03425857424736023, + 0.007910516113042831, + 0.011957545764744282, + 0.015828851610422134, + 0.009512539021670818, + 0.010216373018920422, + 0.01277959905564785, + 0.006580984685570002, + 0.010233702138066292, + 0.009336557239294052, + 0.013671421445906162, + 0.005159817170351744, + 0.003928289748728275, + 0.008135109208524227, + 0.009154903702437878, + 0.005162011366337538, + 0.005655674729496241, + 0.009408438578248024, + 0.00546319130808115, + 0.0018332923064008355, + 0.003060961375012994, + 0.0025809244252741337, + 0.00396687351167202, + 0.0016105794347822666, + 0.001550858374685049, + 0.003929427359253168, + 0.001634606858715415, + 0.003342561423778534, + 0.0045580510050058365, + 0.004381534643471241, + 0.007239026948809624, + 0.002597207436338067, + 0.001541713485494256, + 0.00427129166200757, + 0.0032063480466604233, + 0.015225283801555634, + 0.025939825922250748, + 0.0230838805437088, + 0.03921843692660332, + 0.0196403656154871, + 0.015355243347585201, + 0.019970087334513664, + 0.01892741397023201, + 0.007740136235952377, + 0.027294408529996872, + 0.011028182692825794, + 0.004763818811625242, + 0.02018122747540474, + 0.005044272169470787, + 0.01013004407286644, + 0.007121280767023563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.023294363170862198 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182270228862762 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182270228862762, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182270228862762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13123035430908203, + 0.13951539993286133, + 0.09037283062934875, + 0.17916971445083618, + 0.09203796088695526, + 0.08745531737804413, + 0.09804990142583847, + 0.029303304851055145, + 0.10155242681503296, + 0.051312752068042755 + ], + "credit_map": [ + 0.010347879491746426, + 0.019689839333295822, + 0.015821821987628937, + 0.012714783661067486, + 0.01881181262433529, + 0.013660035096108913, + 0.014516705647110939, + 0.025667468085885048, + 0.01769936829805374, + 0.010666292160749435, + 0.047120656818151474, + 0.008179457858204842, + 0.027407905086874962, + 0.00966669712215662, + 0.008221116848289967, + 0.010553909465670586, + 0.005312897730618715, + 0.018002472817897797, + 0.009925140999257565, + 0.006408196408301592, + 0.0106869637966156, + 0.011485815048217773, + 0.020310385152697563, + 0.008240955881774426, + 0.016108669340610504, + 0.013210969977080822, + 0.01552984956651926, + 0.03194284811615944, + 0.018037233501672745, + 0.04143688082695007, + 0.01615934446454048, + 0.026743926107883453, + 0.012338033877313137, + 0.011849073693156242, + 0.01167422067373991, + 0.011202597990632057, + 0.009347774088382721, + 0.011248549446463585, + 0.011445257812738419, + 0.012932457961142063, + 0.006148610729724169, + 0.009580778889358044, + 0.0142245814204216, + 0.013346333988010883, + 0.008840437978506088, + 0.011003510095179081, + 0.015144173055887222, + 0.009166885167360306, + 0.007928273640573025, + 0.016366256400942802, + 0.013076026923954487, + 0.013424476608633995, + 0.012737021781504154, + 0.010690940544009209, + 0.014883754774928093, + 0.008943147957324982, + 0.0030997309368103743, + 0.004456074442714453, + 0.004425596445798874, + 0.005155184306204319, + 0.0031745433807373047, + 0.002269691787660122, + 0.004012495279312134, + 0.002709989668801427, + 0.008677290752530098, + 0.014793794602155685, + 0.01404412742704153, + 0.018511656671762466, + 0.011862272396683693, + 0.008979402482509613, + 0.01396574079990387, + 0.010718148201704025, + 0.005319293588399887, + 0.011783143505454063, + 0.006593614351004362, + 0.0034678231459110975, + 0.008443739265203476, + 0.0028569477144628763, + 0.007681482005864382, + 0.005166708491742611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.018182270228862762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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, low-angle close-up of the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are dark, rain-soaked wooden beams with visible microscopic cracks and deep-green moss visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single viscous stream of water seep through the cracks and flow across the dark gray clay tiles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a subtle, slow downward tilt following the water stream to a heavy drop that swells and trembles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in soft, diffuse light from a gray overcast sky with a color palette of earthy browns, slate grays, and damp greens?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031943507492542267 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.0031943507492542267, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031943507492542267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1099691390991211, + 0.13549652695655823, + 0.15826578438282013, + 0.07196184247732162, + 0.09143983572721481, + 0.03608130291104317, + 0.05176682025194168, + 0.05458439141511917, + 0.21706923842430115, + 0.07336517423391342 + ], + "credit_map": [ + 0.011882306076586246, + 0.02120412141084671, + 0.012346007861196995, + 0.012978291139006615, + 0.010550180450081825, + 0.012765850871801376, + 0.00931521411985159, + 0.018927155062556267, + 0.020155349746346474, + 0.034029677510261536, + 0.0259132981300354, + 0.008659576997160912, + 0.017950519919395447, + 0.011585787869989872, + 0.008901403285562992, + 0.0083009023219347, + 0.01243277546018362, + 0.03602857515215874, + 0.017574718222022057, + 0.010167309083044529, + 0.01644960604608059, + 0.030008668079972267, + 0.021364407613873482, + 0.014239723794162273, + 0.007696796208620071, + 0.00830824300646782, + 0.006713974755257368, + 0.0152964498847723, + 0.014359593391418457, + 0.00925640482455492, + 0.0043955096043646336, + 0.005934860557317734, + 0.013115772977471352, + 0.011110290884971619, + 0.009901963174343109, + 0.009877062402665615, + 0.009588745422661304, + 0.013060799799859524, + 0.011544236913323402, + 0.013240952044725418, + 0.0036938826087862253, + 0.0031798139680176973, + 0.004792351275682449, + 0.005847647786140442, + 0.0031373563688248396, + 0.005128533113747835, + 0.005175260826945305, + 0.0051264530047774315, + 0.003892345819622278, + 0.006093122065067291, + 0.00725515466183424, + 0.01342532318085432, + 0.004388976842164993, + 0.0036926104221493006, + 0.00783916749060154, + 0.005180112086236477, + 0.005234191659837961, + 0.007486182730644941, + 0.007293961942195892, + 0.01115307305008173, + 0.005250981543213129, + 0.005126244854182005, + 0.007305408827960491, + 0.005734339822083712, + 0.017014814540743828, + 0.02704142965376377, + 0.029042871668934822, + 0.051127441227436066, + 0.02200184389948845, + 0.015598616562783718, + 0.023393496870994568, + 0.03184869885444641, + 0.006295814644545317, + 0.0172763392329216, + 0.009187105111777782, + 0.006020365748554468, + 0.013609116896986961, + 0.0057930476032197475, + 0.009001278318464756, + 0.006182095035910606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0031943507492542267 + } + ], + "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/0000160/group_3/index_tensor([1684, 1684])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_tensor([1684, 1684])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3736f595929e8b780d88e81a777ac52f254f96d0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_tensor([1684, 1684])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot in a hyperrealistic documentary style, the camera moves in a slow, smooth orbit, rotating 360 degrees around an ornate sedan chair as it navigates a rugged mountain path. It's an intricately carved vermilion lacquered wooden sedan, its sides intricately carved with golden dragon and phoenix reliefs. Its roof is covered in bright yellow glazed tiles, and its corners are draped with silk tassels that sway gently as the sedan rolls. Four bearers, dressed in simple dark blue coarse cloth jackets, steadily carry the heavy poles with muscular shoulders. Their weathered faces express concentration, and each step is remarkably steady and powerful, creating a stark contrast to the sedan's rhythmic swaying caused by the uneven road. The afternoon sun penetrated the gaps in the dense and intertwined treetops overhead, forming sharp beams of light, casting ever-shifting and ever-changing dappled light and shadows on the smooth red paint surface of the sedan chair. The light and shadows flowed as the team moved forward, highlighting the texture of the wood and the depth of the carvings, creating an atmosphere of solemnity, primitiveness and dynamic beauty.", + "seed": 23434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23434_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": -0.12467329986393452, + "vlm_logprob": -0.12467329986393452, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.12467329986393452, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028522158041596413 + ], + "opposite_logprob_tokens": [ + -3.778522253036499 + ], + "logprob_reward": -0.028522158041596413, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028522158041596413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1274574100971222, + 0.1006222665309906, + 0.02840670570731163, + 0.03563402593135834, + 0.04818159341812134, + 0.43992045521736145, + 0.11156968772411346, + 0.053255870938301086, + 0.033791348338127136, + 0.021160611882805824 + ], + "credit_map": [ + 0.018350642174482346, + 0.016856005415320396, + 0.01863260380923748, + 0.011855842545628548, + 0.0169630479067564, + 0.016168076545000076, + 0.014838830567896366, + 0.01379235927015543, + 0.011707473546266556, + 0.009273862466216087, + 0.02008783258497715, + 0.008698763325810432, + 0.006564387120306492, + 0.019664254039525986, + 0.009810872375965118, + 0.014814818277955055, + 0.004401395097374916, + 0.003653324209153652, + 0.002740174997597933, + 0.0049524675123393536, + 0.0034254679922014475, + 0.0036222110502421856, + 0.003576114773750305, + 0.002035551005974412, + 0.004896989092230797, + 0.0036626565270125866, + 0.005129768513143063, + 0.0059024877846241, + 0.004817394074052572, + 0.004268888849765062, + 0.003693317761644721, + 0.0032625235617160797, + 0.010488864965736866, + 0.004079840611666441, + 0.0031963528599590063, + 0.003641808405518532, + 0.00410029012709856, + 0.014709641225636005, + 0.004649530164897442, + 0.0033152643591165543, + 0.15202046930789948, + 0.04671687260270119, + 0.042366646230220795, + 0.025570735335350037, + 0.032377488911151886, + 0.055934689939022064, + 0.024349458515644073, + 0.060584090650081635, + 0.023274866864085197, + 0.007934006862342358, + 0.03504016622900963, + 0.008407382294535637, + 0.008769304491579533, + 0.01384724210947752, + 0.006925967987626791, + 0.007370748091489077, + 0.00794918928295374, + 0.004892601165920496, + 0.005367145407944918, + 0.00642384821549058, + 0.006161301862448454, + 0.01268366165459156, + 0.004629658069461584, + 0.005148464348167181, + 0.005446635652333498, + 0.0053527806885540485, + 0.0028137238696217537, + 0.007571428082883358, + 0.0027460430283099413, + 0.0046444786712527275, + 0.0027898952830582857, + 0.002426362130790949, + 0.002852373756468296, + 0.0028213979676365852, + 0.0028130575083196163, + 0.0025270713958889246, + 0.0038978669326752424, + 0.002351014642044902, + 0.002414460526779294, + 0.001483369036577642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.028522158041596413 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023302864283323288 + ], + "opposite_logprob_tokens": [ + -3.7733027935028076 + ], + "logprob_reward": -0.023302864283323288, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023302864283323288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26341280341148376, + 0.22065208852291107, + 0.02702200785279274, + 0.02740737795829773, + 0.013242960907518864, + 0.1022702008485794, + 0.08189759403467178, + 0.15120021998882294, + 0.0139173474162817, + 0.09897742420434952 + ], + "credit_map": [ + 0.036529265344142914, + 0.02798186056315899, + 0.06870338320732117, + 0.023522676900029182, + 0.02088705077767372, + 0.023102618753910065, + 0.016820406541228294, + 0.04586556553840637, + 0.02601112425327301, + 0.02128598652780056, + 0.04252268746495247, + 0.0190097875893116, + 0.014062062837183475, + 0.043540529906749725, + 0.019691290333867073, + 0.03452862799167633, + 0.004710223525762558, + 0.0021062414161860943, + 0.0020822803489863873, + 0.00299708335660398, + 0.0028111739084124565, + 0.006190205458551645, + 0.0034563743975013494, + 0.002668425440788269, + 0.00422328757122159, + 0.0028529984410852194, + 0.0031199180521070957, + 0.004504861310124397, + 0.0038963425904512405, + 0.0028441804461181164, + 0.003637644462287426, + 0.002328145084902644, + 0.0016153736505657434, + 0.0013589489972218871, + 0.0011878380319103599, + 0.0027267413679510355, + 0.0016736974939703941, + 0.0024125322233885527, + 0.0014831407461315393, + 0.000784688163548708, + 0.04014620557427406, + 0.008014371618628502, + 0.009087713435292244, + 0.004263742361217737, + 0.007550419308245182, + 0.013912354595959187, + 0.0038553380873054266, + 0.015440051443874836, + 0.009743050672113895, + 0.007664206437766552, + 0.020055867731571198, + 0.008247394114732742, + 0.007201886270195246, + 0.012637109495699406, + 0.008350013755261898, + 0.007998058572411537, + 0.03271430358290672, + 0.010976293124258518, + 0.011613411828875542, + 0.018321428447961807, + 0.01625538431107998, + 0.01830899529159069, + 0.012187544256448746, + 0.030822860077023506, + 0.002097887685522437, + 0.0017343598883599043, + 0.0013895726297050714, + 0.0023449042346328497, + 0.0012435406679287553, + 0.0025569358840584755, + 0.0014780597994104028, + 0.0010720877908170223, + 0.011949126608669758, + 0.01779518648982048, + 0.00601592380553484, + 0.013943362049758434, + 0.019253887236118317, + 0.00929643027484417, + 0.01543592382222414, + 0.00528757693246007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.023302864283323288 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038105446845293045 + ], + "opposite_logprob_tokens": [ + -3.2881054878234863 + ], + "logprob_reward": -0.038105446845293045, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038105446845293045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26866570115089417, + 0.12623780965805054, + 0.10178714245557785, + 0.0610194057226181, + 0.060604725033044815, + 0.11730197817087173, + 0.09118662774562836, + 0.030511708930134773, + 0.04005810245871544, + 0.10262681543827057 + ], + "credit_map": [ + 0.037177491933107376, + 0.02535572275519371, + 0.0717720165848732, + 0.02584693394601345, + 0.018841136246919632, + 0.024159550666809082, + 0.018590455874800682, + 0.046922385692596436, + 0.015666253864765167, + 0.01096926350146532, + 0.02515297755599022, + 0.01221053209155798, + 0.007889399304986, + 0.020504703745245934, + 0.014666078612208366, + 0.019178593531250954, + 0.02343452163040638, + 0.006119683850556612, + 0.006565596908330917, + 0.008215642534196377, + 0.0090719573199749, + 0.030621709302067757, + 0.008328499272465706, + 0.009429528377950191, + 0.009364703670144081, + 0.006398837547749281, + 0.006667399778962135, + 0.007721050176769495, + 0.005838195327669382, + 0.00861256942152977, + 0.0076850419864058495, + 0.008731603622436523, + 0.0068058231845498085, + 0.0057458593510091305, + 0.005123835522681475, + 0.010047131218016148, + 0.006823607254773378, + 0.013723697513341904, + 0.005693139508366585, + 0.006641628686338663, + 0.04471605643630028, + 0.009051566012203693, + 0.012804451398551464, + 0.0057489038445055485, + 0.007156054954975843, + 0.01581464521586895, + 0.004836886655539274, + 0.0171734020113945, + 0.010939639993011951, + 0.010108118876814842, + 0.0150893060490489, + 0.011290504597127438, + 0.008839389309287071, + 0.01590685173869133, + 0.01169351115822792, + 0.007319304626435041, + 0.005244294181466103, + 0.0028342404402792454, + 0.0036508168559521437, + 0.002945203799754381, + 0.0037128624971956015, + 0.004369406495243311, + 0.0036979352589696646, + 0.004056947305798531, + 0.005303485784679651, + 0.005801504943519831, + 0.0037428070791065693, + 0.009540465660393238, + 0.0033055925741791725, + 0.005723302718251944, + 0.0038637849502265453, + 0.002777154790237546, + 0.01240623090416193, + 0.01786220446228981, + 0.006846063770353794, + 0.014442007057368755, + 0.017985576763749123, + 0.009674189612269402, + 0.01726382225751877, + 0.006146715953946114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.038105446845293045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813597619533539 + ], + "opposite_logprob_tokens": [ + -1.4063597917556763 + ], + "logprob_reward": -0.2813597619533539, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813597619533539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44617733359336853, + 0.09515570104122162, + 0.07680463790893555, + 0.06145817041397095, + 0.03734177350997925, + 0.043714553117752075, + 0.12588895857334137, + 0.045700859278440475, + 0.030910661444067955, + 0.0368473194539547 + ], + "credit_map": [ + 0.06217560917139053, + 0.04686935991048813, + 0.11569229513406754, + 0.042384352535009384, + 0.03339270502328873, + 0.039105772972106934, + 0.02840743586421013, + 0.07814980298280716, + 0.012743567116558552, + 0.007530189119279385, + 0.013013867661356926, + 0.010376005433499813, + 0.006128829903900623, + 0.02489844337105751, + 0.009942788630723953, + 0.010522015392780304, + 0.015494643710553646, + 0.0077926176600158215, + 0.004932774696499109, + 0.0046799322590231895, + 0.007522766478359699, + 0.01899450458586216, + 0.00935608521103859, + 0.00803131703287363, + 0.010251609608530998, + 0.006076943129301071, + 0.00599642051383853, + 0.011743228882551193, + 0.007887559942901134, + 0.007218656595796347, + 0.006429241970181465, + 0.005854510702192783, + 0.005922134965658188, + 0.003907593432813883, + 0.00353460805490613, + 0.008029764518141747, + 0.0041961390525102615, + 0.004186888225376606, + 0.003843165934085846, + 0.003721476998180151, + 0.012465385720133781, + 0.006388487759977579, + 0.004244678188115358, + 0.00379277509637177, + 0.004313483368605375, + 0.003942524548619986, + 0.0028030690737068653, + 0.005764143541455269, + 0.015441679395735264, + 0.014777909964323044, + 0.021918902173638344, + 0.01636737771332264, + 0.01250323373824358, + 0.02037316933274269, + 0.015192936174571514, + 0.009313755668699741, + 0.008784055709838867, + 0.005888768937438726, + 0.004853429738432169, + 0.003919878508895636, + 0.0036927596665918827, + 0.009368039667606354, + 0.00277827144600451, + 0.006415656302124262, + 0.004558699671179056, + 0.004007813986390829, + 0.004082065541297197, + 0.004434679634869099, + 0.002617766847833991, + 0.005147362593561411, + 0.002863311441615224, + 0.0031989619601517916, + 0.004703095182776451, + 0.005968938115984201, + 0.003566039726138115, + 0.007158529944717884, + 0.005994231440126896, + 0.0028113904409110546, + 0.00463644415140152, + 0.0020086518488824368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.2813597619533539 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207626819610596 + ], + "opposite_logprob_tokens": [ + -1.502076268196106 + ], + "logprob_reward": -0.25207626819610596, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207626819610596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3526969850063324, + 0.06487555056810379, + 0.14109310507774353, + 0.023326242342591286, + 0.035532716661691666, + 0.16286520659923553, + 0.14171108603477478, + 0.031919267028570175, + 0.02778010256588459, + 0.018199719488620758 + ], + "credit_map": [ + 0.04781896993517876, + 0.03849983587861061, + 0.08970536291599274, + 0.03210101276636124, + 0.028882671147584915, + 0.03221957013010979, + 0.02383977174758911, + 0.05962981656193733, + 0.008279172703623772, + 0.007227157708257437, + 0.015259147621691227, + 0.005307579413056374, + 0.0038578847888857126, + 0.00878837238997221, + 0.008152562193572521, + 0.008003675378859043, + 0.03566720336675644, + 0.007164139300584793, + 0.008741140365600586, + 0.008846752345561981, + 0.01091478206217289, + 0.04716239869594574, + 0.010847168043255806, + 0.011749541386961937, + 0.003167450660839677, + 0.002499655121937394, + 0.0028149557765573263, + 0.003678854089230299, + 0.002962701488286257, + 0.002364198910072446, + 0.0029709998052567244, + 0.002867430681362748, + 0.005090781953185797, + 0.0036020430270582438, + 0.002475317334756255, + 0.006227307487279177, + 0.0030551254749298096, + 0.00820466410368681, + 0.004270879551768303, + 0.0026065984275192022, + 0.062430769205093384, + 0.013177864253520966, + 0.018148919567465782, + 0.007600704673677683, + 0.009186330251395702, + 0.020013341680169106, + 0.0071816211566329, + 0.025125667452812195, + 0.023871708661317825, + 0.007822761312127113, + 0.051136329770088196, + 0.010151089169085026, + 0.007212809287011623, + 0.022610163316130638, + 0.006375954486429691, + 0.012530284933745861, + 0.00642311479896307, + 0.002509834710508585, + 0.0029907256830483675, + 0.0042853932827711105, + 0.002821842208504677, + 0.0037502881605178118, + 0.0031324028968811035, + 0.00600566528737545, + 0.004456590861082077, + 0.002905946457758546, + 0.0022083488292992115, + 0.003013691632077098, + 0.0025544066447764635, + 0.004412769805639982, + 0.00532573601230979, + 0.002902616746723652, + 0.0027242812793701887, + 0.002634849166497588, + 0.0019916703458875418, + 0.00289648724719882, + 0.00217215227894485, + 0.0018470020731911063, + 0.002551843412220478, + 0.0013814350822940469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.25207626819610596 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a hyperrealistic documentary style, the camera moves in a slow, smooth orbit, rotating 360 degrees around an ornate sedan chair as it navigates a rugged mountain path. It's an intricately carved vermilion lacquered wooden sedan, its sides intricately carved with golden dragon and phoenix reliefs. Its roof is covered in bright yellow glazed tiles, and its corners are draped with silk tassels that sway gently as the sedan rolls. Four bearers, dressed in simple dark blue coarse cloth jackets, steadily carry the heavy poles with muscular shoulders. Their weathered faces express concentration, and each step is remarkably steady and powerful, creating a stark contrast to the sedan's rhythmic swaying caused by the uneven road. The afternoon sun penetrated the gaps in the dense and intertwined treetops overhead, forming sharp beams of light, casting ever-shifting and ever-changing dappled light and shadows on the smooth red paint surface of the sedan chair. The light and shadows flowed as the team moved forward, highlighting the texture of the wood and the depth of the carvings, creating an atmosphere of solemnity, primitiveness and dynamic beauty.", + "seed": 23435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23435_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": -0.103740244358778, + "vlm_logprob": -0.103740244358778, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.103740244358778, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01389579102396965 + ], + "opposite_logprob_tokens": [ + -4.7638959884643555 + ], + "logprob_reward": -0.01389579102396965, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01389579102396965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07059357315301895, + 0.06646539270877838, + 0.05123280733823776, + 0.49607571959495544, + 0.03142251446843147, + 0.07926766574382782, + 0.049179501831531525, + 0.05556940287351608, + 0.058280009776353836, + 0.041913412511348724 + ], + "credit_map": [ + 0.008178593590855598, + 0.0074370973743498325, + 0.005655115470290184, + 0.0106273228302598, + 0.006036914419382811, + 0.009202596731483936, + 0.012600593268871307, + 0.010855339467525482, + 0.004554587882012129, + 0.008462577126920223, + 0.008773610927164555, + 0.010137644596397877, + 0.008921480737626553, + 0.008358515799045563, + 0.009649957530200481, + 0.007607016246765852, + 0.0035285090561956167, + 0.006954960059374571, + 0.003270311513915658, + 0.007882682606577873, + 0.006049398798495531, + 0.010135113261640072, + 0.008012465201318264, + 0.005399366375058889, + 0.027583273127675056, + 0.020829593762755394, + 0.019595958292484283, + 0.04483853280544281, + 0.031064007431268692, + 0.19825586676597595, + 0.12986622750759125, + 0.02404228039085865, + 0.0029258376453071833, + 0.004092731513082981, + 0.0030508795753121376, + 0.004408515058457851, + 0.004334018100053072, + 0.00564458267763257, + 0.003469036193564534, + 0.003496917663142085, + 0.008154723793268204, + 0.009629526175558567, + 0.011112491600215435, + 0.011715896427631378, + 0.008607774041593075, + 0.00797981582581997, + 0.00922571960836649, + 0.012841726653277874, + 0.00453302962705493, + 0.006865669507533312, + 0.003930300939828157, + 0.011650161817669868, + 0.0038235937245190144, + 0.005740266293287277, + 0.010045208036899567, + 0.002591273747384548, + 0.005757301114499569, + 0.007848779670894146, + 0.004847520496696234, + 0.00809870008379221, + 0.009073194116353989, + 0.006417461205273867, + 0.009192625992000103, + 0.00433382298797369, + 0.006064619868993759, + 0.010203255340456963, + 0.00910953525453806, + 0.005892932880669832, + 0.006746112369000912, + 0.008407571353018284, + 0.00521512795239687, + 0.006640858948230743, + 0.005560646299272776, + 0.004968845285475254, + 0.002923091407865286, + 0.004615257494151592, + 0.0047644320875406265, + 0.005896750371903181, + 0.007830679416656494, + 0.005353711079806089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.01389579102396965 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.09590552747249603, + 0.061735883355140686, + 0.129204660654068, + 0.3504130244255066, + 0.04982089623808861, + 0.09704941511154175, + 0.05926692113280296, + 0.049319539219141006, + 0.06951706856489182, + 0.037767067551612854 + ], + "credit_map": [ + 0.009432290680706501, + 0.011525026522576809, + 0.009069732390344143, + 0.014055644161999226, + 0.010308430530130863, + 0.015122879296541214, + 0.014322243630886078, + 0.0120692802593112, + 0.004387915134429932, + 0.008200607262551785, + 0.005972690414637327, + 0.008100955747067928, + 0.007994012907147408, + 0.007361023221164942, + 0.010829508304595947, + 0.008889173157513142, + 0.012259968556463718, + 0.016442211344838142, + 0.013163135387003422, + 0.02461116574704647, + 0.01909770257771015, + 0.01620318554341793, + 0.016167040914297104, + 0.011260257102549076, + 0.02074684388935566, + 0.012517161667346954, + 0.01432726439088583, + 0.03429423272609711, + 0.020607050508260727, + 0.14733941853046417, + 0.08449287712574005, + 0.016088159754872322, + 0.004933517891913652, + 0.010232482105493546, + 0.005183109547942877, + 0.0070589082315564156, + 0.006496399641036987, + 0.006933246273547411, + 0.00489136204123497, + 0.004091872368007898, + 0.011902681551873684, + 0.008731575682759285, + 0.012085197493433952, + 0.01939118094742298, + 0.00800755899399519, + 0.007689937949180603, + 0.015321576967835426, + 0.013919700868427753, + 0.0046846759505569935, + 0.007893189787864685, + 0.004029588308185339, + 0.014665898866951466, + 0.005636270623654127, + 0.005728423595428467, + 0.01362384483218193, + 0.003005025442689657, + 0.006317903753370047, + 0.008675415068864822, + 0.003801657585427165, + 0.006839992478489876, + 0.007622874341905117, + 0.006182370241731405, + 0.006711555644869804, + 0.0031677691731601954, + 0.0071596368215978146, + 0.009776093065738678, + 0.010051255114376545, + 0.007110397331416607, + 0.010832489468157291, + 0.010627501644194126, + 0.0064196232706308365, + 0.007540075574070215, + 0.00409027561545372, + 0.0057440632954239845, + 0.002725966740399599, + 0.004475245252251625, + 0.005943962838500738, + 0.004371302668005228, + 0.006652522832155228, + 0.0037637308705598116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.01819748803973198 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588379353284836 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588379353284836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588379353284836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09048938006162643, + 0.03388035669922829, + 0.03515218198299408, + 0.5804762244224548, + 0.02809561789035797, + 0.11449669301509857, + 0.02939538098871708, + 0.02965078130364418, + 0.03815298154950142, + 0.020210392773151398 + ], + "credit_map": [ + 0.011250816285610199, + 0.005874886643141508, + 0.006425881292670965, + 0.013174958527088165, + 0.007567848544567823, + 0.019024508073925972, + 0.013635880313813686, + 0.013534595258533955, + 0.003224317217245698, + 0.003836824791505933, + 0.003049961058422923, + 0.005395217798650265, + 0.00460951728746295, + 0.003259178251028061, + 0.005885262042284012, + 0.004620077554136515, + 0.002532082377001643, + 0.0035213232040405273, + 0.003283924190327525, + 0.007061423268169165, + 0.004167964216321707, + 0.006057459861040115, + 0.005362510681152344, + 0.0031654899939894676, + 0.03399432823061943, + 0.01554777566343546, + 0.022266535088419914, + 0.051175765693187714, + 0.03261900693178177, + 0.24569344520568848, + 0.1508444845676422, + 0.028334906324744225, + 0.0019318076083436608, + 0.002611876232549548, + 0.0024174232967197895, + 0.0035759189631789923, + 0.003484017215669155, + 0.004963192623108625, + 0.005507766269147396, + 0.003603614866733551, + 0.010209870524704456, + 0.00656086765229702, + 0.018682867288589478, + 0.020630016922950745, + 0.009340821765363216, + 0.01300606969743967, + 0.015293546952307224, + 0.02077263966202736, + 0.0028416207060217857, + 0.0031162493396550417, + 0.0033246472012251616, + 0.00392333697527647, + 0.004147875122725964, + 0.0033023254945874214, + 0.006559077650308609, + 0.002180250594392419, + 0.003689899342134595, + 0.003306453349068761, + 0.004010038450360298, + 0.004404034931212664, + 0.0033136666752398014, + 0.0037972426507622004, + 0.004590899683535099, + 0.0025385471526533365, + 0.004668415058404207, + 0.005727075506001711, + 0.004327047616243362, + 0.004228627774864435, + 0.003927754238247871, + 0.005898897536098957, + 0.003621314885094762, + 0.005753848701715469, + 0.002797321882098913, + 0.0027607260271906853, + 0.0018679609056562185, + 0.002875151112675667, + 0.0031806682236492634, + 0.0024842931888997555, + 0.0024396905209869146, + 0.001804580446332693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.020588379353284836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899527996778488 + ], + "opposite_logprob_tokens": [ + -2.5789952278137207 + ], + "logprob_reward": -0.07899527996778488, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899527996778488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2945796847343445, + 0.059818465262651443, + 0.10477767139673233, + 0.1848195344209671, + 0.07529860734939575, + 0.053855810314416885, + 0.034898798912763596, + 0.07252073287963867, + 0.0777658149600029, + 0.041664883494377136 + ], + "credit_map": [ + 0.04411569982767105, + 0.024510813876986504, + 0.03641926869750023, + 0.03185627609491348, + 0.024505844339728355, + 0.060623977333307266, + 0.035151902586221695, + 0.0373959019780159, + 0.005389463156461716, + 0.004662602674216032, + 0.004589491989463568, + 0.006186663173139095, + 0.00792540144175291, + 0.011877785436809063, + 0.009458496235311031, + 0.009728560224175453, + 0.008011842146515846, + 0.0074018919840455055, + 0.01327474694699049, + 0.01441195048391819, + 0.011253774166107178, + 0.026115858927369118, + 0.015497990883886814, + 0.008809614926576614, + 0.011377724818885326, + 0.009798546321690083, + 0.008487209677696228, + 0.019216489046812057, + 0.012258429080247879, + 0.0732218325138092, + 0.04167993739247322, + 0.008779363706707954, + 0.006988306995481253, + 0.008431492373347282, + 0.007732659112662077, + 0.006204263307154179, + 0.006610522046685219, + 0.01809651218354702, + 0.012013380415737629, + 0.009221469983458519, + 0.009529686532914639, + 0.005653799045830965, + 0.0049439407885074615, + 0.011409434489905834, + 0.004527458921074867, + 0.003991450648754835, + 0.009220069274306297, + 0.004579971544444561, + 0.0031140237115323544, + 0.004906425718218088, + 0.0048774657770991325, + 0.005955887958407402, + 0.0026864216197282076, + 0.004064873792231083, + 0.004601510241627693, + 0.004692189861088991, + 0.009419400244951248, + 0.011596989817917347, + 0.006409517955034971, + 0.00966175552457571, + 0.007366684265434742, + 0.009669564664363861, + 0.011513900011777878, + 0.0068829236552119255, + 0.008836699649691582, + 0.009748359210789204, + 0.009066350758075714, + 0.006659028586000204, + 0.015233635902404785, + 0.014081872068345547, + 0.005180697422474623, + 0.00895917322486639, + 0.005549110006541014, + 0.005549205467104912, + 0.004082673694938421, + 0.0054013109765946865, + 0.005347741302102804, + 0.005428650416433811, + 0.005103454925119877, + 0.005202735774219036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.07899527996778488 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38702428340911865 + ], + "opposite_logprob_tokens": [ + -1.1370242834091187 + ], + "logprob_reward": -0.38702428340911865, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38702428340911865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06022277846932411, + 0.04591261222958565, + 0.06258714199066162, + 0.5761430263519287, + 0.04920927435159683, + 0.07258068770170212, + 0.049331262707710266, + 0.02310151979327202, + 0.033871252089738846, + 0.027040448039770126 + ], + "credit_map": [ + 0.005678100045770407, + 0.006064784713089466, + 0.005801745690405369, + 0.012965504080057144, + 0.005310225300490856, + 0.007227948401123285, + 0.009049727581441402, + 0.00812474638223648, + 0.003003378165885806, + 0.005170551594346762, + 0.0038471841253340244, + 0.008815042674541473, + 0.0051436168141663074, + 0.004367250949144363, + 0.007825363427400589, + 0.007740224711596966, + 0.00527073722332716, + 0.006116084288805723, + 0.004778056871145964, + 0.008543651551008224, + 0.007497129961848259, + 0.010631234385073185, + 0.011378624476492405, + 0.008371624164283276, + 0.034170836210250854, + 0.015182522125542164, + 0.021696709096431732, + 0.0503201000392437, + 0.031919002532958984, + 0.2417016178369522, + 0.15287363529205322, + 0.028278615325689316, + 0.004252128768712282, + 0.009934077039361, + 0.003701276844367385, + 0.007126753218472004, + 0.005580107215791941, + 0.008569888770580292, + 0.006210254039615393, + 0.003834789153188467, + 0.00913883838802576, + 0.005778385326266289, + 0.008983006700873375, + 0.01203394029289484, + 0.006293005309998989, + 0.007458684500306845, + 0.012382317334413528, + 0.010512510314583778, + 0.004938007332384586, + 0.006912549026310444, + 0.005468389019370079, + 0.007592411711812019, + 0.0052748993039131165, + 0.006931967101991177, + 0.008686121553182602, + 0.0035269169602543116, + 0.0027804779820144176, + 0.0033287229016423225, + 0.003260105848312378, + 0.003937285393476486, + 0.0024662360083311796, + 0.002628665417432785, + 0.0029284905176609755, + 0.0017715365393087268, + 0.0031160779763013124, + 0.004878266714513302, + 0.004279520828276873, + 0.003110468853265047, + 0.006172063760459423, + 0.005744765512645245, + 0.0035829604603350163, + 0.002987127983942628, + 0.003367443336173892, + 0.0033372724428772926, + 0.0025592732708901167, + 0.0028565742541104555, + 0.003164770547300577, + 0.003969904035329819, + 0.0044046975672245026, + 0.0033805109560489655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.38702428340911865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a hyperrealistic documentary style, the camera moves in a slow, smooth orbit, rotating 360 degrees around an ornate sedan chair as it navigates a rugged mountain path. It's an intricately carved vermilion lacquered wooden sedan, its sides intricately carved with golden dragon and phoenix reliefs. Its roof is covered in bright yellow glazed tiles, and its corners are draped with silk tassels that sway gently as the sedan rolls. Four bearers, dressed in simple dark blue coarse cloth jackets, steadily carry the heavy poles with muscular shoulders. Their weathered faces express concentration, and each step is remarkably steady and powerful, creating a stark contrast to the sedan's rhythmic swaying caused by the uneven road. The afternoon sun penetrated the gaps in the dense and intertwined treetops overhead, forming sharp beams of light, casting ever-shifting and ever-changing dappled light and shadows on the smooth red paint surface of the sedan chair. The light and shadows flowed as the team moved forward, highlighting the texture of the wood and the depth of the carvings, creating an atmosphere of solemnity, primitiveness and dynamic beauty.", + "seed": 23436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23436_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": -0.09925602972507477, + "vlm_logprob": -0.09925602972507477, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.09925602972507477, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0218119528144598 + ], + "opposite_logprob_tokens": [ + -4.0218119621276855 + ], + "logprob_reward": -0.0218119528144598, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0218119528144598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11430191248655319, + 0.10012991726398468, + 0.03497914969921112, + 0.0672832578420639, + 0.053603604435920715, + 0.11839131265878677, + 0.1544845998287201, + 0.13282747566699982, + 0.11627621203660965, + 0.10772256553173065 + ], + "credit_map": [ + 0.01973583549261093, + 0.014100657775998116, + 0.014442573301494122, + 0.013414483517408371, + 0.010324030183255672, + 0.010126931592822075, + 0.016583111137151718, + 0.015574291348457336, + 0.011751946061849594, + 0.009441503323614597, + 0.014308596029877663, + 0.018754923716187477, + 0.009615919552743435, + 0.008832077495753765, + 0.015976810827851295, + 0.011448131874203682, + 0.0038712783716619015, + 0.003611601423472166, + 0.004559284076094627, + 0.009764430113136768, + 0.00235080998390913, + 0.0028655887581408024, + 0.0048193419352173805, + 0.0031368141062557697, + 0.005722458474338055, + 0.013163596391677856, + 0.011079746298491955, + 0.008028749376535416, + 0.007461316883563995, + 0.00650927796959877, + 0.00860734935849905, + 0.0067107598297297955, + 0.008261163718998432, + 0.003660862799733877, + 0.013683951459825039, + 0.007025650702416897, + 0.003719378262758255, + 0.003978678025305271, + 0.008699819445610046, + 0.004574100486934185, + 0.013715834356844425, + 0.01712256856262684, + 0.013870050199329853, + 0.014327848330140114, + 0.020326929166913033, + 0.01364070363342762, + 0.012823548167943954, + 0.012563844211399555, + 0.012314697727560997, + 0.009725701995193958, + 0.009708738885819912, + 0.009589619934558868, + 0.009540262632071972, + 0.029711401090025902, + 0.02104921080172062, + 0.052844975143671036, + 0.015214890241622925, + 0.015844320878386497, + 0.01217014342546463, + 0.015160550363361835, + 0.015175625681877136, + 0.023220505565404892, + 0.012696943245828152, + 0.02334449253976345, + 0.008743996731936932, + 0.015109062194824219, + 0.01249114703387022, + 0.01688241958618164, + 0.011705326847732067, + 0.017149487510323524, + 0.019450318068265915, + 0.014744458720088005, + 0.010617896914482117, + 0.01318164449185133, + 0.009839720092713833, + 0.014267544262111187, + 0.016441751271486282, + 0.016687888652086258, + 0.01812112331390381, + 0.00856500118970871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0218119528144598 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016084907576441765 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.016084907576441765, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016084907576441765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.070302814245224, + 0.04948094114661217, + 0.023646144196391106, + 0.07328937202692032, + 0.11618933081626892, + 0.0308759193867445, + 0.18856477737426758, + 0.1292278915643692, + 0.2174505740404129, + 0.10097219794988632 + ], + "credit_map": [ + 0.013819030486047268, + 0.007576745469123125, + 0.010374556295573711, + 0.0075338371098041534, + 0.007734636776149273, + 0.005261378828436136, + 0.011717264540493488, + 0.006285369861871004, + 0.006873063277453184, + 0.005801025312393904, + 0.007838871330022812, + 0.007832092233002186, + 0.007063358090817928, + 0.0034721149131655693, + 0.006189925596117973, + 0.0044104959815740585, + 0.0036596995778381824, + 0.0031869520898908377, + 0.003515414195135236, + 0.0030891227070242167, + 0.002475632121786475, + 0.002931061666458845, + 0.0025750668719410896, + 0.0022131972946226597, + 0.0074151670560240746, + 0.02487638033926487, + 0.007605967577546835, + 0.006834933999925852, + 0.004817549139261246, + 0.00422633858397603, + 0.010288668796420097, + 0.0072243716567754745, + 0.012740801088511944, + 0.00946834310889244, + 0.02021016739308834, + 0.019634881988167763, + 0.012632296420633793, + 0.011871880851686, + 0.02113003470003605, + 0.008500941097736359, + 0.0029077688232064247, + 0.004036104306578636, + 0.0039808559231460094, + 0.003944735042750835, + 0.0033823770936578512, + 0.004014129284769297, + 0.003729794640094042, + 0.004880155436694622, + 0.01261365506798029, + 0.008070641197264194, + 0.009366431273519993, + 0.00845668837428093, + 0.012940507382154465, + 0.040705252438783646, + 0.025933122262358665, + 0.07047849893569946, + 0.01654973439872265, + 0.01300749834626913, + 0.015336847864091396, + 0.01455896906554699, + 0.010392806492745876, + 0.02654849737882614, + 0.012143940664827824, + 0.02068961225450039, + 0.017612555995583534, + 0.02204638347029686, + 0.01940167509019375, + 0.027639169245958328, + 0.02256791852414608, + 0.03355150297284126, + 0.04493700712919235, + 0.02969438210129738, + 0.007457252591848373, + 0.007794140372425318, + 0.008798215538263321, + 0.013446652330458164, + 0.010794567875564098, + 0.016596833243966103, + 0.01976151205599308, + 0.016323037445545197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.016084907576441765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635536715388298 + ], + "opposite_logprob_tokens": [ + -3.651355266571045 + ], + "logprob_reward": -0.02635536715388298, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635536715388298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11203759908676147, + 0.10347319394350052, + 0.05523889511823654, + 0.07451318204402924, + 0.06501723825931549, + 0.035581137984991074, + 0.1319328397512436, + 0.15020513534545898, + 0.11745574325323105, + 0.1545451134443283 + ], + "credit_map": [ + 0.019456971436738968, + 0.015359388664364815, + 0.015256506390869617, + 0.016745932400226593, + 0.00995570421218872, + 0.007812798023223877, + 0.017147427424788475, + 0.01030285470187664, + 0.014477076940238476, + 0.011931288987398148, + 0.01604953594505787, + 0.016004975885152817, + 0.008201994001865387, + 0.008873378857970238, + 0.015387588180601597, + 0.012547334656119347, + 0.014638434164226055, + 0.005437024403363466, + 0.00887257605791092, + 0.01015094481408596, + 0.003923120442777872, + 0.004630211740732193, + 0.004018277861177921, + 0.003568302607163787, + 0.007510458584874868, + 0.02371140383183956, + 0.008516984060406685, + 0.008360018953680992, + 0.006533957086503506, + 0.005433922633528709, + 0.010366329923272133, + 0.004080096259713173, + 0.008915677666664124, + 0.004192796535789967, + 0.014792595990002155, + 0.009382310323417187, + 0.004317084327340126, + 0.007286103907972574, + 0.010967621579766273, + 0.005163042806088924, + 0.0038509920705109835, + 0.0048542083241045475, + 0.00510428985580802, + 0.003933052998036146, + 0.005544296000152826, + 0.004214623477309942, + 0.005250518675893545, + 0.0028291575144976377, + 0.012887552380561829, + 0.01325209066271782, + 0.008062821812927723, + 0.007985074073076248, + 0.014359772205352783, + 0.0230787955224514, + 0.016580332070589066, + 0.035726405680179596, + 0.016710776835680008, + 0.017544014379382133, + 0.01920776627957821, + 0.02864217571914196, + 0.015050975605845451, + 0.01950877532362938, + 0.015860777348279953, + 0.017679866403341293, + 0.009803282096982002, + 0.012412059120833874, + 0.013155123218894005, + 0.015184530057013035, + 0.010844402946531773, + 0.020132074132561684, + 0.015303174965083599, + 0.02062108926475048, + 0.013230863027274609, + 0.01621267758309841, + 0.015814924612641335, + 0.01951446756720543, + 0.020085791125893593, + 0.023130321875214577, + 0.030718272551894188, + 0.015837788581848145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.02635536715388298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521568834781647 + ], + "opposite_logprob_tokens": [ + -1.5021568536758423 + ], + "logprob_reward": -0.2521568834781647, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521568834781647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19405032694339752, + 0.0576474629342556, + 0.03118976764380932, + 0.041766662150621414, + 0.026462271809577942, + 0.09096870571374893, + 0.1304267942905426, + 0.23771539330482483, + 0.11779079586267471, + 0.07198182493448257 + ], + "credit_map": [ + 0.039574865251779556, + 0.035924509167671204, + 0.027952013537287712, + 0.0273983646184206, + 0.01164648775011301, + 0.009292159229516983, + 0.022584963589906693, + 0.019676964730024338, + 0.007130136713385582, + 0.005161081440746784, + 0.006948002148419619, + 0.010488077066838741, + 0.00800329353660345, + 0.005014078225940466, + 0.007947909645736217, + 0.006954890675842762, + 0.00472227344289422, + 0.004297731444239616, + 0.006483742501586676, + 0.00443421071395278, + 0.0026207217015326023, + 0.002061773557215929, + 0.004275966435670853, + 0.0022933476138859987, + 0.0038587339222431183, + 0.014747388660907745, + 0.005445125978440046, + 0.004900231491774321, + 0.0032303135376423597, + 0.002746397629380226, + 0.004995846655219793, + 0.0018426260212436318, + 0.0031885092612355947, + 0.0024606981314718723, + 0.003336969530209899, + 0.0023125563748180866, + 0.0028063750360161066, + 0.005991226062178612, + 0.0029284770134836435, + 0.0034374601673334837, + 0.010060441680252552, + 0.013527910225093365, + 0.009805876761674881, + 0.00963528361171484, + 0.016543090343475342, + 0.010898614302277565, + 0.01161725539714098, + 0.008880234323441982, + 0.008875948376953602, + 0.008710244670510292, + 0.007060196716338396, + 0.007798491511493921, + 0.010662119835615158, + 0.02605121023952961, + 0.018818557262420654, + 0.04245002567768097, + 0.025834159925580025, + 0.026220442727208138, + 0.03400121629238129, + 0.05291873961687088, + 0.016034817323088646, + 0.027316292747855186, + 0.025155285373330116, + 0.030234429985284805, + 0.009245536290109158, + 0.010778466239571571, + 0.009473169222474098, + 0.012719841673970222, + 0.009359372779726982, + 0.01922179013490677, + 0.026708997786045074, + 0.020283609628677368, + 0.006324403453618288, + 0.005106369033455849, + 0.008935447782278061, + 0.013580098748207092, + 0.0060597676783800125, + 0.011667132377624512, + 0.009281643666327, + 0.011026964522898197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.2521568834781647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17987103760242462 + ], + "opposite_logprob_tokens": [ + -1.8048710823059082 + ], + "logprob_reward": -0.17987103760242462, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17987103760242462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054671015590429306, + 0.04105082526803017, + 0.0532563216984272, + 0.08319064229726791, + 0.10018343478441238, + 0.052859727293252945, + 0.22853998839855194, + 0.12046311050653458, + 0.1541036069393158, + 0.11168135702610016 + ], + "credit_map": [ + 0.008289546705782413, + 0.007544659078121185, + 0.0062425327487289906, + 0.006260639987885952, + 0.005434836260974407, + 0.005430311895906925, + 0.008885654620826244, + 0.0065828305669128895, + 0.005561901722103357, + 0.003990429453551769, + 0.005559837911278009, + 0.006999120581895113, + 0.0038079393561929464, + 0.0042448146268725395, + 0.006937015801668167, + 0.003949764650315046, + 0.00972579326480627, + 0.0069640325382351875, + 0.008928367868065834, + 0.007740025874227285, + 0.0043988656252622604, + 0.0036015817895531654, + 0.008434991352260113, + 0.0034626610577106476, + 0.006743573118001223, + 0.01016355399042368, + 0.010486959479749203, + 0.011236710473895073, + 0.007965156808495522, + 0.005760179832577705, + 0.01662452705204487, + 0.014209980145096779, + 0.012237193062901497, + 0.005754264071583748, + 0.02022097259759903, + 0.019808316603302956, + 0.0074457633309066296, + 0.007447618991136551, + 0.019715068861842155, + 0.007554234936833382, + 0.0053951130248606205, + 0.007757913786917925, + 0.005832429975271225, + 0.006708440836519003, + 0.008140255697071552, + 0.006508737802505493, + 0.007311901077628136, + 0.005204934626817703, + 0.01624033972620964, + 0.011999152600765228, + 0.012643415480852127, + 0.012273446656763554, + 0.015990393236279488, + 0.047261420637369156, + 0.030894314870238304, + 0.08123748749494553, + 0.017013894394040108, + 0.01541426032781601, + 0.016456015408039093, + 0.016739346086978912, + 0.011440977454185486, + 0.015303408727049828, + 0.013899575918912888, + 0.014195624738931656, + 0.009794582612812519, + 0.01014683023095131, + 0.014676753431558609, + 0.01331416517496109, + 0.012510471977293491, + 0.026330389082431793, + 0.04289485886693001, + 0.02443554438650608, + 0.010003400966525078, + 0.015138102695345879, + 0.013196348212659359, + 0.01651855930685997, + 0.01334358099848032, + 0.016408076509833336, + 0.014362337999045849, + 0.01271094661206007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.17987103760242462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot in a hyperrealistic documentary style, the camera moves in a slow, smooth orbit, rotating 360 degrees around an ornate sedan chair as it navigates a rugged mountain path. It's an intricately carved vermilion lacquered wooden sedan, its sides intricately carved with golden dragon and phoenix reliefs. Its roof is covered in bright yellow glazed tiles, and its corners are draped with silk tassels that sway gently as the sedan rolls. Four bearers, dressed in simple dark blue coarse cloth jackets, steadily carry the heavy poles with muscular shoulders. Their weathered faces express concentration, and each step is remarkably steady and powerful, creating a stark contrast to the sedan's rhythmic swaying caused by the uneven road. The afternoon sun penetrated the gaps in the dense and intertwined treetops overhead, forming sharp beams of light, casting ever-shifting and ever-changing dappled light and shadows on the smooth red paint surface of the sedan chair. The light and shadows flowed as the team moved forward, highlighting the texture of the wood and the depth of the carvings, creating an atmosphere of solemnity, primitiveness and dynamic beauty.", + "seed": 23437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_3/index_1684_seed_23437_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": -0.1149118509143591, + "vlm_logprob": -0.1149118509143591, + "vlm_yes_no_margin": 2.924999988079071, + "vlm_token_credit": -0.1149118509143591, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019259357824921608 + ], + "opposite_logprob_tokens": [ + -4.144259452819824 + ], + "logprob_reward": -0.019259357824921608, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019259357824921608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6759592294692993, + 0.027408825233578682, + 0.069175124168396, + 0.017957963049411774, + 0.03300418704748154, + 0.02222956344485283, + 0.031791120767593384, + 0.05082085356116295, + 0.031736791133880615, + 0.03991633653640747 + ], + "credit_map": [ + 0.04132336378097534, + 0.058930035680532455, + 0.047155749052762985, + 0.17803248763084412, + 0.028082432225346565, + 0.09626567363739014, + 0.045211952179670334, + 0.18095757067203522, + 0.0033504562452435493, + 0.0030131794046610594, + 0.004290511831641197, + 0.0024406122975051403, + 0.0024676979519426823, + 0.003650347702205181, + 0.004206668585538864, + 0.003989353310316801, + 0.005889744497835636, + 0.007732085883617401, + 0.013452649116516113, + 0.008738155476748943, + 0.005566827952861786, + 0.007680879440158606, + 0.006592324003577232, + 0.01352245919406414, + 0.0021041594445705414, + 0.001626132638193667, + 0.003142726607620716, + 0.003027789993211627, + 0.002194172004237771, + 0.0019651693291962147, + 0.0021389597095549107, + 0.0017588521586731076, + 0.003013043897226453, + 0.00264094490557909, + 0.0028420512098819017, + 0.004224023316055536, + 0.003930413629859686, + 0.005023809615522623, + 0.007849412970244884, + 0.0034804875031113625, + 0.003526629414409399, + 0.0024400262627750635, + 0.0029524979181587696, + 0.0029175900854170322, + 0.0017649262445047498, + 0.0025392910465598106, + 0.0029220706783235073, + 0.003166532376781106, + 0.0035336969885975122, + 0.00310496985912323, + 0.004213681444525719, + 0.004748618230223656, + 0.004257747437804937, + 0.0031007924117147923, + 0.004933497402817011, + 0.003898115362972021, + 0.004859468899667263, + 0.006007499992847443, + 0.005233495030552149, + 0.009369002655148506, + 0.00519502954557538, + 0.006776699796319008, + 0.008606011047959328, + 0.004773647990077734, + 0.0037946493830531836, + 0.00397630175575614, + 0.0038260009605437517, + 0.005539023317396641, + 0.0032913261093199253, + 0.0030020601116120815, + 0.005113163962960243, + 0.003194265766069293, + 0.0034237587824463844, + 0.0049056694842875, + 0.004825674928724766, + 0.006646537221968174, + 0.004734768066555262, + 0.0057264347560703754, + 0.005193373188376427, + 0.0044601187109947205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.019259357824921608 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020589197054505348 + ], + "opposite_logprob_tokens": [ + -3.8955891132354736 + ], + "logprob_reward": -0.020589197054505348, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020589197054505348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43258118629455566, + 0.033699508756399155, + 0.047598548233509064, + 0.015682239085435867, + 0.05272626876831055, + 0.0442998968064785, + 0.07966633886098862, + 0.054344117641448975, + 0.1342938244342804, + 0.10510808974504471 + ], + "credit_map": [ + 0.029582522809505463, + 0.043296929448843, + 0.03804831579327583, + 0.13865871727466583, + 0.020468859001994133, + 0.0409744530916214, + 0.03148260340094566, + 0.09006878733634949, + 0.00337498658336699, + 0.003027684520930052, + 0.004558575339615345, + 0.004936505109071732, + 0.003063310869038105, + 0.0030093572568148375, + 0.006900963839143515, + 0.004828125704079866, + 0.0045106783509254456, + 0.003989690449088812, + 0.00811601523309946, + 0.005581418517976999, + 0.005224204156547785, + 0.007019716314971447, + 0.005910852458328009, + 0.007245970889925957, + 0.002296130172908306, + 0.001550559070892632, + 0.0024313523899763823, + 0.002866529393941164, + 0.0016528192209079862, + 0.001432460150681436, + 0.0019471304258331656, + 0.0015052588423714042, + 0.007466252893209457, + 0.004384177271276712, + 0.0066916137002408504, + 0.0067067742347717285, + 0.007751480210572481, + 0.007652672473341227, + 0.006112302653491497, + 0.005960999988019466, + 0.006508080754429102, + 0.004532989580184221, + 0.005344226490706205, + 0.00594303198158741, + 0.004601188004016876, + 0.006963920313864946, + 0.0049400292336940765, + 0.005466432776302099, + 0.011167231947183609, + 0.007697931956499815, + 0.007190707139670849, + 0.008115275762975216, + 0.01476750336587429, + 0.012024062685668468, + 0.009413698688149452, + 0.009289931505918503, + 0.0037036717403680086, + 0.006763216573745012, + 0.009615255519747734, + 0.01009826548397541, + 0.00592941464856267, + 0.006246862001717091, + 0.007251064293086529, + 0.004736365284770727, + 0.011980581097304821, + 0.015708772465586662, + 0.018105745315551758, + 0.0160641148686409, + 0.018284140154719353, + 0.021545061841607094, + 0.01905759610235691, + 0.013547803275287151, + 0.009600954130291939, + 0.007487171329557896, + 0.014296483248472214, + 0.009712142869830132, + 0.01321638748049736, + 0.015525457449257374, + 0.01157634798437357, + 0.0236931461840868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.020589197054505348 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588496699929237 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588496699929237, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588496699929237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36958110332489014, + 0.21729160845279694, + 0.0684482529759407, + 0.027407778427004814, + 0.07951919734477997, + 0.0940161794424057, + 0.03899182751774788, + 0.020538073033094406, + 0.035608645528554916, + 0.04859732836484909 + ], + "credit_map": [ + 0.02244960516691208, + 0.02044757455587387, + 0.025335265323519707, + 0.040242888033390045, + 0.01489083282649517, + 0.11581630259752274, + 0.02111748233437538, + 0.10928118973970413, + 0.02401401288807392, + 0.018793169409036636, + 0.0346684530377388, + 0.03214742988348007, + 0.019483031705021858, + 0.03718245029449463, + 0.020462781190872192, + 0.03054027445614338, + 0.006098227575421333, + 0.007746761664748192, + 0.01307679433375597, + 0.010536018759012222, + 0.006982322782278061, + 0.00804835557937622, + 0.009353769943118095, + 0.006605999544262886, + 0.003340143011882901, + 0.0027514600660651922, + 0.004010388161987066, + 0.003972676116973162, + 0.002651277231052518, + 0.0035536771174520254, + 0.004094194155186415, + 0.0030339614022523165, + 0.00966315995901823, + 0.007644466124475002, + 0.0074810441583395, + 0.01005048118531704, + 0.009977911598980427, + 0.009997492656111717, + 0.0170753113925457, + 0.007629324682056904, + 0.012204498052597046, + 0.008649674244225025, + 0.009864050894975662, + 0.012975580058991909, + 0.008462898433208466, + 0.014186489395797253, + 0.015676705166697502, + 0.011996287852525711, + 0.003697894513607025, + 0.003967396914958954, + 0.0032894432079046965, + 0.005887148901820183, + 0.004931959789246321, + 0.005062737036496401, + 0.005918186157941818, + 0.006237061694264412, + 0.002942511346191168, + 0.0022821153979748487, + 0.0025083974469453096, + 0.002213462023064494, + 0.0019068854162469506, + 0.0029602295253425837, + 0.002980856690555811, + 0.0027436153031885624, + 0.003755370620638132, + 0.0029205044265836477, + 0.0046465033665299416, + 0.005504181142896414, + 0.004321071319282055, + 0.005094876512885094, + 0.004749643616378307, + 0.004616494756191969, + 0.006263168528676033, + 0.0050302911549806595, + 0.005626475438475609, + 0.00761825405061245, + 0.0057549141347408295, + 0.0057321698404848576, + 0.006165679078549147, + 0.006406375207006931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.020588496699929237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708701193332672 + ], + "opposite_logprob_tokens": [ + -2.127087116241455 + ], + "logprob_reward": -0.12708701193332672, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708701193332672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.779632031917572, + 0.04788818210363388, + 0.017169512808322906, + 0.018116986379027367, + 0.015045939944684505, + 0.03504513204097748, + 0.027734732255339622, + 0.010069508105516434, + 0.0241746436804533, + 0.025123320519924164 + ], + "credit_map": [ + 0.04472021013498306, + 0.06156434118747711, + 0.05172361433506012, + 0.18125231564044952, + 0.028588715940713882, + 0.14223648607730865, + 0.047960877418518066, + 0.2215854823589325, + 0.00620572455227375, + 0.004380970261991024, + 0.007562672253698111, + 0.0073940386064350605, + 0.0038542908150702715, + 0.00690626073628664, + 0.004902299493551254, + 0.006681922357529402, + 0.001980069326236844, + 0.0027439494151622057, + 0.0023449049331247807, + 0.002344986656680703, + 0.0024383533746004105, + 0.002148258965462446, + 0.0015592722920700908, + 0.0016097172629088163, + 0.002231582533568144, + 0.0012647425755858421, + 0.0028117659967392683, + 0.0037463419139385223, + 0.002237541601061821, + 0.0016204452840611339, + 0.002338662277907133, + 0.0018659057095646858, + 0.0021549018565565348, + 0.0016976840561255813, + 0.001699357875622809, + 0.0011821255320683122, + 0.001971632242202759, + 0.0030840677209198475, + 0.001460463390685618, + 0.0017957076197490096, + 0.004955862648785114, + 0.003318530274555087, + 0.0033706717658787966, + 0.003589654341340065, + 0.0035418840125203133, + 0.006500918418169022, + 0.0054779695346951485, + 0.004289643373340368, + 0.0026376324240118265, + 0.0032670702785253525, + 0.0025653671473264694, + 0.0038108383305370808, + 0.0035154817160218954, + 0.004480772651731968, + 0.003619208000600338, + 0.003838363103568554, + 0.00104046321939677, + 0.001076280022971332, + 0.0010423357598483562, + 0.001148556126281619, + 0.0015263785608112812, + 0.0015902576269581914, + 0.0015637798933312297, + 0.001081456313841045, + 0.0029672100208699703, + 0.0026213342789560556, + 0.0025583170354366302, + 0.00479913642629981, + 0.0027019972912967205, + 0.001995127648115158, + 0.0042262147180736065, + 0.0023053069598972797, + 0.0034780981950461864, + 0.002539318986237049, + 0.003408523742109537, + 0.0034715307410806417, + 0.001984266797080636, + 0.0021232024300843477, + 0.002486162120476365, + 0.005632217042148113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.12708701193332672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- 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 ornate sedan chair with vermilion lacquered wood and golden dragon and phoenix carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sedan chair's roof covered in bright yellow glazed tiles with silk tassels at the corners?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are four bearers wearing dark blue coarse cloth jackets carrying the sedan chair on a rugged mountain path?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature afternoon sunlight filtering through dense treetops to create dappled light and shadows on the sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move in a slow, smooth orbit rotating 360 degrees around the sedan chair?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38703519105911255 + ], + "opposite_logprob_tokens": [ + -1.1370351314544678 + ], + "logprob_reward": -0.38703519105911255, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38703519105911255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2994716167449951, + 0.12639454007148743, + 0.07049398124217987, + 0.037140410393476486, + 0.07534041255712509, + 0.10495560616254807, + 0.08573535084724426, + 0.03300752490758896, + 0.09073584526777267, + 0.07672473043203354 + ], + "credit_map": [ + 0.024557651951909065, + 0.01583344303071499, + 0.025706008076667786, + 0.03280403092503548, + 0.01156076043844223, + 0.08168280124664307, + 0.014105334877967834, + 0.09322160482406616, + 0.014386683702468872, + 0.012193663977086544, + 0.023096377030014992, + 0.017661120742559433, + 0.009637941606342793, + 0.01733521744608879, + 0.011292480863630772, + 0.02079104818403721, + 0.005741093773394823, + 0.008903875946998596, + 0.015312375500798225, + 0.008935383521020412, + 0.0058881863951683044, + 0.005924344528466463, + 0.007250902708619833, + 0.012537823989987373, + 0.0036013489589095116, + 0.0032745152711868286, + 0.004521874710917473, + 0.004704302176833153, + 0.00476088048890233, + 0.00598528515547514, + 0.005199084058403969, + 0.0050931209698319435, + 0.011448152363300323, + 0.005605367012321949, + 0.010181444697082043, + 0.00851366762071848, + 0.00847766175866127, + 0.013693040236830711, + 0.009128502570092678, + 0.008292573504149914, + 0.015676291659474373, + 0.012577177956700325, + 0.012082688510417938, + 0.011536889709532261, + 0.010457228869199753, + 0.014741846360266209, + 0.014196204952895641, + 0.013687274418771267, + 0.008243691176176071, + 0.010618211701512337, + 0.009367682039737701, + 0.015641607344150543, + 0.006850750185549259, + 0.008099082857370377, + 0.015722736716270447, + 0.011191588826477528, + 0.0031410648953169584, + 0.004209926351904869, + 0.0035232622176408768, + 0.0069845388643443584, + 0.0035045617260038853, + 0.003819284262135625, + 0.004616145510226488, + 0.0032087438739836216, + 0.009608197025954723, + 0.01137495506554842, + 0.01233724970370531, + 0.015682648867368698, + 0.009435567073523998, + 0.008223769254982471, + 0.012417113408446312, + 0.011656341142952442, + 0.008625810965895653, + 0.00830021407455206, + 0.00879979319870472, + 0.009554644115269184, + 0.007187293842434883, + 0.009899977594614029, + 0.011908682063221931, + 0.01244831457734108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.38703519105911255 + } + ], + "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/0000160/group_4/index_tensor([3213, 3213])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_tensor([3213, 3213])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..47099d9083c36394db425e03ad05251750372778 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_tensor([3213, 3213])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "As the camera begins a slow, dramatic push-in, a medium shot from a slightly high angle unfolds across a realistic globe suspended in a desolate black void. The globe's surface features a semi-gloss finish, with deep sapphire oceans and continents detailed in realistic topographic hues of forest green, sandy yellow, and earthy brown, accented with delicate latitude and vertical lines. A single spotlight shining from above casts a bright highlight on the northern hemisphere, creating sharp shadows that emphasize its perfectly spherical shape. The transformation subtly begins as the solid textures of the continents begin to shimmer and lose their contours, liquefying as if made of hot wax. Distinctive land masses melt, their vibrant colors melting and flowing into the ocean in viscous, shimmering streams. As the camera steadily approaches, the scene focuses on a mesmerizing chaos: greens, browns, and yellows blend into deep blue, forming vibrant, swirling pools of mixed colors that create a hypnotic marble pattern across the globe's surface, transforming the familiar map into a fluid, abstract painting of beautiful decay.", + "seed": 1310, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_3213_seed_1310_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": -0.06107242973521352, + "vlm_logprob": -0.06107242973521352, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.06107242973521352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.037833333015441895 + ], + "opposite_logprob_tokens": [ + -3.4128332138061523 + ], + "logprob_reward": -0.037833333015441895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.037833333015441895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07903964072465897, + 0.05725580453872681, + 0.14738762378692627, + 0.42978399991989136, + 0.03534255549311638, + 0.044013142585754395, + 0.035069506615400314, + 0.059965282678604126, + 0.014549287967383862, + 0.09759315848350525 + ], + "credit_map": [ + 0.008262296207249165, + 0.007536308839917183, + 0.013988547958433628, + 0.004761029500514269, + 0.008981692604720592, + 0.015475754626095295, + 0.00973521638661623, + 0.010298794135451317, + 0.004159949719905853, + 0.00399575661867857, + 0.009106316603720188, + 0.0024380413815379143, + 0.004260745830833912, + 0.012833795510232449, + 0.0052779060788452625, + 0.015183288604021072, + 0.00783857237547636, + 0.00846683606505394, + 0.02304454892873764, + 0.0047828941605985165, + 0.014870279468595982, + 0.0456886813044548, + 0.008035879582166672, + 0.03465993329882622, + 0.01053047925233841, + 0.0161972064524889, + 0.03342248126864433, + 0.011505953036248684, + 0.012688636779785156, + 0.1641455739736557, + 0.015189532190561295, + 0.16610412299633026, + 0.0034748748876154423, + 0.003879322437569499, + 0.00481744110584259, + 0.004280140157788992, + 0.005110437981784344, + 0.005015882663428783, + 0.0034624277614057064, + 0.005302027799189091, + 0.004041770938783884, + 0.003521386068314314, + 0.0030910836067050695, + 0.0031460504978895187, + 0.00798732228577137, + 0.009193281643092632, + 0.004420959856361151, + 0.008611286990344524, + 0.0022474611178040504, + 0.003071926534175873, + 0.005262441001832485, + 0.00312304450199008, + 0.0037437956780195236, + 0.007661290932446718, + 0.004047309048473835, + 0.005912236403673887, + 0.0056698741391301155, + 0.003387513803318143, + 0.009811942465603352, + 0.0049986946396529675, + 0.0046617030166089535, + 0.013495078310370445, + 0.0051165116019546986, + 0.012823968194425106, + 0.0015468671917915344, + 0.0013666872400790453, + 0.0016860353061929345, + 0.002333252225071192, + 0.002159818774089217, + 0.0016913124127313495, + 0.0015803567366674542, + 0.002184959128499031, + 0.010013531893491745, + 0.011354528367519379, + 0.00895124115049839, + 0.019508181139826775, + 0.009327935054898262, + 0.012000512331724167, + 0.012521274387836456, + 0.013915956020355225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.037833333015441895 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023336289450526237 + ], + "opposite_logprob_tokens": [ + -3.773336172103882 + ], + "logprob_reward": -0.023336289450526237, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023336289450526237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09005362540483475, + 0.05220099166035652, + 0.03763440251350403, + 0.26250216364860535, + 0.13821206986904144, + 0.12246543914079666, + 0.16147005558013916, + 0.04618147388100624, + 0.047962456941604614, + 0.04131735488772392 + ], + "credit_map": [ + 0.008149960078299046, + 0.006145837716758251, + 0.01229209266602993, + 0.005286196246743202, + 0.01142800785601139, + 0.018472937867045403, + 0.00948294810950756, + 0.018795648589730263, + 0.004832737147808075, + 0.00497491704300046, + 0.00882389210164547, + 0.004180822055786848, + 0.00475891912356019, + 0.011996721848845482, + 0.005195511505007744, + 0.007437468972057104, + 0.004501545801758766, + 0.0035886524710804224, + 0.005381890572607517, + 0.0026318340096622705, + 0.003245098516345024, + 0.006933120079338551, + 0.0036552781239151955, + 0.007696980610489845, + 0.021042153239250183, + 0.0235889945179224, + 0.030570492148399353, + 0.02283172495663166, + 0.02588551864027977, + 0.057440850883722305, + 0.02196640707552433, + 0.0591760091483593, + 0.012476368807256222, + 0.011428148485720158, + 0.01805741712450981, + 0.010615297593176365, + 0.02332131564617157, + 0.025671059265732765, + 0.014146772213280201, + 0.022495681419968605, + 0.010893629863858223, + 0.014321881346404552, + 0.010463444516062737, + 0.008474143221974373, + 0.024173012003302574, + 0.018949126824736595, + 0.013452917337417603, + 0.02173728123307228, + 0.014352917671203613, + 0.0224860068410635, + 0.026524342596530914, + 0.017726179212331772, + 0.02100542187690735, + 0.01867828331887722, + 0.014369049109518528, + 0.02632785402238369, + 0.005869615823030472, + 0.003935142420232296, + 0.009324394166469574, + 0.004963033366948366, + 0.006038957741111517, + 0.005607747472822666, + 0.005315214395523071, + 0.005127366632223129, + 0.006221983581781387, + 0.0049919565208256245, + 0.006572154816240072, + 0.00729125551879406, + 0.007219314109534025, + 0.004784409888088703, + 0.005109823774546385, + 0.00577155826613307, + 0.004031333141028881, + 0.0038082378450781107, + 0.004439676646143198, + 0.006766502745449543, + 0.004668745677918196, + 0.005404147319495678, + 0.006163254380226135, + 0.006035455968230963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.023336289450526237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609722524881363 + ], + "opposite_logprob_tokens": [ + -4.141097068786621 + ], + "logprob_reward": -0.01609722524881363, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609722524881363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0841270163655281, + 0.06953954696655273, + 0.03221507743000984, + 0.556333065032959, + 0.061993345618247986, + 0.07819287478923798, + 0.031454410403966904, + 0.04610930755734444, + 0.013445074670016766, + 0.02659033052623272 + ], + "credit_map": [ + 0.012659519910812378, + 0.008035518229007721, + 0.012682264670729637, + 0.004260125104337931, + 0.008124323561787605, + 0.016966763883829117, + 0.008639483712613583, + 0.012759010307490826, + 0.007202924694865942, + 0.005320827476680279, + 0.01432298868894577, + 0.0031480733305215836, + 0.006120114121586084, + 0.017388269305229187, + 0.003352338680997491, + 0.012684009037911892, + 0.002432148903608322, + 0.0028310087509453297, + 0.00511703547090292, + 0.0018190444679930806, + 0.002963660517707467, + 0.007402685005217791, + 0.0021653061266988516, + 0.0074841841123998165, + 0.015532155521214008, + 0.021781075745821, + 0.05063290148973465, + 0.014821227639913559, + 0.016685746610164642, + 0.20705349743366241, + 0.020137852057814598, + 0.20968855917453766, + 0.006256270222365856, + 0.005823447834700346, + 0.009484515525400639, + 0.006244494579732418, + 0.008752335794270039, + 0.008339250460267067, + 0.0068214694038033485, + 0.010271559469401836, + 0.006205182522535324, + 0.006469822954386473, + 0.005302969831973314, + 0.005299543496221304, + 0.016711466014385223, + 0.014198132790625095, + 0.008771378546953201, + 0.015234374441206455, + 0.002528988989070058, + 0.003278614953160286, + 0.005627716425806284, + 0.00376891135238111, + 0.005089952144771814, + 0.0034274321515113115, + 0.0026196229737251997, + 0.005113170947879553, + 0.006583122536540031, + 0.00427960604429245, + 0.009861818514764309, + 0.005170310381799936, + 0.005899331998080015, + 0.004944040440022945, + 0.004835579544305801, + 0.004535496234893799, + 0.0015496495179831982, + 0.0013458681060001254, + 0.0019261111738160253, + 0.0017548436298966408, + 0.0016586350975558162, + 0.001464401837438345, + 0.001558010815642774, + 0.002187553560361266, + 0.002587216906249523, + 0.0023052224423736334, + 0.003333509434014559, + 0.004147843457758427, + 0.0036576315760612488, + 0.0036917540710419416, + 0.004498492926359177, + 0.0023686583153903484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.01609722524881363 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559776902198792 + ], + "opposite_logprob_tokens": [ + -1.6005977392196655 + ], + "logprob_reward": -0.22559776902198792, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559776902198792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05612055957317352, + 0.23354022204875946, + 0.07878045737743378, + 0.39749225974082947, + 0.023119088262319565, + 0.030965980142354965, + 0.030144518241286278, + 0.09171532839536667, + 0.00885730516165495, + 0.04926425218582153 + ], + "credit_map": [ + 0.008981025777757168, + 0.005067320074886084, + 0.009808708913624287, + 0.0033896174281835556, + 0.005042994394898415, + 0.00965536292642355, + 0.005350275430828333, + 0.00882525835186243, + 0.013439971953630447, + 0.01571892946958542, + 0.04660334810614586, + 0.012231359258294106, + 0.03383899852633476, + 0.06128568947315216, + 0.013116772286593914, + 0.037305183708667755, + 0.004697883035987616, + 0.004500298295170069, + 0.012567565776407719, + 0.0028943216893821955, + 0.006909122224897146, + 0.02314029447734356, + 0.004659320693463087, + 0.019411657005548477, + 0.011332602240145206, + 0.015939341858029366, + 0.0337337888777256, + 0.01097074244171381, + 0.011642481200397015, + 0.1477442979812622, + 0.014261615462601185, + 0.15186737477779388, + 0.0018856602255254984, + 0.0017735533183440566, + 0.002779053058475256, + 0.0029943501576781273, + 0.003249221947044134, + 0.00316137308254838, + 0.003860687604174018, + 0.0034151903819292784, + 0.0022072584833949804, + 0.0027653274592012167, + 0.004240764304995537, + 0.0025173677131533623, + 0.005473720841109753, + 0.004697022959589958, + 0.0034309574402868748, + 0.005633561406284571, + 0.0026442930102348328, + 0.0036423129495233297, + 0.003163947258144617, + 0.003720941487699747, + 0.0035568696912378073, + 0.00514961127191782, + 0.0032021321821957827, + 0.005064412485808134, + 0.004129032604396343, + 0.003249000059440732, + 0.004336894024163485, + 0.0037918235175311565, + 0.006622441112995148, + 0.0346008762717247, + 0.003340683178976178, + 0.03164459019899368, + 0.0008780393982306123, + 0.0010126344859600067, + 0.0011908103479072452, + 0.000934201234485954, + 0.001600839663296938, + 0.0008915818179957569, + 0.0010725477477535605, + 0.001276650233194232, + 0.0060243308544158936, + 0.00537033611908555, + 0.005306808277964592, + 0.010999859310686588, + 0.004048929549753666, + 0.006025180220603943, + 0.005485102534294128, + 0.006003709509968758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.22559776902198792 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024975319392979145 + ], + "opposite_logprob_tokens": [ + -6.002497673034668 + ], + "logprob_reward": -0.0024975319392979145, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024975319392979145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10210055112838745, + 0.09791824966669083, + 0.12673576176166534, + 0.05717191845178604, + 0.09028882533311844, + 0.04997866228222847, + 0.07861752808094025, + 0.10745062679052353, + 0.039935000240802765, + 0.2498028427362442 + ], + "credit_map": [ + 0.007841450162231922, + 0.006247573532164097, + 0.020367108285427094, + 0.004096202086657286, + 0.0063728759996593, + 0.030613340437412262, + 0.00630964757874608, + 0.020252350717782974, + 0.009288082830607891, + 0.009248496033251286, + 0.014676122926175594, + 0.0062443530187010765, + 0.016481438651680946, + 0.016263125464320183, + 0.008531501516699791, + 0.01718512922525406, + 0.009043403901159763, + 0.005426547024399042, + 0.016646012663841248, + 0.00476787518709898, + 0.005116935819387436, + 0.044430430978536606, + 0.007150974124670029, + 0.034153588116168976, + 0.004159681499004364, + 0.004829341545701027, + 0.006329412572085857, + 0.00622522784397006, + 0.005277183372527361, + 0.010156533680856228, + 0.0063974070362746716, + 0.013797130435705185, + 0.006573879159986973, + 0.006260474678128958, + 0.013727283105254173, + 0.005538412369787693, + 0.015496993437409401, + 0.01744772680103779, + 0.009974198415875435, + 0.015269863419234753, + 0.0037310596089810133, + 0.0033421723637729883, + 0.006666368804872036, + 0.003657738910987973, + 0.005777656100690365, + 0.011377102695405483, + 0.005788056645542383, + 0.009638508781790733, + 0.005289525259286165, + 0.010722668841481209, + 0.012006568722426891, + 0.008400644175708294, + 0.009282167069613934, + 0.012510077096521854, + 0.008156596682965755, + 0.012249277904629707, + 0.007447689771652222, + 0.005143420305103064, + 0.008549102582037449, + 0.0074050286784768105, + 0.007488162722438574, + 0.03106324002146721, + 0.011334212496876717, + 0.02901976741850376, + 0.004512574058026075, + 0.0036202892661094666, + 0.005375179927796125, + 0.006233525462448597, + 0.006189161911606789, + 0.004186877515166998, + 0.0046625458635389805, + 0.005154844373464584, + 0.03846122696995735, + 0.025783721357584, + 0.03093920461833477, + 0.05315764993429184, + 0.02666281908750534, + 0.030996263027191162, + 0.02358534187078476, + 0.020216606557369232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.0024975319392979145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera begins a slow, dramatic push-in, a medium shot from a slightly high angle unfolds across a realistic globe suspended in a desolate black void. The globe's surface features a semi-gloss finish, with deep sapphire oceans and continents detailed in realistic topographic hues of forest green, sandy yellow, and earthy brown, accented with delicate latitude and vertical lines. A single spotlight shining from above casts a bright highlight on the northern hemisphere, creating sharp shadows that emphasize its perfectly spherical shape. The transformation subtly begins as the solid textures of the continents begin to shimmer and lose their contours, liquefying as if made of hot wax. Distinctive land masses melt, their vibrant colors melting and flowing into the ocean in viscous, shimmering streams. As the camera steadily approaches, the scene focuses on a mesmerizing chaos: greens, browns, and yellows blend into deep blue, forming vibrant, swirling pools of mixed colors that create a hypnotic marble pattern across the globe's surface, transforming the familiar map into a fluid, abstract painting of beautiful decay.", + "seed": 1311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_3213_seed_1311_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": -0.05682246037758887, + "vlm_logprob": -0.05682246037758887, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.05682246037758887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03203502669930458 + ], + "opposite_logprob_tokens": [ + -3.5320351123809814 + ], + "logprob_reward": -0.03203502669930458, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03203502669930458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21009936928749084, + 0.037948329001665115, + 0.08993799984455109, + 0.02040211297571659, + 0.0761956050992012, + 0.17599859833717346, + 0.07882647961378098, + 0.055179957300424576, + 0.08571121096611023, + 0.16970038414001465 + ], + "credit_map": [ + 0.04256763309240341, + 0.015102880075573921, + 0.0321519710123539, + 0.014214209280908108, + 0.015815680846571922, + 0.015231937170028687, + 0.03441014140844345, + 0.040604885667562485, + 0.006192481145262718, + 0.004380641970783472, + 0.005284473765641451, + 0.0052728671580553055, + 0.003736523212864995, + 0.0035131978802382946, + 0.004161233548074961, + 0.005406905896961689, + 0.01017195824533701, + 0.00659518176689744, + 0.011752602644264698, + 0.008044380694627762, + 0.007480646017938852, + 0.017489876598119736, + 0.011895140632987022, + 0.016508210450410843, + 0.002166199265047908, + 0.002361753024160862, + 0.002468652790412307, + 0.0031791809014976025, + 0.0024559898301959038, + 0.0023497785441577435, + 0.0027295351028442383, + 0.0026910214219242334, + 0.008273941464722157, + 0.007701549679040909, + 0.008904683403670788, + 0.007821325212717056, + 0.007723123300820589, + 0.012060936540365219, + 0.012935729697346687, + 0.010774308815598488, + 0.011883876286447048, + 0.01459298562258482, + 0.044490065425634384, + 0.010828012600541115, + 0.010200578719377518, + 0.01503255870193243, + 0.02032340131700039, + 0.04864709451794624, + 0.004564960487186909, + 0.004178447648882866, + 0.01972731575369835, + 0.005254417657852173, + 0.012691481970250607, + 0.0034285024739801884, + 0.024425026029348373, + 0.004556319676339626, + 0.0052862828597426414, + 0.008086823858320713, + 0.006965418346226215, + 0.004146132152527571, + 0.008732649497687817, + 0.005161371547728777, + 0.007235088851302862, + 0.009566185064613819, + 0.011566024273633957, + 0.01119239442050457, + 0.007531445007771254, + 0.005145288072526455, + 0.015932651236653328, + 0.01304535660892725, + 0.007140704430639744, + 0.014157337136566639, + 0.030173564329743385, + 0.0184242632240057, + 0.022811511531472206, + 0.013569977134466171, + 0.016859330236911774, + 0.029456976801156998, + 0.018701311200857162, + 0.01970343105494976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.03203502669930458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0161137655377388 + ], + "opposite_logprob_tokens": [ + -4.141113758087158 + ], + "logprob_reward": -0.0161137655377388, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0161137655377388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14057736098766327, + 0.08933110535144806, + 0.015282735228538513, + 0.073578841984272, + 0.04901965707540512, + 0.1083596721291542, + 0.35712307691574097, + 0.013047480955719948, + 0.10883000493049622, + 0.044850099831819534 + ], + "credit_map": [ + 0.014180039055645466, + 0.0038471270818263292, + 0.044078487902879715, + 0.006769075989723206, + 0.006932215765118599, + 0.009136590175330639, + 0.0113466652110219, + 0.04428716003894806, + 0.010527071543037891, + 0.00820150040090084, + 0.009536084719002247, + 0.01128295436501503, + 0.010592050850391388, + 0.009302854537963867, + 0.016402872279286385, + 0.013485717587172985, + 0.0015237980987876654, + 0.0020590031053870916, + 0.0015381068224087358, + 0.0015389823820441961, + 0.0015731375897303224, + 0.00256119342520833, + 0.001937317312695086, + 0.0025511959102004766, + 0.004926944151520729, + 0.006547448690980673, + 0.005290464032441378, + 0.008650979027152061, + 0.009665058925747871, + 0.011977004818618298, + 0.013724889606237411, + 0.01279605645686388, + 0.004815940745174885, + 0.0055952831171453, + 0.006377986166626215, + 0.006102958228439093, + 0.005591917783021927, + 0.004688809625804424, + 0.009703398682177067, + 0.006143362261354923, + 0.014494444243609905, + 0.009602641686797142, + 0.017620839178562164, + 0.012645483016967773, + 0.007931461557745934, + 0.010610626079142094, + 0.010462112724781036, + 0.02499205991625786, + 0.024151954799890518, + 0.014753819443285465, + 0.09805460274219513, + 0.014841620810329914, + 0.045499540865421295, + 0.014714186079800129, + 0.1236620619893074, + 0.021445272490382195, + 0.0009589797118678689, + 0.0015868226764723659, + 0.0013396026333793998, + 0.001148342969827354, + 0.0020453014876693487, + 0.0015420421259477735, + 0.0021492675878107548, + 0.002277121879160404, + 0.007664254400879145, + 0.005187280476093292, + 0.005419383756816387, + 0.00363549729809165, + 0.007025882601737976, + 0.03608602285385132, + 0.00655096722766757, + 0.03726072236895561, + 0.0046839602291584015, + 0.004051681607961655, + 0.005611749831587076, + 0.0060786716639995575, + 0.00640499871224165, + 0.005031611304730177, + 0.00686482060700655, + 0.006122605409473181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0161137655377388 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751937326043844 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006751937326043844, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751937326043844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04792027547955513, + 0.07619068771600723, + 0.0332886204123497, + 0.060824278742074966, + 0.07831456512212753, + 0.12569811940193176, + 0.3160261809825897, + 0.0344194695353508, + 0.17662915587425232, + 0.05068862810730934 + ], + "credit_map": [ + 0.006765749305486679, + 0.0034275369253009558, + 0.007297218311578035, + 0.004025642294436693, + 0.004473854787647724, + 0.00587737699970603, + 0.007794549223035574, + 0.008258343674242496, + 0.009075130335986614, + 0.006049735005944967, + 0.006427733693271875, + 0.0064265187829732895, + 0.010569990612566471, + 0.007703954819589853, + 0.01845400407910347, + 0.011483626440167427, + 0.0035506407730281353, + 0.002701322315260768, + 0.004327805712819099, + 0.0035079605877399445, + 0.003143141744658351, + 0.0056655979715287685, + 0.004637780599296093, + 0.005754371173679829, + 0.004544632509350777, + 0.006150939501821995, + 0.00477576395496726, + 0.005684581585228443, + 0.006375020369887352, + 0.012023260816931725, + 0.010337105952203274, + 0.010932973586022854, + 0.007172673940658569, + 0.010625024326145649, + 0.009743282571434975, + 0.006451180204749107, + 0.00863823015242815, + 0.014144760556519032, + 0.012230407446622849, + 0.009309015236794949, + 0.00868954136967659, + 0.01247571874409914, + 0.029370291158556938, + 0.010232977569103241, + 0.009245487861335278, + 0.009366140700876713, + 0.019011585041880608, + 0.02730638161301613, + 0.02216208167374134, + 0.013553694821894169, + 0.0868925228714943, + 0.013213732279837132, + 0.039478983730077744, + 0.012805270962417126, + 0.10931520164012909, + 0.01860467717051506, + 0.003691731719300151, + 0.004280354827642441, + 0.003342149779200554, + 0.0028848573565483093, + 0.0056181373074650764, + 0.003500727005302906, + 0.0043383631855249405, + 0.0067631457932293415, + 0.0061552501283586025, + 0.007313775364309549, + 0.007131726946681738, + 0.004127766005694866, + 0.009838073514401913, + 0.06320666521787643, + 0.009893422946333885, + 0.06896249204874039, + 0.0073000723496079445, + 0.0054060653783380985, + 0.006231061648577452, + 0.006069938652217388, + 0.00528856972232461, + 0.007772672921419144, + 0.005497994367033243, + 0.007122253067791462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.006751937326043844 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558940947055817 + ], + "opposite_logprob_tokens": [ + -1.600589394569397 + ], + "logprob_reward": -0.22558940947055817, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558940947055817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15779657661914825, + 0.04779626429080963, + 0.07049913704395294, + 0.04075748100876808, + 0.16047637164592743, + 0.1954086571931839, + 0.16334030032157898, + 0.03559901937842369, + 0.05387672781944275, + 0.07444943487644196 + ], + "credit_map": [ + 0.02593713440001011, + 0.010793372057378292, + 0.0218668133020401, + 0.015279759652912617, + 0.01294754073023796, + 0.01713169552385807, + 0.03015017881989479, + 0.023690076544880867, + 0.005624684970825911, + 0.004469133913516998, + 0.004750863648951054, + 0.005145365372300148, + 0.008287572301924229, + 0.004125526174902916, + 0.00920714158564806, + 0.006185973063111305, + 0.007601351011544466, + 0.005989797879010439, + 0.007798139005899429, + 0.007616299670189619, + 0.005889067426323891, + 0.012884194031357765, + 0.010020784102380276, + 0.012699505314230919, + 0.003774927230551839, + 0.004864501766860485, + 0.004570420831441879, + 0.005251302849501371, + 0.004902757238596678, + 0.005394823383539915, + 0.006648920476436615, + 0.005349823739379644, + 0.013195894658565521, + 0.016491461545228958, + 0.01786966808140278, + 0.01915733888745308, + 0.018138650804758072, + 0.026894154027104378, + 0.03029470704495907, + 0.018434500321745872, + 0.012757138349115849, + 0.009750141762197018, + 0.043614983558654785, + 0.010260343551635742, + 0.009911340661346912, + 0.023465558886528015, + 0.018082275986671448, + 0.0675668865442276, + 0.01159934513270855, + 0.007087172474712133, + 0.039659131318330765, + 0.0063996571116149426, + 0.022633137181401253, + 0.007015669252723455, + 0.0577876940369606, + 0.011158503592014313, + 0.003111364785581827, + 0.0027132767718285322, + 0.003457573475316167, + 0.002784284995868802, + 0.004738620948046446, + 0.0050039091147482395, + 0.00722158420830965, + 0.006568403448909521, + 0.0036191341932862997, + 0.004515598528087139, + 0.004286637529730797, + 0.0033256378956139088, + 0.011918634176254272, + 0.007574601098895073, + 0.005687218625098467, + 0.012949258089065552, + 0.008751478977501392, + 0.005831767339259386, + 0.009790853597223759, + 0.011887081898748875, + 0.007933606393635273, + 0.01066820789128542, + 0.012247368693351746, + 0.007339067291468382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.22558940947055817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036221628542989492 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036221628542989492, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036221628542989492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19781699776649475, + 0.08873769640922546, + 0.06408093124628067, + 0.04516448825597763, + 0.057088423520326614, + 0.18410846590995789, + 0.11753831803798676, + 0.04118683561682701, + 0.1324668675661087, + 0.07181102782487869 + ], + "credit_map": [ + 0.02279002219438553, + 0.009151830337941647, + 0.05303165316581726, + 0.013736516237258911, + 0.011444834060966969, + 0.013698797672986984, + 0.019889038056135178, + 0.054074306041002274, + 0.010241560637950897, + 0.007814058102667332, + 0.006390269845724106, + 0.01086443942040205, + 0.012485898099839687, + 0.008625257760286331, + 0.018165582790970802, + 0.01415062416344881, + 0.0074881287291646, + 0.005384016316384077, + 0.008924949914216995, + 0.007248736917972565, + 0.005954001098871231, + 0.011010102927684784, + 0.0073889559134840965, + 0.01068203803151846, + 0.002009220654144883, + 0.0036924041341990232, + 0.003846235340461135, + 0.006122487597167492, + 0.006148192100226879, + 0.0062545305117964745, + 0.009318024851381779, + 0.007773392368108034, + 0.007191153708845377, + 0.00637023476883769, + 0.005127619951963425, + 0.006230691913515329, + 0.0039065442979335785, + 0.008224371820688248, + 0.009097163565456867, + 0.010940643958747387, + 0.0157768614590168, + 0.02103899046778679, + 0.04144391790032387, + 0.020091727375984192, + 0.01421779952943325, + 0.010528762824833393, + 0.03037610463798046, + 0.030634302645921707, + 0.013910448178648949, + 0.00846567377448082, + 0.023111088201403618, + 0.008894906379282475, + 0.015120954252779484, + 0.009628204628825188, + 0.028925016522407532, + 0.00948202796280384, + 0.0036907701287418604, + 0.004541851580142975, + 0.0039024045690894127, + 0.003765583038330078, + 0.00717928959056735, + 0.0037673532497137785, + 0.007642264943569899, + 0.0066973199136555195, + 0.007419251836836338, + 0.009503941982984543, + 0.007226996123790741, + 0.00480958167463541, + 0.012537588365375996, + 0.03696844354271889, + 0.008521825075149536, + 0.04547923058271408, + 0.010851211845874786, + 0.007811254356056452, + 0.010002218186855316, + 0.008870390243828297, + 0.006049757357686758, + 0.0123172951862216, + 0.0074362619780004025, + 0.008472643792629242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0036221628542989492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera begins a slow, dramatic push-in, a medium shot from a slightly high angle unfolds across a realistic globe suspended in a desolate black void. The globe's surface features a semi-gloss finish, with deep sapphire oceans and continents detailed in realistic topographic hues of forest green, sandy yellow, and earthy brown, accented with delicate latitude and vertical lines. A single spotlight shining from above casts a bright highlight on the northern hemisphere, creating sharp shadows that emphasize its perfectly spherical shape. The transformation subtly begins as the solid textures of the continents begin to shimmer and lose their contours, liquefying as if made of hot wax. Distinctive land masses melt, their vibrant colors melting and flowing into the ocean in viscous, shimmering streams. As the camera steadily approaches, the scene focuses on a mesmerizing chaos: greens, browns, and yellows blend into deep blue, forming vibrant, swirling pools of mixed colors that create a hypnotic marble pattern across the globe's surface, transforming the familiar map into a fluid, abstract painting of beautiful decay.", + "seed": 1312, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_3213_seed_1312_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": -0.07579867327585817, + "vlm_logprob": -0.07579867327585817, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07579867327585817, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0936320573091507 + ], + "opposite_logprob_tokens": [ + -2.4686319828033447 + ], + "logprob_reward": -0.0936320573091507, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0936320573091507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.584588348865509, + 0.041664399206638336, + 0.03833642229437828, + 0.013686000369489193, + 0.038293831050395966, + 0.07586797326803207, + 0.0165519118309021, + 0.0626138225197792, + 0.0665549710392952, + 0.06184236705303192 + ], + "credit_map": [ + 0.06389082223176956, + 0.009531762450933456, + 0.23110532760620117, + 0.0176090095192194, + 0.012681905180215836, + 0.05249590799212456, + 0.011900882236659527, + 0.1853727251291275, + 0.005987989716231823, + 0.003757837228477001, + 0.006220400333404541, + 0.0042100888676941395, + 0.0040351627394557, + 0.005972716957330704, + 0.0033495582174509764, + 0.008130643516778946, + 0.004602815955877304, + 0.004051075782626867, + 0.007979718036949635, + 0.004767061676830053, + 0.004035842604935169, + 0.004057338926941156, + 0.0038857401814311743, + 0.0049568237736821175, + 0.0019699661061167717, + 0.0020299935713410378, + 0.0014429735019803047, + 0.0009716743370518088, + 0.002423152793198824, + 0.001740302424877882, + 0.0012930603697896004, + 0.00181487575173378, + 0.005139590241014957, + 0.005222357343882322, + 0.0028839975129812956, + 0.0035346257500350475, + 0.005501165054738522, + 0.006034941412508488, + 0.003762277076020837, + 0.0062148733995854855, + 0.005283927544951439, + 0.013154975138604641, + 0.00409121997654438, + 0.0031775212846696377, + 0.027127902954816818, + 0.009582214057445526, + 0.004332315642386675, + 0.00911789108067751, + 0.00252905604429543, + 0.0020068807061761618, + 0.002066162647679448, + 0.0017233381513506174, + 0.003372243605554104, + 0.0014700748724862933, + 0.0021553735714405775, + 0.0012287801364436746, + 0.009343286976218224, + 0.015143784694373608, + 0.005340988282114267, + 0.005147796124219894, + 0.007609367370605469, + 0.005115167703479528, + 0.006695523392409086, + 0.00821789912879467, + 0.007518279366195202, + 0.005286844912916422, + 0.004142191726714373, + 0.007578140590339899, + 0.02700134925544262, + 0.008280064910650253, + 0.004148499108850956, + 0.0025995986070483923, + 0.007892061956226826, + 0.00868729967623949, + 0.010178216733038425, + 0.010310378856956959, + 0.005181425251066685, + 0.007175371050834656, + 0.006881569977849722, + 0.005536037962883711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.0936320573091507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332673966884613 + ], + "opposite_logprob_tokens": [ + -3.7733266353607178 + ], + "logprob_reward": -0.02332673966884613, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332673966884613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5999096035957336, + 0.0502142496407032, + 0.012297255918383598, + 0.017570948228240013, + 0.017852632328867912, + 0.1693011224269867, + 0.028723247349262238, + 0.0246637761592865, + 0.03779229521751404, + 0.04167483001947403 + ], + "credit_map": [ + 0.06956297904253006, + 0.008427505381405354, + 0.24856096506118774, + 0.015778396278619766, + 0.008866584859788418, + 0.054241735488176346, + 0.00979568064212799, + 0.18467579782009125, + 0.005514869932085276, + 0.0029280432499945164, + 0.01414045412093401, + 0.0031872084364295006, + 0.002440102631226182, + 0.004121749196201563, + 0.0022507139947265387, + 0.015631111338734627, + 0.001545610255561769, + 0.0008609611541032791, + 0.0015253201127052307, + 0.0018581284675747156, + 0.0015003335429355502, + 0.0020905814599245787, + 0.0012586336815729737, + 0.001657686778344214, + 0.002409448614344001, + 0.0017154351808130741, + 0.0024083785247057676, + 0.0021030432544648647, + 0.002560346620157361, + 0.002298815408721566, + 0.0019492459250614047, + 0.002126233885064721, + 0.002349935472011566, + 0.0023645709734410048, + 0.00156599550973624, + 0.0015064787585288286, + 0.002167705912142992, + 0.0029163463041186333, + 0.0022990142460912466, + 0.002682585036382079, + 0.01583496853709221, + 0.023147013038396835, + 0.007986122742295265, + 0.007190800737589598, + 0.029323440045118332, + 0.03834878280758858, + 0.00910714641213417, + 0.03836284577846527, + 0.0033088051714003086, + 0.0032195921521633863, + 0.004103229381144047, + 0.0045509906485676765, + 0.0050534410402178764, + 0.0028064458165317774, + 0.003039108356460929, + 0.002641634549945593, + 0.003249585162848234, + 0.004260600544512272, + 0.0032425939571112394, + 0.00385082489810884, + 0.0026171498466283083, + 0.0018635564483702183, + 0.002678347285836935, + 0.0029011208098381758, + 0.003758637700229883, + 0.0030668063554912806, + 0.0024257733020931482, + 0.00465087965130806, + 0.014740138314664364, + 0.004794037900865078, + 0.002852845937013626, + 0.001503175706602633, + 0.0052995034493505955, + 0.005160580854862928, + 0.006873311009258032, + 0.008116748183965683, + 0.003506700275465846, + 0.004333252552896738, + 0.00475822389125824, + 0.0036265095695853233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.02332673966884613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004654762800782919 + ], + "opposite_logprob_tokens": [ + -5.379654884338379 + ], + "logprob_reward": -0.004654762800782919, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004654762800782919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5859007835388184, + 0.09792820364236832, + 0.014879774302244186, + 0.014328635297715664, + 0.024903591722249985, + 0.17230059206485748, + 0.010901876725256443, + 0.04215557873249054, + 0.009113556705415249, + 0.027587277814745903 + ], + "credit_map": [ + 0.06892697513103485, + 0.011402946896851063, + 0.23611021041870117, + 0.015518479980528355, + 0.013032344169914722, + 0.048302676528692245, + 0.009658245369791985, + 0.1829490065574646, + 0.00860076304525137, + 0.005552438087761402, + 0.02518385462462902, + 0.005678784567862749, + 0.0061210948042571545, + 0.009359528310596943, + 0.0058861118741333485, + 0.03154565021395683, + 0.0016530327266082168, + 0.0012791180051863194, + 0.0027173920534551144, + 0.001981934765353799, + 0.0015274790348485112, + 0.0027153585106134415, + 0.0014474118361249566, + 0.001558050629682839, + 0.0017282931366935372, + 0.0016690430929884315, + 0.0016879532486200333, + 0.0015952505636960268, + 0.0028035964351147413, + 0.0017178766429424286, + 0.0015976453432813287, + 0.001528979162685573, + 0.0031067361123859882, + 0.0033463852014392614, + 0.0020344608929008245, + 0.0017319608014076948, + 0.0036012257914990187, + 0.0038754099514335394, + 0.002770531689748168, + 0.004436887335032225, + 0.01457590889185667, + 0.02537337690591812, + 0.009271657094359398, + 0.008197044022381306, + 0.05815178528428078, + 0.021123934537172318, + 0.009727075695991516, + 0.02587985247373581, + 0.0015666018007323146, + 0.0015851182397454977, + 0.0016584501136094332, + 0.0015296940691769123, + 0.001397681306116283, + 0.0010614723432809114, + 0.0011349368141964078, + 0.0009679239592514932, + 0.005180294159799814, + 0.008319145999848843, + 0.003591432236135006, + 0.0036402747500687838, + 0.005344009026885033, + 0.0044745695777237415, + 0.005428926553577185, + 0.006176931783556938, + 0.001075501088052988, + 0.0012304757256060839, + 0.0010119660291820765, + 0.001107953255996108, + 0.002006324240937829, + 0.0010643224231898785, + 0.0008930933545343578, + 0.000723921344615519, + 0.002487308345735073, + 0.0026866737753152847, + 0.0053098974749445915, + 0.006277360022068024, + 0.0021019328851252794, + 0.0025206655263900757, + 0.003809769870713353, + 0.0023936762008816004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.004654762800782919 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521210312843323 + ], + "opposite_logprob_tokens": [ + -1.5021209716796875 + ], + "logprob_reward": -0.2521210312843323, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521210312843323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41948622465133667, + 0.07935737073421478, + 0.019321123138070107, + 0.019377682358026505, + 0.05458144471049309, + 0.25620412826538086, + 0.016679346561431885, + 0.02673139050602913, + 0.03647731989622116, + 0.07178391516208649 + ], + "credit_map": [ + 0.046007610857486725, + 0.006632375065237284, + 0.16903093457221985, + 0.012209051288664341, + 0.008560117334127426, + 0.03687027469277382, + 0.008836529217660427, + 0.13133937120437622, + 0.007193201687186956, + 0.005402325186878443, + 0.018252979964017868, + 0.008813976310193539, + 0.005642256699502468, + 0.008013194426894188, + 0.00781174935400486, + 0.01822769269347191, + 0.0026443609967827797, + 0.0015114956768229604, + 0.00473902327939868, + 0.0022108398843556643, + 0.0016979528591036797, + 0.0025240611284971237, + 0.0018864035373553634, + 0.0021069878712296486, + 0.0026294435374438763, + 0.002726476639509201, + 0.001861784141510725, + 0.001763649401254952, + 0.003668098011985421, + 0.0026732354890555143, + 0.0019487187964841723, + 0.0021062782034277916, + 0.007213120348751545, + 0.00730052450671792, + 0.0038395537994802, + 0.005260451231151819, + 0.0072845169343054295, + 0.007882771082222462, + 0.006243512500077486, + 0.009556992910802364, + 0.025026965886354446, + 0.04510902613401413, + 0.014025215990841389, + 0.013490783050656319, + 0.07280251383781433, + 0.023939643055200577, + 0.019890761002898216, + 0.04191923886537552, + 0.0016307902988046408, + 0.002107989974319935, + 0.0020063247065991163, + 0.0022443050984293222, + 0.0032647133339196444, + 0.0018070973455905914, + 0.002212619176134467, + 0.0014055088395252824, + 0.002938607009127736, + 0.003223035717383027, + 0.0033738957718014717, + 0.002537541789934039, + 0.004746434278786182, + 0.003614611690863967, + 0.0028809895738959312, + 0.003416277701035142, + 0.0032675040420144796, + 0.0033550504595041275, + 0.002445101272314787, + 0.005779758561402559, + 0.01344586443156004, + 0.004410246387124062, + 0.00249885069206357, + 0.0012749488232657313, + 0.006386220455169678, + 0.006864756811410189, + 0.012892559170722961, + 0.015944303944706917, + 0.00507714506238699, + 0.0071073793806135654, + 0.01025392860174179, + 0.00725761940702796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.2521210312843323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258775316178799 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258775316178799, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258775316178799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2738720178604126, + 0.09896114468574524, + 0.02583920769393444, + 0.024881821125745773, + 0.033057283610105515, + 0.39410021901130676, + 0.015485968440771103, + 0.02117573656141758, + 0.05680805817246437, + 0.05581851676106453 + ], + "credit_map": [ + 0.03246364742517471, + 0.008749358355998993, + 0.1080361008644104, + 0.009009328670799732, + 0.006979180034250021, + 0.025805603712797165, + 0.008980350568890572, + 0.07384845614433289, + 0.010808733291924, + 0.006516411434859037, + 0.022650497034192085, + 0.0072204372845590115, + 0.007317558396607637, + 0.010085844434797764, + 0.006298975553363562, + 0.028062688186764717, + 0.0031109522096812725, + 0.0026387812104076147, + 0.005520485807210207, + 0.0026883091777563095, + 0.0027948503848165274, + 0.0039983754977583885, + 0.0024739427026361227, + 0.002613512799143791, + 0.0038725961931049824, + 0.0027374664787203074, + 0.003680530935525894, + 0.003778493031859398, + 0.0025039496831595898, + 0.0027749415021389723, + 0.0025717844255268574, + 0.002962063066661358, + 0.004931263159960508, + 0.004920406732708216, + 0.0025148990098387003, + 0.0025039121974259615, + 0.005137324333190918, + 0.004364317283034325, + 0.0033141050953418016, + 0.005371056497097015, + 0.03507979214191437, + 0.06699217855930328, + 0.019656479358673096, + 0.018628401681780815, + 0.13014072179794312, + 0.04008366912603378, + 0.023671530187129974, + 0.059847455471754074, + 0.001445553731173277, + 0.001791645074263215, + 0.0014406966511160135, + 0.0017058298690244555, + 0.0037431882228702307, + 0.0019223205745220184, + 0.0019698920659720898, + 0.0014668423682451248, + 0.0028186868876218796, + 0.0042113312520086765, + 0.0027964983601123095, + 0.0021843300200998783, + 0.0027934846002608538, + 0.0019499013433232903, + 0.0020789504051208496, + 0.002342553809285164, + 0.005246669054031372, + 0.007222713902592659, + 0.004023345652967691, + 0.007123442366719246, + 0.019178180024027824, + 0.006723863072693348, + 0.0038344520144164562, + 0.0034553934819996357, + 0.00534791499376297, + 0.005286631640046835, + 0.010774840600788593, + 0.014840713702142239, + 0.0038602116983383894, + 0.004020405001938343, + 0.00814862921833992, + 0.0035391703713685274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.005258775316178799 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera begins a slow, dramatic push-in, a medium shot from a slightly high angle unfolds across a realistic globe suspended in a desolate black void. The globe's surface features a semi-gloss finish, with deep sapphire oceans and continents detailed in realistic topographic hues of forest green, sandy yellow, and earthy brown, accented with delicate latitude and vertical lines. A single spotlight shining from above casts a bright highlight on the northern hemisphere, creating sharp shadows that emphasize its perfectly spherical shape. The transformation subtly begins as the solid textures of the continents begin to shimmer and lose their contours, liquefying as if made of hot wax. Distinctive land masses melt, their vibrant colors melting and flowing into the ocean in viscous, shimmering streams. As the camera steadily approaches, the scene focuses on a mesmerizing chaos: greens, browns, and yellows blend into deep blue, forming vibrant, swirling pools of mixed colors that create a hypnotic marble pattern across the globe's surface, transforming the familiar map into a fluid, abstract painting of beautiful decay.", + "seed": 1313, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_4/index_3213_seed_1313_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": -0.06955316765233874, + "vlm_logprob": -0.06955316765233874, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06955316765233874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023793289437890053 + ], + "opposite_logprob_tokens": [ + -3.8987932205200195 + ], + "logprob_reward": -0.023793289437890053, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023793289437890053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09875677525997162, + 0.057523805648088455, + 0.07192502915859222, + 0.2068903148174286, + 0.10717132687568665, + 0.06504253298044205, + 0.14164148271083832, + 0.13337965309619904, + 0.07448966801166534, + 0.04317934438586235 + ], + "credit_map": [ + 0.00822011660784483, + 0.008268012665212154, + 0.010886795818805695, + 0.00839208159595728, + 0.011308511719107628, + 0.018180232495069504, + 0.019262388348579407, + 0.014238636940717697, + 0.0046420469880104065, + 0.005712392274290323, + 0.0063386703841388226, + 0.005930854938924313, + 0.007024451624602079, + 0.010665646754205227, + 0.006467901170253754, + 0.01074184663593769, + 0.010383925400674343, + 0.005631121341139078, + 0.009803332388401031, + 0.006988792214542627, + 0.006407782901078463, + 0.01283908262848854, + 0.007491322699934244, + 0.012379673309624195, + 0.015600427985191345, + 0.012994327582418919, + 0.03415761888027191, + 0.030102208256721497, + 0.02703079767525196, + 0.036396242678165436, + 0.012757578864693642, + 0.03785112500190735, + 0.00760653056204319, + 0.00903373397886753, + 0.022314827889204025, + 0.01111058983951807, + 0.010918139480054379, + 0.01814187504351139, + 0.010556114837527275, + 0.01748952642083168, + 0.003313907189294696, + 0.004446030594408512, + 0.006527465768158436, + 0.004162751138210297, + 0.0054177772253751755, + 0.017631465569138527, + 0.006220824550837278, + 0.017322320491075516, + 0.011990709230303764, + 0.01909461058676243, + 0.0219300277531147, + 0.01574004627764225, + 0.016518665477633476, + 0.019972849637269974, + 0.02050717920064926, + 0.015887398272752762, + 0.015126393176615238, + 0.016608230769634247, + 0.01440498512238264, + 0.01614229567348957, + 0.021751560270786285, + 0.0204633679240942, + 0.016828879714012146, + 0.012053964659571648, + 0.007840626873075962, + 0.007739389315247536, + 0.012572082690894604, + 0.008627249859273434, + 0.005948558449745178, + 0.012579369358718395, + 0.00780278816819191, + 0.011379608884453773, + 0.003872684435918927, + 0.004109077155590057, + 0.007727464195340872, + 0.004444476682692766, + 0.0038918331265449524, + 0.006201827432960272, + 0.007741753477603197, + 0.005190232302993536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.023793289437890053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033729564398527145 + ], + "opposite_logprob_tokens": [ + -3.4087295532226562 + ], + "logprob_reward": -0.033729564398527145, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033729564398527145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09162589907646179, + 0.08042341470718384, + 0.07633969932794571, + 0.2508625388145447, + 0.13575215637683868, + 0.05882929638028145, + 0.06167345494031906, + 0.09566780924797058, + 0.10120724886655807, + 0.047618504613637924 + ], + "credit_map": [ + 0.007819144055247307, + 0.007950766943395138, + 0.009293913841247559, + 0.009043826721608639, + 0.010058512911200523, + 0.015733681619167328, + 0.018025465309619904, + 0.013700595125555992, + 0.007665668148547411, + 0.007553276140242815, + 0.006149240769445896, + 0.006353999022394419, + 0.009195243939757347, + 0.016978532075881958, + 0.009741285815834999, + 0.01678616926074028, + 0.006950365845113993, + 0.0053493850864470005, + 0.010104273445904255, + 0.013462869450449944, + 0.007572980597615242, + 0.01345042884349823, + 0.007868079468607903, + 0.011581308208405972, + 0.02293413318693638, + 0.0163064356893301, + 0.05294456332921982, + 0.033736202865839005, + 0.022188374772667885, + 0.04167675971984863, + 0.021736741065979004, + 0.039339322596788406, + 0.009888785891234875, + 0.008786983788013458, + 0.03229179605841637, + 0.012329566292464733, + 0.010903776623308659, + 0.023325344547629356, + 0.014816489070653915, + 0.023409422487020493, + 0.004396605305373669, + 0.003939876798540354, + 0.007096383720636368, + 0.005046345759183168, + 0.006012867204844952, + 0.013546863570809364, + 0.005829267203807831, + 0.012961084023118019, + 0.006809794344007969, + 0.008121909573674202, + 0.009225653484463692, + 0.0056500062346458435, + 0.008261769078671932, + 0.009141705930233002, + 0.006894772406667471, + 0.007567844353616238, + 0.01120756845921278, + 0.01167339738458395, + 0.012920026667416096, + 0.008271313272416592, + 0.013850954361259937, + 0.013445388525724411, + 0.015985164791345596, + 0.008313996717333794, + 0.01294502429664135, + 0.007744255941361189, + 0.019802095368504524, + 0.012661932967603207, + 0.0069420430809259415, + 0.01635061204433441, + 0.015086904168128967, + 0.0096743768081069, + 0.003410032019019127, + 0.0032400202471762896, + 0.012018859386444092, + 0.005179294850677252, + 0.004313374403864145, + 0.007334448862820864, + 0.006528664845973253, + 0.005593809299170971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.033729564398527145 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052655343897640705 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.0052655343897640705, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052655343897640705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12074533849954605, + 0.11328735947608948, + 0.09667704999446869, + 0.1464613378047943, + 0.10160035640001297, + 0.0844837948679924, + 0.09477082639932632, + 0.09738799929618835, + 0.04105440899729729, + 0.10353148728609085 + ], + "credit_map": [ + 0.011422133073210716, + 0.009691224433481693, + 0.012180694378912449, + 0.011850693263113499, + 0.01642361842095852, + 0.022169943898916245, + 0.01586013101041317, + 0.021146884188055992, + 0.007158328779041767, + 0.011374013498425484, + 0.008273913525044918, + 0.01150326058268547, + 0.014059790410101414, + 0.022406240925192833, + 0.015394879505038261, + 0.023116938769817352, + 0.00996338576078415, + 0.007506015244871378, + 0.014464160427451134, + 0.010280176997184753, + 0.014455927535891533, + 0.015927806496620178, + 0.011060670018196106, + 0.01301890891045332, + 0.010807208716869354, + 0.010766997933387756, + 0.026883909478783607, + 0.023841945454478264, + 0.015593607909977436, + 0.022556466981768608, + 0.01341173890978098, + 0.02259945496916771, + 0.006933802738785744, + 0.0075548384338617325, + 0.01637893170118332, + 0.008129778318107128, + 0.009847934357821941, + 0.019452596083283424, + 0.017121709883213043, + 0.01618075929582119, + 0.004306792747229338, + 0.005617902148514986, + 0.00976294744759798, + 0.0063014402985572815, + 0.006565665826201439, + 0.023252347484230995, + 0.00603222893550992, + 0.022644473239779472, + 0.006783524062484503, + 0.008635814301669598, + 0.011561977677047253, + 0.010504008270800114, + 0.010452826507389545, + 0.012232241220772266, + 0.021268082782626152, + 0.01333235576748848, + 0.008585505187511444, + 0.012483689934015274, + 0.015044484287500381, + 0.011133060790598392, + 0.018212981522083282, + 0.013726899400353432, + 0.00970599427819252, + 0.008495383895933628, + 0.0037681974936276674, + 0.00307298987172544, + 0.006824638694524765, + 0.004893524572253227, + 0.0040827845223248005, + 0.0068565914407372475, + 0.0058424463495612144, + 0.00571323512122035, + 0.015054463408887386, + 0.008191702887415886, + 0.01729479245841503, + 0.009135501459240913, + 0.009576394222676754, + 0.014587757177650928, + 0.011339383199810982, + 0.018351493403315544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0052655343897640705 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813540995121002 + ], + "opposite_logprob_tokens": [ + -1.4063540697097778 + ], + "logprob_reward": -0.2813540995121002, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813540995121002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07144223153591156, + 0.18401999771595, + 0.10762188583612442, + 0.17214535176753998, + 0.10045982152223587, + 0.11528116464614868, + 0.05580895394086838, + 0.049242667853832245, + 0.050086941570043564, + 0.09389099478721619 + ], + "credit_map": [ + 0.004115920979529619, + 0.005063389427959919, + 0.0062470766715705395, + 0.0068473489955067635, + 0.0073672193102538586, + 0.015905344858765602, + 0.009447412565350533, + 0.016448521986603737, + 0.018184363842010498, + 0.019060784950852394, + 0.015514136292040348, + 0.016361601650714874, + 0.02218785136938095, + 0.03363778442144394, + 0.02365533821284771, + 0.03541814535856247, + 0.01138609554618597, + 0.010315511375665665, + 0.016115007922053337, + 0.015006848610937595, + 0.01136686559766531, + 0.015955660492181778, + 0.009808382019400597, + 0.01766752079129219, + 0.015184170566499233, + 0.009304908104240894, + 0.04057377204298973, + 0.03106565587222576, + 0.015242857858538628, + 0.02458072081208229, + 0.014355655759572983, + 0.0218376312404871, + 0.009647889994084835, + 0.0080692358314991, + 0.022800857201218605, + 0.012169552966952324, + 0.01076572947204113, + 0.014270830899477005, + 0.010663938708603382, + 0.012071791104972363, + 0.006361434236168861, + 0.006878250744193792, + 0.010666501708328724, + 0.007698452565819025, + 0.013042512349784374, + 0.02894977107644081, + 0.014376971870660782, + 0.02730727382004261, + 0.004021802451461554, + 0.004332378972321749, + 0.008628969080746174, + 0.0068421862088143826, + 0.005603650584816933, + 0.0074430303648114204, + 0.011836142279207706, + 0.007100793533027172, + 0.0053693135268986225, + 0.005478702951222658, + 0.006485466379672289, + 0.006085776723921299, + 0.006968603003770113, + 0.0076260995119810104, + 0.0062812697142362595, + 0.00494743837043643, + 0.004880157765001059, + 0.004394129384309053, + 0.0070978617295622826, + 0.006169370841234922, + 0.004643702879548073, + 0.007278435863554478, + 0.00727692199870944, + 0.008346357382833958, + 0.010641063563525677, + 0.0072165499441325665, + 0.016040654852986336, + 0.007565757725387812, + 0.007316807750612497, + 0.012686296366155148, + 0.015337497927248478, + 0.017086369916796684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.2813540995121002 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\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 medium shot of a realistic globe suspended in a desolate black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera begin with a slow push-in from a slightly high angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the globe's surface semi-gloss with deep sapphire oceans and continents in forest green, sandy yellow, and earthy brown?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single spotlight from above cast a bright highlight on the northern hemisphere with sharp shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the continents melt and flow into the ocean, forming vibrant, swirling pools of mixed colors that create a marble pattern?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036233505234122276 + ], + "opposite_logprob_tokens": [ + -5.6286234855651855 + ], + "logprob_reward": -0.0036233505234122276, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036233505234122276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12580625712871552, + 0.09441298991441727, + 0.11139489710330963, + 0.12231983989477158, + 0.09730859845876694, + 0.07466381788253784, + 0.07178393751382828, + 0.149288609623909, + 0.07715287804603577, + 0.07586818933486938 + ], + "credit_map": [ + 0.014335071668028831, + 0.012560428120195866, + 0.01309802196919918, + 0.009969521313905716, + 0.01632147654891014, + 0.022946512326598167, + 0.016575699672102928, + 0.019999533891677856, + 0.009696543216705322, + 0.009023398160934448, + 0.007519521750509739, + 0.008047050796449184, + 0.010126613080501556, + 0.018346214666962624, + 0.01258013490587473, + 0.01907351240515709, + 0.0198188666254282, + 0.009783000685274601, + 0.01698332466185093, + 0.013700663112103939, + 0.014285726472735405, + 0.015347912907600403, + 0.008578876033425331, + 0.012896524742245674, + 0.011039636097848415, + 0.012276815250515938, + 0.022239483892917633, + 0.017932405695319176, + 0.015683673322200775, + 0.017140870913863182, + 0.013012158684432507, + 0.012994799762964249, + 0.006304558366537094, + 0.00781691912561655, + 0.018215063959360123, + 0.008535252884030342, + 0.010943055152893066, + 0.01841622032225132, + 0.011338118463754654, + 0.01573941297829151, + 0.003645184449851513, + 0.00417860085144639, + 0.00898177083581686, + 0.007890625856816769, + 0.010800805874168873, + 0.01264046411961317, + 0.013683789409697056, + 0.012842576950788498, + 0.004575964529067278, + 0.006178486160933971, + 0.010731645859777927, + 0.00911000557243824, + 0.007115491200238466, + 0.009573815390467644, + 0.01579318568110466, + 0.00870534498244524, + 0.017346007749438286, + 0.015011388808488846, + 0.019193626940250397, + 0.01667976751923561, + 0.027085868641734123, + 0.024997128173708916, + 0.016007529571652412, + 0.012967289425432682, + 0.009943966753780842, + 0.012554371729493141, + 0.009482904337346554, + 0.009370812214910984, + 0.007077148649841547, + 0.01102390605956316, + 0.009690447710454464, + 0.008009321056306362, + 0.010494327172636986, + 0.006839803885668516, + 0.016901183873414993, + 0.006265018135309219, + 0.006168211344629526, + 0.0098197590559721, + 0.01030238252133131, + 0.00907750055193901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.0036233505234122276 + } + ], + "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/0000160/group_7/index_tensor([8595, 8595])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_tensor([8595, 8595])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9863a5d41769b288548ae73ef1e07c27735cbdaf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_tensor([8595, 8595])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-speed camera, in a single, seamless shot, begins from a dynamic, low-angle upward sweep, closely following a massive queen bee, draped in an ornate golden carapace. Its massive bulk casts sharp shadows across the vast, futuristic amphitheater. A massive, circular spotlight overhead casts a harsh, cold white light, revealing every intricate pattern on its gleaming metallic carapace. Its enormous compound eyes gleam with a cold, calculating radiance, while its translucent wings blur in their rapid vibrations. As the camera steadily advances, the queen bee lunges forward, its massive, razor-sharp jaws closing with unstoppable force, precise and brutal, instantly biting down upon the head and thorax of its prey—a hideous, beetle-like creature. In extreme slow motion, time seems frozen as streams of emerald green bodily fluids burst and splash like pearls from the opponent's wounds, each drop gleaming eerily under the cold light. Simultaneously, the opponent's hard exoskeleton shatters, sending countless tiny fragments scattering in all directions, creating a stunning scene imbued with a violent aesthetic. The camera follows the largest fragment as it spins and flies, then quickly pulls back to focus on the queen bee, her mandibles retracted, standing proudly, her body stained with the enemy's bodily fluids. Against the blurred backdrop of an energy shield and the silhouettes of a frenzied crowd, this undeniable end to this fiercely competitive match is clearly announced.", + "seed": 20937, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20937_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": -0.07208040487021208, + "vlm_logprob": -0.07208040487021208, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.07208040487021208, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009582829661667347 + ], + "opposite_logprob_tokens": [ + -5.134582996368408 + ], + "logprob_reward": -0.009582829661667347, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009582829661667347, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12919050455093384, + 0.0760614350438118, + 0.14490848779678345, + 0.08414886146783829, + 0.0375635027885437, + 0.12022572755813599, + 0.10838635265827179, + 0.06539623439311981, + 0.13241109251976013, + 0.101707823574543 + ], + "credit_map": [ + 0.008503147400915623, + 0.022260230034589767, + 0.015582816675305367, + 0.013402551412582397, + 0.029109736904501915, + 0.01781047135591507, + 0.01092543825507164, + 0.01159610878676176, + 0.006761282216757536, + 0.007128039840608835, + 0.021437574177980423, + 0.005915714893490076, + 0.006760071963071823, + 0.010515597648918629, + 0.006167702842503786, + 0.011375446803867817, + 0.01713448017835617, + 0.03225364536046982, + 0.015106978826224804, + 0.012502025812864304, + 0.02225106582045555, + 0.019247153773903847, + 0.01240629144012928, + 0.0140068419277668, + 0.012770667672157288, + 0.015291423536837101, + 0.009771707467734814, + 0.010590062476694584, + 0.009088443592190742, + 0.009107735008001328, + 0.007877368479967117, + 0.00965145230293274, + 0.005635766312479973, + 0.005251489579677582, + 0.003314352361485362, + 0.004432425368577242, + 0.0052131530828773975, + 0.002704853191971779, + 0.006962218787521124, + 0.004049243405461311, + 0.013869849033653736, + 0.02044076845049858, + 0.00889089796692133, + 0.019326014444231987, + 0.020701957866549492, + 0.008990230970084667, + 0.0138180460780859, + 0.014187962748110294, + 0.011513203382492065, + 0.017094163224101067, + 0.010428914800286293, + 0.02659495919942856, + 0.01403531339019537, + 0.0075014191679656506, + 0.00968360435217619, + 0.01153477095067501, + 0.011149666272103786, + 0.007059422321617603, + 0.011577912606298923, + 0.0066212438978254795, + 0.006416947580873966, + 0.008349820971488953, + 0.005883511155843735, + 0.008337713778018951, + 0.02299986407160759, + 0.007791261188685894, + 0.023174919188022614, + 0.030702315270900726, + 0.008309025317430496, + 0.01957891881465912, + 0.009997090324759483, + 0.00985769834369421, + 0.01651340164244175, + 0.012524875812232494, + 0.013202948495745659, + 0.013660961762070656, + 0.007290182635188103, + 0.019969500601291656, + 0.009404043667018414, + 0.009141900576651096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.009582829661667347 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543829157948494 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543829157948494, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543829157948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42800360918045044, + 0.11171825230121613, + 0.04149401932954788, + 0.057944633066654205, + 0.06984587758779526, + 0.045217763632535934, + 0.03181101754307747, + 0.05781345069408417, + 0.0955093577504158, + 0.060641929507255554 + ], + "credit_map": [ + 0.03389966860413551, + 0.07627927511930466, + 0.053214237093925476, + 0.02861323207616806, + 0.07682932913303375, + 0.07527351379394531, + 0.03217950463294983, + 0.05171489343047142, + 0.010317191481590271, + 0.011560383252799511, + 0.025440657511353493, + 0.010411418974399567, + 0.00938464142382145, + 0.015832271426916122, + 0.01005706749856472, + 0.018714623525738716, + 0.004108928143978119, + 0.006361991632729769, + 0.004847618751227856, + 0.003254840848967433, + 0.005008340813219547, + 0.004854999948292971, + 0.005132890306413174, + 0.007924415171146393, + 0.007639950606971979, + 0.006030447315424681, + 0.009537034668028355, + 0.009353451430797577, + 0.008322864770889282, + 0.005570380017161369, + 0.006197369657456875, + 0.005293143447488546, + 0.013809471391141415, + 0.010678542777895927, + 0.009521973319351673, + 0.008883252739906311, + 0.006554424297064543, + 0.005255673546344042, + 0.009097733534872532, + 0.00604481203481555, + 0.005694522522389889, + 0.0061819786205887794, + 0.0052418699488043785, + 0.010081220418214798, + 0.005491478368639946, + 0.003136070678010583, + 0.0036369552835822105, + 0.005753675475716591, + 0.004817496053874493, + 0.0054309070110321045, + 0.003435684135183692, + 0.0049166749231517315, + 0.0034074087161570787, + 0.0020593199878931046, + 0.0043335361406207085, + 0.003409991506487131, + 0.010462752543389797, + 0.005512466188520193, + 0.011836305260658264, + 0.005974447820335627, + 0.008077056147158146, + 0.0068611521273851395, + 0.004941337276250124, + 0.004147940780967474, + 0.009154753759503365, + 0.007728370372205973, + 0.017490049824118614, + 0.027467742562294006, + 0.005867939908057451, + 0.014961951412260532, + 0.006258102599531412, + 0.0065804580226540565, + 0.00951287243515253, + 0.00671761017292738, + 0.0064411163330078125, + 0.009455101564526558, + 0.007545724045485258, + 0.00817098468542099, + 0.006022024434059858, + 0.0067764995619654655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.012543829157948494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288540542125702 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.023288540542125702, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288540542125702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05309761315584183, + 0.08509321510791779, + 0.24032068252563477, + 0.09655226767063141, + 0.04549508914351463, + 0.10275965183973312, + 0.022000355646014214, + 0.18997427821159363, + 0.11168840527534485, + 0.053018487989902496 + ], + "credit_map": [ + 0.0045448835007846355, + 0.007983585819602013, + 0.00546678202226758, + 0.006798757240176201, + 0.008622344583272934, + 0.010725010186433792, + 0.004473945591598749, + 0.004482307471334934, + 0.008147169835865498, + 0.008022317662835121, + 0.02480282261967659, + 0.0069014099426567554, + 0.007242741994559765, + 0.010066797025501728, + 0.007991781458258629, + 0.011918174102902412, + 0.02650590054690838, + 0.05519161373376846, + 0.022084854543209076, + 0.023031005635857582, + 0.030453501269221306, + 0.03177127242088318, + 0.026668930426239967, + 0.024613603949546814, + 0.009760233573615551, + 0.008198278956115246, + 0.014235319569706917, + 0.016186198219656944, + 0.014295781031250954, + 0.009079733863472939, + 0.013710388913750648, + 0.011086336337029934, + 0.0054603321477770805, + 0.005988124292343855, + 0.0059232464991509914, + 0.0056664301082491875, + 0.004468193743377924, + 0.0031599903013557196, + 0.006488376762717962, + 0.008340397849678993, + 0.015638507902622223, + 0.014695890247821808, + 0.014416943304240704, + 0.01943078264594078, + 0.013011392205953598, + 0.006374140735715628, + 0.009861244820058346, + 0.009330750443041325, + 0.002848717151209712, + 0.004448905121535063, + 0.002674676710739732, + 0.002815724117681384, + 0.0026838360354304314, + 0.001605172990821302, + 0.00230175512842834, + 0.0026215692050755024, + 0.038561008870601654, + 0.016100116074085236, + 0.05612938478589058, + 0.01749185472726822, + 0.01064347568899393, + 0.02397332526743412, + 0.011837989091873169, + 0.015237112529575825, + 0.01593361794948578, + 0.008286863565444946, + 0.01908741518855095, + 0.02195483259856701, + 0.008602214977145195, + 0.01715232990682125, + 0.008869868703186512, + 0.011801265180110931, + 0.008049041964113712, + 0.006824089214205742, + 0.006217343732714653, + 0.006650615483522415, + 0.0049959756433963776, + 0.007848083972930908, + 0.006630540359765291, + 0.00580279715359211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.023288540542125702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369256854057312 + ], + "opposite_logprob_tokens": [ + -3.4086925983428955 + ], + "logprob_reward": -0.03369256854057312, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369256854057312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05662098899483681, + 0.053665678948163986, + 0.11271809786558151, + 0.05723217874765396, + 0.025511084124445915, + 0.058588311076164246, + 0.033303651958703995, + 0.4890090823173523, + 0.08566994965076447, + 0.02768094278872013 + ], + "credit_map": [ + 0.0037549531552940607, + 0.010119774378836155, + 0.005591357126832008, + 0.006057417020201683, + 0.011996322311460972, + 0.00810723565518856, + 0.00536844227463007, + 0.005625488236546516, + 0.005175672937184572, + 0.004893619567155838, + 0.013946736231446266, + 0.00418487936258316, + 0.004621917847543955, + 0.0062701283022761345, + 0.00558013329282403, + 0.008992590941488743, + 0.012172296643257141, + 0.024645300582051277, + 0.010933035984635353, + 0.01044610608369112, + 0.01576385460793972, + 0.014825924299657345, + 0.011909831315279007, + 0.012021759524941444, + 0.006572060286998749, + 0.00911263283342123, + 0.008498867973685265, + 0.008892283774912357, + 0.005844346713274717, + 0.00634946720674634, + 0.006816606502979994, + 0.005145912524312735, + 0.00395930465310812, + 0.0032786373049020767, + 0.0022945960517972708, + 0.002768776146695018, + 0.0025788333732634783, + 0.0020369752310216427, + 0.0042814044281840324, + 0.004312556702643633, + 0.00790524110198021, + 0.008108886890113354, + 0.008105548098683357, + 0.010822933167219162, + 0.00582476332783699, + 0.004180682357400656, + 0.006043801084160805, + 0.007596452720463276, + 0.0056174579076468945, + 0.0047285608015954494, + 0.005298176314681768, + 0.005836362484842539, + 0.002900214632973075, + 0.0025891007389873266, + 0.0032559914980083704, + 0.003077786648645997, + 0.09798558056354523, + 0.04091184586286545, + 0.13257507979869843, + 0.037845201790332794, + 0.04413369297981262, + 0.0613805390894413, + 0.028786160051822662, + 0.045390963554382324, + 0.013504019938409328, + 0.00786413624882698, + 0.014018458314239979, + 0.019666610285639763, + 0.005866374354809523, + 0.010263568721711636, + 0.006830508820712566, + 0.007656277623027563, + 0.003524447325617075, + 0.004506321623921394, + 0.0041959346272051334, + 0.004276455845683813, + 0.002340354723855853, + 0.0035793182905763388, + 0.0026359145995229483, + 0.002622193656861782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.03369256854057312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2812942564487457 + ], + "opposite_logprob_tokens": [ + -1.4062942266464233 + ], + "logprob_reward": -0.2812942564487457, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2812942564487457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07652943581342697, + 0.1485602855682373, + 0.25485455989837646, + 0.03644213452935219, + 0.027587078511714935, + 0.06251353025436401, + 0.023778479546308517, + 0.20996370911598206, + 0.12651225924491882, + 0.03325847536325455 + ], + "credit_map": [ + 0.006340201012790203, + 0.007841936312615871, + 0.013180959969758987, + 0.008608364500105381, + 0.007972272112965584, + 0.014392754063010216, + 0.009509499184787273, + 0.008683462627232075, + 0.014050806872546673, + 0.012422372587025166, + 0.03810407221317291, + 0.01147062610834837, + 0.013469116762280464, + 0.021351156756281853, + 0.012380869127810001, + 0.02531127817928791, + 0.032881807535886765, + 0.05968097224831581, + 0.026134122163057327, + 0.022304216399788857, + 0.032508473843336105, + 0.033880867063999176, + 0.0248789694160223, + 0.02258513867855072, + 0.002931330818682909, + 0.004038031212985516, + 0.004476457368582487, + 0.004736856557428837, + 0.007515236735343933, + 0.004284248221665621, + 0.00553977582603693, + 0.002920201513916254, + 0.004156035371124744, + 0.0035123368725180626, + 0.002583459485322237, + 0.0031827944330871105, + 0.003390289144590497, + 0.002045382047072053, + 0.005212974268943071, + 0.003503808518871665, + 0.008843925781548023, + 0.009704184718430042, + 0.006265702657401562, + 0.010119191370904446, + 0.009407227858901024, + 0.005177027080208063, + 0.005879837088286877, + 0.007116439752280712, + 0.0033269894775003195, + 0.004239100497215986, + 0.0019403110491111875, + 0.0042595271952450275, + 0.002654481679201126, + 0.0017874776385724545, + 0.002749469131231308, + 0.0028211248572915792, + 0.0419175885617733, + 0.018436329439282417, + 0.061723366379737854, + 0.019625237211585045, + 0.012113183736801147, + 0.02653420902788639, + 0.011290710419416428, + 0.01832311414182186, + 0.025571806356310844, + 0.011389028280973434, + 0.02021845430135727, + 0.02172670140862465, + 0.009077304042875767, + 0.015470458194613457, + 0.009578677825629711, + 0.013479849323630333, + 0.005242282524704933, + 0.006082370411604643, + 0.004510046914219856, + 0.0037513114511966705, + 0.002566240495070815, + 0.006221499759703875, + 0.0023427254054695368, + 0.0025420021265745163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.2812942564487457 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed camera, in a single, seamless shot, begins from a dynamic, low-angle upward sweep, closely following a massive queen bee, draped in an ornate golden carapace. Its massive bulk casts sharp shadows across the vast, futuristic amphitheater. A massive, circular spotlight overhead casts a harsh, cold white light, revealing every intricate pattern on its gleaming metallic carapace. Its enormous compound eyes gleam with a cold, calculating radiance, while its translucent wings blur in their rapid vibrations. As the camera steadily advances, the queen bee lunges forward, its massive, razor-sharp jaws closing with unstoppable force, precise and brutal, instantly biting down upon the head and thorax of its prey—a hideous, beetle-like creature. In extreme slow motion, time seems frozen as streams of emerald green bodily fluids burst and splash like pearls from the opponent's wounds, each drop gleaming eerily under the cold light. Simultaneously, the opponent's hard exoskeleton shatters, sending countless tiny fragments scattering in all directions, creating a stunning scene imbued with a violent aesthetic. The camera follows the largest fragment as it spins and flies, then quickly pulls back to focus on the queen bee, her mandibles retracted, standing proudly, her body stained with the enemy's bodily fluids. Against the blurred backdrop of an energy shield and the silhouettes of a frenzied crowd, this undeniable end to this fiercely competitive match is clearly announced.", + "seed": 20938, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20938_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": -0.18718750029802322, + "vlm_logprob": -0.18718750029802322, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.18718750029802322, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031629133969545364 + ], + "opposite_logprob_tokens": [ + -3.6566290855407715 + ], + "logprob_reward": -0.031629133969545364, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.031629133969545364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17031130194664001, + 0.0936734601855278, + 0.06465205550193787, + 0.10532385110855103, + 0.05954568833112717, + 0.053011246025562286, + 0.06856117397546768, + 0.30770042538642883, + 0.04078070819377899, + 0.03644004464149475 + ], + "credit_map": [ + 0.018355557695031166, + 0.022610854357481003, + 0.017809659242630005, + 0.028869422152638435, + 0.01580415479838848, + 0.019880378618836403, + 0.02536625973880291, + 0.02161501534283161, + 0.009101614356040955, + 0.013838727958500385, + 0.009228305891156197, + 0.008828656747937202, + 0.016368981450796127, + 0.017724674195051193, + 0.010317837819457054, + 0.008264654316008091, + 0.01077544316649437, + 0.007867837324738503, + 0.006544078700244427, + 0.008381202816963196, + 0.0067908260971307755, + 0.006748269312083721, + 0.006142782978713512, + 0.011401612311601639, + 0.011845594272017479, + 0.01650264486670494, + 0.010241172276437283, + 0.011274345219135284, + 0.01975051499903202, + 0.010699396021664143, + 0.00777257326990366, + 0.01723760925233364, + 0.010116973891854286, + 0.008972552604973316, + 0.00775620061904192, + 0.005201851949095726, + 0.006561710499227047, + 0.005330098792910576, + 0.005742642562836409, + 0.0098636569455266, + 0.007490892894566059, + 0.005412215366959572, + 0.006833407562226057, + 0.007190051954239607, + 0.009429576806724072, + 0.004690722096711397, + 0.005931743886321783, + 0.006032638251781464, + 0.007209902163594961, + 0.006598878651857376, + 0.007343449164181948, + 0.009581681340932846, + 0.012434803880751133, + 0.010161462239921093, + 0.00779323885217309, + 0.0074377600103616714, + 0.052710872143507004, + 0.030053360387682915, + 0.054547179490327835, + 0.028139451518654823, + 0.03165460377931595, + 0.04148068651556969, + 0.030288243666291237, + 0.03882604092359543, + 0.006808709818869829, + 0.0033844292629510164, + 0.0034892887342721224, + 0.00394175061956048, + 0.00376218743622303, + 0.0077086323872208595, + 0.008169218897819519, + 0.003516490338370204, + 0.005743482615798712, + 0.004017589148133993, + 0.003572179237380624, + 0.003934718202799559, + 0.003776902798563242, + 0.005452260375022888, + 0.005734953097999096, + 0.004207958467304707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.031629133969545364 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"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": [ + 0.1624777913093567, + 0.04810001328587532, + 0.02869492769241333, + 0.04084736108779907, + 0.18602445721626282, + 0.03233204782009125, + 0.06949105858802795, + 0.17812128365039825, + 0.0779457613825798, + 0.17596535384655 + ], + "credit_map": [ + 0.015557903796434402, + 0.01941969431936741, + 0.016918964684009552, + 0.02415645308792591, + 0.027564777061343193, + 0.018624287098646164, + 0.020906854420900345, + 0.019328860566020012, + 0.005525055341422558, + 0.006968312431126833, + 0.0037356186658143997, + 0.004302342887967825, + 0.008698547258973122, + 0.007868312299251556, + 0.005003546364605427, + 0.005998275242745876, + 0.003668443765491247, + 0.004801968578249216, + 0.0031781026627868414, + 0.0028677124064415693, + 0.004707727115601301, + 0.0025967671535909176, + 0.002654176903888583, + 0.004220031201839447, + 0.004999712575227022, + 0.005128198768943548, + 0.005698007065802813, + 0.0043557616882026196, + 0.005565397907048464, + 0.003747175447642803, + 0.0036775001790374517, + 0.007675607223063707, + 0.03706185519695282, + 0.02753291092813015, + 0.020590294152498245, + 0.017385389655828476, + 0.01816468872129917, + 0.01612136699259281, + 0.015235272236168385, + 0.033932678401470184, + 0.006005266215652227, + 0.0034707682207226753, + 0.00406025443226099, + 0.003826054045930505, + 0.003767543938010931, + 0.0030012628994882107, + 0.0038415167946368456, + 0.004359381273388863, + 0.012649371288716793, + 0.006764881778508425, + 0.007866713218390942, + 0.008507790043950081, + 0.00789373368024826, + 0.011547046713531017, + 0.007135558873414993, + 0.007125958800315857, + 0.03621938079595566, + 0.019269993528723717, + 0.029579460620880127, + 0.01600450463593006, + 0.018430082127451897, + 0.024280158802866936, + 0.015537613071501255, + 0.018800098448991776, + 0.013179326429963112, + 0.006792524829506874, + 0.007856719195842743, + 0.007279958110302687, + 0.008434831164777279, + 0.015771497040987015, + 0.012936576269567013, + 0.005694330669939518, + 0.02553352154791355, + 0.017892656847834587, + 0.018557701259851456, + 0.016430117189884186, + 0.02493909001350403, + 0.025864116847515106, + 0.03084098920226097, + 0.01590716652572155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 9 + ], + "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 main subject a massive queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495632812380791 + ], + "opposite_logprob_tokens": [ + -2.9299564361572266 + ], + "logprob_reward": -0.05495632812380791, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495632812380791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12358220666646957, + 0.07024355232715607, + 0.038798198103904724, + 0.17995619773864746, + 0.025552714243531227, + 0.06331072002649307, + 0.05769314989447594, + 0.16568145155906677, + 0.14042340219020844, + 0.13475842773914337 + ], + "credit_map": [ + 0.013767648488283157, + 0.01708202436566353, + 0.011950507760047913, + 0.02407006174325943, + 0.00910674687474966, + 0.014221214689314365, + 0.017994733527302742, + 0.015389276668429375, + 0.01192043349146843, + 0.007743251975625753, + 0.007848971523344517, + 0.00736209424212575, + 0.007361431606113911, + 0.011659882962703705, + 0.0070259179919958115, + 0.009321565739810467, + 0.005523930303752422, + 0.005577474366873503, + 0.004704513121396303, + 0.004708366934210062, + 0.004522757139056921, + 0.0039042187854647636, + 0.0035805345978587866, + 0.006276403553783894, + 0.02745216339826584, + 0.02413182333111763, + 0.024018922820687294, + 0.014581573195755482, + 0.0298294797539711, + 0.016126707196235657, + 0.014632301405072212, + 0.02918323315680027, + 0.004394225310534239, + 0.0035087917931377888, + 0.0029587813187390566, + 0.0028025538194924593, + 0.0027451550122350454, + 0.0024996339343488216, + 0.002131144981831312, + 0.004512427840381861, + 0.011025610379874706, + 0.005185181275010109, + 0.007373434025794268, + 0.007661702111363411, + 0.007201641798019409, + 0.006178725510835648, + 0.005871964618563652, + 0.012812460772693157, + 0.013219166547060013, + 0.004728384781628847, + 0.010843619704246521, + 0.003525197971612215, + 0.004456121940165758, + 0.007944915443658829, + 0.006117631681263447, + 0.006858116947114468, + 0.027293968945741653, + 0.01604648306965828, + 0.031363386660814285, + 0.010628036223351955, + 0.015031200833618641, + 0.023847214877605438, + 0.015200123190879822, + 0.026271024718880653, + 0.021084189414978027, + 0.013723315671086311, + 0.01363750733435154, + 0.013585809618234634, + 0.017007099464535713, + 0.01953697018325329, + 0.028936387971043587, + 0.012912135571241379, + 0.01842757873237133, + 0.01874879002571106, + 0.014161351136863232, + 0.009027480147778988, + 0.01959267444908619, + 0.019343245774507523, + 0.024551501497626305, + 0.010905792005360126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.05495632812380791 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306419938802719 + ], + "opposite_logprob_tokens": [ + -3.1680641174316406 + ], + "logprob_reward": -0.04306419938802719, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306419938802719, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09544535726308823, + 0.024236837401986122, + 0.02672901563346386, + 0.09290474653244019, + 0.017079420387744904, + 0.014707265421748161, + 0.045090045779943466, + 0.08595450967550278, + 0.12241234630346298, + 0.47544044256210327 + ], + "credit_map": [ + 0.009741775691509247, + 0.010770870372653008, + 0.009314336813986301, + 0.013358277268707752, + 0.012445949018001556, + 0.013204189948737621, + 0.014619054272770882, + 0.011990907602012157, + 0.0031333048827946186, + 0.003233059076592326, + 0.002993244444951415, + 0.002795788459479809, + 0.0031733845826238394, + 0.0032687122002243996, + 0.0028935251757502556, + 0.0027458176482468843, + 0.005407869815826416, + 0.0030312605667859316, + 0.0029990780167281628, + 0.002457431284710765, + 0.0020853118039667606, + 0.0028327691834419966, + 0.002081173937767744, + 0.005834123119711876, + 0.011097513139247894, + 0.013550689443945885, + 0.01014637015759945, + 0.008392061106860638, + 0.018460843712091446, + 0.009433580562472343, + 0.00797221064567566, + 0.013851472176611423, + 0.0032112691551446915, + 0.001883731340058148, + 0.0023395095486193895, + 0.0016518732300028205, + 0.0017522461712360382, + 0.001574033172801137, + 0.001455251476727426, + 0.0032115071080625057, + 0.002847803058102727, + 0.0019138576462864876, + 0.0015221924986690283, + 0.0014042970724403858, + 0.0021630132105201483, + 0.0013660629047080874, + 0.002022265922278166, + 0.0014677734579890966, + 0.010483276098966599, + 0.00532024260610342, + 0.006825100630521774, + 0.006130479741841555, + 0.003693841863423586, + 0.005612077657133341, + 0.0037974209990352392, + 0.0032276089768856764, + 0.014980699867010117, + 0.009129191748797894, + 0.01576136238873005, + 0.006189640611410141, + 0.007442492060363293, + 0.011862555518746376, + 0.007480867672711611, + 0.013107703998684883, + 0.017933934926986694, + 0.009160473011434078, + 0.010280078276991844, + 0.012458769604563713, + 0.010786302387714386, + 0.01845904067158699, + 0.03007389046251774, + 0.01325984951108694, + 0.06372211128473282, + 0.0506756491959095, + 0.045348428189754486, + 0.036022085696458817, + 0.07015150040388107, + 0.05890607088804245, + 0.10301176458597183, + 0.0476028248667717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04306419938802719 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"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": [ + -0.6933579444885254 + ], + "opposite_logprob_tokens": [ + -0.6933579444885254 + ], + "logprob_reward": -0.6933579444885254, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933579444885254, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15167492628097534, + 0.10373149812221527, + 0.06104645878076553, + 0.2287464141845703, + 0.08085664361715317, + 0.03396747261285782, + 0.08371728658676147, + 0.058197781443595886, + 0.09005695581436157, + 0.10800451785326004 + ], + "credit_map": [ + 0.01608021929860115, + 0.017610572278499603, + 0.01774592511355877, + 0.021824195981025696, + 0.010259605944156647, + 0.021007154136896133, + 0.02740359492599964, + 0.019743671640753746, + 0.009068388491868973, + 0.01728173904120922, + 0.010977313853800297, + 0.012969265691936016, + 0.01562531851232052, + 0.017041699960827827, + 0.01172968652099371, + 0.009038085117936134, + 0.007655569817870855, + 0.01113936211913824, + 0.006238555070012808, + 0.006765453610569239, + 0.009507764130830765, + 0.005921404343098402, + 0.00613328255712986, + 0.0076850601471960545, + 0.04574224725365639, + 0.029721058905124664, + 0.02597932517528534, + 0.019777046516537666, + 0.03280865401029587, + 0.019750237464904785, + 0.021141964942216873, + 0.03382587060332298, + 0.016237063333392143, + 0.013393993489444256, + 0.007147364318370819, + 0.010095982812345028, + 0.008965793997049332, + 0.007321679033339024, + 0.005404038820415735, + 0.012290729209780693, + 0.006334022153168917, + 0.004413323476910591, + 0.0036567538045346737, + 0.004741053096950054, + 0.004318639170378447, + 0.003438089042901993, + 0.0035077473148703575, + 0.003557844553142786, + 0.010361536405980587, + 0.010449221357703209, + 0.01235263328999281, + 0.012722245417535305, + 0.010855067521333694, + 0.011624517850577831, + 0.008561762981116772, + 0.006790307816118002, + 0.010704766027629375, + 0.005879475735127926, + 0.012111750431358814, + 0.0049590906128287315, + 0.00476858438923955, + 0.008510598912835121, + 0.004989583510905504, + 0.0062739308923482895, + 0.013384083285927773, + 0.006339076440781355, + 0.010818816721439362, + 0.009117232635617256, + 0.007776794023811817, + 0.014301015064120293, + 0.019577622413635254, + 0.008742306381464005, + 0.0156998373568058, + 0.013814005069434643, + 0.009852821938693523, + 0.007950345054268837, + 0.018374672159552574, + 0.014522054232656956, + 0.019255509600043297, + 0.008535263128578663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.6933579444885254 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-speed camera, in a single, seamless shot, begins from a dynamic, low-angle upward sweep, closely following a massive queen bee, draped in an ornate golden carapace. Its massive bulk casts sharp shadows across the vast, futuristic amphitheater. A massive, circular spotlight overhead casts a harsh, cold white light, revealing every intricate pattern on its gleaming metallic carapace. Its enormous compound eyes gleam with a cold, calculating radiance, while its translucent wings blur in their rapid vibrations. As the camera steadily advances, the queen bee lunges forward, its massive, razor-sharp jaws closing with unstoppable force, precise and brutal, instantly biting down upon the head and thorax of its prey—a hideous, beetle-like creature. In extreme slow motion, time seems frozen as streams of emerald green bodily fluids burst and splash like pearls from the opponent's wounds, each drop gleaming eerily under the cold light. Simultaneously, the opponent's hard exoskeleton shatters, sending countless tiny fragments scattering in all directions, creating a stunning scene imbued with a violent aesthetic. The camera follows the largest fragment as it spins and flies, then quickly pulls back to focus on the queen bee, her mandibles retracted, standing proudly, her body stained with the enemy's bodily fluids. Against the blurred backdrop of an energy shield and the silhouettes of a frenzied crowd, this undeniable end to this fiercely competitive match is clearly announced.", + "seed": 20939, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20939_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": -0.10146423541009426, + "vlm_logprob": -0.10146423541009426, + "vlm_yes_no_margin": 3.1999999403953554, + "vlm_token_credit": -0.10146423541009426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013889795169234276 + ], + "opposite_logprob_tokens": [ + -4.638889789581299 + ], + "logprob_reward": -0.013889795169234276, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013889795169234276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05416905879974365, + 0.039917826652526855, + 0.03419458866119385, + 0.09115564078092575, + 0.026318082585930824, + 0.01881466992199421, + 0.45906397700309753, + 0.18023231625556946, + 0.07349485903978348, + 0.022638890892267227 + ], + "credit_map": [ + 0.008112696930766106, + 0.005978408269584179, + 0.009191116318106651, + 0.006684030871838331, + 0.00491571007296443, + 0.007361970841884613, + 0.005922704469412565, + 0.006002427544444799, + 0.007536400109529495, + 0.0030039490666240454, + 0.004565467592328787, + 0.0032958253286778927, + 0.002967492211610079, + 0.005221896339207888, + 0.003991250414401293, + 0.009335548616945744, + 0.0052220383659005165, + 0.0037233876064419746, + 0.005668529774993658, + 0.003728854935616255, + 0.003653026418760419, + 0.003753436729311943, + 0.004415205214172602, + 0.004030110780149698, + 0.010053165256977081, + 0.010904707945883274, + 0.01590307056903839, + 0.010753964073956013, + 0.010908630676567554, + 0.011346051469445229, + 0.01043185405433178, + 0.010854198597371578, + 0.003185139736160636, + 0.004006576258689165, + 0.00286956038326025, + 0.003509691683575511, + 0.002579155145213008, + 0.003380559617653489, + 0.0034255317877978086, + 0.0033618686720728874, + 0.0020862433593720198, + 0.0022723348811268806, + 0.003950454294681549, + 0.0017966064624488354, + 0.0016138071659952402, + 0.002699768403545022, + 0.0015362363774329424, + 0.0028592192102223635, + 0.013524782843887806, + 0.005571211688220501, + 0.03265800699591637, + 0.009135240688920021, + 0.009389578364789486, + 0.18535169959068298, + 0.01232330221682787, + 0.19111016392707825, + 0.01617446355521679, + 0.005232024006545544, + 0.011346527375280857, + 0.007204126100987196, + 0.004989715293049812, + 0.06299640983343124, + 0.013541324064135551, + 0.05874773859977722, + 0.007423542905598879, + 0.012588892132043839, + 0.007955043576657772, + 0.00670368317514658, + 0.012167219072580338, + 0.006820312701165676, + 0.012960588559508324, + 0.006875583902001381, + 0.0026351320557296276, + 0.0025366083718836308, + 0.0027201471384614706, + 0.0024211546406149864, + 0.0033835151698440313, + 0.002446544822305441, + 0.0032730521634221077, + 0.0032227386254817247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.013889795169234276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187537789344788 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187537789344788, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187537789344788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1303226798772812, + 0.03685162216424942, + 0.020312918350100517, + 0.030143780633807182, + 0.047672338783741, + 0.055587511509656906, + 0.5719507336616516, + 0.04428388178348541, + 0.04616137221455574, + 0.01671312004327774 + ], + "credit_map": [ + 0.023775532841682434, + 0.015419949777424335, + 0.025639628991484642, + 0.015075339004397392, + 0.009006310254335403, + 0.013400218449532986, + 0.011469272896647453, + 0.01653643324971199, + 0.007645732723176479, + 0.0026758078020066023, + 0.0037752934731543064, + 0.002472769469022751, + 0.0022342801094055176, + 0.004323896486312151, + 0.0031852805987000465, + 0.010538562200963497, + 0.0037010025698691607, + 0.0014306126395240426, + 0.00559946708381176, + 0.0019170705927535892, + 0.001626153476536274, + 0.001702723209746182, + 0.0026147475000470877, + 0.0017211443046107888, + 0.003862973302602768, + 0.0037139435298740864, + 0.005331643857061863, + 0.003521895967423916, + 0.0031224037520587444, + 0.002941835206001997, + 0.003259875113144517, + 0.004389211535453796, + 0.006483373232185841, + 0.005750579759478569, + 0.008993216790258884, + 0.004497475456446409, + 0.004430374596267939, + 0.006187547463923693, + 0.003498794510960579, + 0.007830977439880371, + 0.007161810994148254, + 0.005839679390192032, + 0.006754889618605375, + 0.006547777447849512, + 0.0035150228068232536, + 0.013284057378768921, + 0.004001888446509838, + 0.008482392877340317, + 0.017651764675974846, + 0.006636614445596933, + 0.04020802676677704, + 0.011775396764278412, + 0.011830342002213001, + 0.2320922464132309, + 0.01577790454030037, + 0.23597849905490875, + 0.0044330875389277935, + 0.0022365630138665438, + 0.0037958603352308273, + 0.002354554831981659, + 0.002851039171218872, + 0.013748059049248695, + 0.003235122188925743, + 0.011629598215222359, + 0.00741101847961545, + 0.00506733451038599, + 0.004147056490182877, + 0.003985558170825243, + 0.0049790265038609505, + 0.005026187747716904, + 0.007184959482401609, + 0.008360232226550579, + 0.0018508309731259942, + 0.0018746970454230905, + 0.0016687465831637383, + 0.0020821369253098965, + 0.0024679985363036394, + 0.0019311595242470503, + 0.0024290126748383045, + 0.002408539177849889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.018187537789344788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183909356594086 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183909356594086, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018183909356594086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026779239997267723, + 0.019942382350564003, + 0.011756157502532005, + 0.03860927373170853, + 0.017919043079018593, + 0.07232317328453064, + 0.6720185279846191, + 0.06530088931322098, + 0.01747909188270569, + 0.057872235774993896 + ], + "credit_map": [ + 0.002993973670527339, + 0.0018990543903782964, + 0.005875257309526205, + 0.004702326841652393, + 0.0025060337502509356, + 0.0027045314200222492, + 0.002535154577344656, + 0.0035629081539809704, + 0.0021919328719377518, + 0.001121402601711452, + 0.0034045574720948935, + 0.0019104069797322154, + 0.0027838232927024364, + 0.0024342420510947704, + 0.0037311995401978493, + 0.0023648166097700596, + 0.0019222270930185914, + 0.0011380708310753107, + 0.002586107002571225, + 0.001374083454720676, + 0.0010318298591300845, + 0.0011764428345486522, + 0.0012745310086756945, + 0.001252865302376449, + 0.003632614854723215, + 0.004293173085898161, + 0.0064723980613052845, + 0.00440997863188386, + 0.004328773822635412, + 0.008069979958236217, + 0.0035364022478461266, + 0.00386595050804317, + 0.0020759773906320333, + 0.002812744816765189, + 0.0017429784638807178, + 0.0023381090722978115, + 0.0023973837960511446, + 0.00235129427164793, + 0.0017568061593919992, + 0.002443749690428376, + 0.010423856787383556, + 0.005357109941542149, + 0.008068150840699673, + 0.0046773203648626804, + 0.0056511713191866875, + 0.022672999650239944, + 0.004093164578080177, + 0.011379397474229336, + 0.01981285586953163, + 0.007433939725160599, + 0.04794206842780113, + 0.013019674457609653, + 0.012921137735247612, + 0.2720738351345062, + 0.017621565610170364, + 0.28119343519210815, + 0.019897770136594772, + 0.0029656768310815096, + 0.004451222252100706, + 0.0018238862976431847, + 0.0029000546783208847, + 0.023445026949048042, + 0.004644155036658049, + 0.005173097364604473, + 0.0027556028217077255, + 0.002133567351847887, + 0.0017529097385704517, + 0.001969056436792016, + 0.002046175766736269, + 0.002082519233226776, + 0.0019257832318544388, + 0.0028134773019701242, + 0.0061221313662827015, + 0.0050466409884393215, + 0.006734197027981281, + 0.006780408788472414, + 0.006536690052598715, + 0.007822226732969284, + 0.007579693105071783, + 0.011250248178839684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.018183909356594086 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700078159570694 + ], + "opposite_logprob_tokens": [ + -2.695007801055908 + ], + "logprob_reward": -0.0700078159570694, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700078159570694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03674043342471123, + 0.028877994045615196, + 0.021566500887274742, + 0.02809300646185875, + 0.03001268394291401, + 0.06112091988325119, + 0.7061963081359863, + 0.04641350731253624, + 0.016017619520425797, + 0.024961013346910477 + ], + "credit_map": [ + 0.0055318367667496204, + 0.0033310833387076855, + 0.005538030061870813, + 0.004772251471877098, + 0.0031293309293687344, + 0.004474462475627661, + 0.005255029071122408, + 0.004708411637693644, + 0.00603331346064806, + 0.002184319542720914, + 0.0028632746543735266, + 0.0024127427022904158, + 0.0017594534438103437, + 0.003042597323656082, + 0.0025294767692685127, + 0.008052815683186054, + 0.004256625659763813, + 0.0015077886637300253, + 0.006689004134386778, + 0.0018427979666739702, + 0.0016817317809909582, + 0.0016468774992972612, + 0.0023773780558258295, + 0.0015642992220818996, + 0.0034711456391960382, + 0.003302403725683689, + 0.004636720288544893, + 0.00362465251237154, + 0.003017500741407275, + 0.003599742194637656, + 0.003105604089796543, + 0.003335239365696907, + 0.0033315175678581, + 0.00400542700663209, + 0.005564433988183737, + 0.0036422263365238905, + 0.002900515915825963, + 0.003996141720563173, + 0.0024422109127044678, + 0.004130212590098381, + 0.00835195928812027, + 0.0051485830917954445, + 0.006338754668831825, + 0.0053749470971524715, + 0.0033272060099989176, + 0.019084135070443153, + 0.004141132812947035, + 0.00935420673340559, + 0.020844288170337677, + 0.00787363387644291, + 0.05043449252843857, + 0.013635287992656231, + 0.013642273843288422, + 0.28500381112098694, + 0.0186929889023304, + 0.29606956243515015, + 0.010729549452662468, + 0.002291696844622493, + 0.00387244182638824, + 0.003560114884749055, + 0.0021871249191462994, + 0.01598540134727955, + 0.004038872197270393, + 0.003748307004570961, + 0.0017619666177779436, + 0.0016856801230460405, + 0.0014788047410547733, + 0.0015649506822228432, + 0.0024158123414963484, + 0.0016229472821578383, + 0.003727884963154793, + 0.00175957300234586, + 0.00265416968613863, + 0.0019951940048485994, + 0.0030277392361313105, + 0.0024852913338690996, + 0.0027931733056902885, + 0.0029707662761211395, + 0.004253979306668043, + 0.004780702758580446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0700078159570694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705211877822876 + ], + "opposite_logprob_tokens": [ + -1.137052059173584 + ], + "logprob_reward": -0.38705211877822876, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.38705211877822876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2407270222902298, + 0.04669493809342384, + 0.04338670149445534, + 0.04205351322889328, + 0.0749417245388031, + 0.12841010093688965, + 0.32757726311683655, + 0.03991249203681946, + 0.030209673568606377, + 0.026086561381816864 + ], + "credit_map": [ + 0.0544755756855011, + 0.03709321841597557, + 0.042839955538511276, + 0.023486752063035965, + 0.012836390174925327, + 0.018563253805041313, + 0.01726505346596241, + 0.0341668464243412, + 0.004762179218232632, + 0.0034064892679452896, + 0.006596497725695372, + 0.00574677437543869, + 0.007173897232860327, + 0.006767780054360628, + 0.006515068002045155, + 0.005726251285523176, + 0.007116437889635563, + 0.0038467999547719955, + 0.007799537852406502, + 0.005027369130402803, + 0.0043588969856500626, + 0.003725546645000577, + 0.006668758578598499, + 0.004843356087803841, + 0.006000480614602566, + 0.00392635352909565, + 0.008886179886758327, + 0.003698247717693448, + 0.003625779179856181, + 0.006946476176381111, + 0.003719799220561981, + 0.005250198766589165, + 0.007126869633793831, + 0.012026824057102203, + 0.009510417468845844, + 0.010001988150179386, + 0.010756311938166618, + 0.008855901658535004, + 0.007681415416300297, + 0.008981995284557343, + 0.019017869606614113, + 0.00986616313457489, + 0.013259174302220345, + 0.008814914152026176, + 0.0102535430341959, + 0.04009707272052765, + 0.007605582941323519, + 0.019495775923132896, + 0.009621737524867058, + 0.004709295462816954, + 0.023149341344833374, + 0.007140252739191055, + 0.008007816970348358, + 0.13027624785900116, + 0.010246828198432922, + 0.13442572951316833, + 0.0028725862503051758, + 0.0021129127126187086, + 0.0032092779874801636, + 0.0016865413635969162, + 0.0026608637999743223, + 0.010966107249259949, + 0.003033299930393696, + 0.013370899483561516, + 0.0039841956458985806, + 0.0023702082689851522, + 0.0027051204815506935, + 0.002556353807449341, + 0.002766456687822938, + 0.0036838201340287924, + 0.004710079170763493, + 0.007433441933244467, + 0.0025239004753530025, + 0.003388399025425315, + 0.002365282503888011, + 0.002794459694996476, + 0.0037026298232376575, + 0.002907299902290106, + 0.003561032470315695, + 0.00484355678781867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.38705211877822876 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-speed camera, in a single, seamless shot, begins from a dynamic, low-angle upward sweep, closely following a massive queen bee, draped in an ornate golden carapace. Its massive bulk casts sharp shadows across the vast, futuristic amphitheater. A massive, circular spotlight overhead casts a harsh, cold white light, revealing every intricate pattern on its gleaming metallic carapace. Its enormous compound eyes gleam with a cold, calculating radiance, while its translucent wings blur in their rapid vibrations. As the camera steadily advances, the queen bee lunges forward, its massive, razor-sharp jaws closing with unstoppable force, precise and brutal, instantly biting down upon the head and thorax of its prey—a hideous, beetle-like creature. In extreme slow motion, time seems frozen as streams of emerald green bodily fluids burst and splash like pearls from the opponent's wounds, each drop gleaming eerily under the cold light. Simultaneously, the opponent's hard exoskeleton shatters, sending countless tiny fragments scattering in all directions, creating a stunning scene imbued with a violent aesthetic. The camera follows the largest fragment as it spins and flies, then quickly pulls back to focus on the queen bee, her mandibles retracted, standing proudly, her body stained with the enemy's bodily fluids. Against the blurred backdrop of an energy shield and the silhouettes of a frenzied crowd, this undeniable end to this fiercely competitive match is clearly announced.", + "seed": 20940, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_7/index_8595_seed_20940_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": -0.14559229612350463, + "vlm_logprob": -0.14559229612350463, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.14559229612350463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04274465888738632 + ], + "opposite_logprob_tokens": [ + -3.2927446365356445 + ], + "logprob_reward": -0.04274465888738632, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04274465888738632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10323148220777512, + 0.06544715911149979, + 0.17719176411628723, + 0.16823354363441467, + 0.08585470169782639, + 0.0940917432308197, + 0.08730854839086533, + 0.07846003025770187, + 0.08666538447141647, + 0.05351566523313522 + ], + "credit_map": [ + 0.01294676773250103, + 0.01531913597136736, + 0.012722176499664783, + 0.019407590851187706, + 0.009670918807387352, + 0.01057716179639101, + 0.011441604234278202, + 0.011146125383675098, + 0.006284005474299192, + 0.011434873566031456, + 0.008305339142680168, + 0.011499269865453243, + 0.006482435390353203, + 0.0043368833139538765, + 0.008435581810772419, + 0.00866877380758524, + 0.01735512726008892, + 0.02016269601881504, + 0.02509940043091774, + 0.02543681673705578, + 0.015980223193764687, + 0.022508487105369568, + 0.02469446137547493, + 0.025954561308026314, + 0.027249248698353767, + 0.01567525789141655, + 0.0216303002089262, + 0.01768120378255844, + 0.021227408200502396, + 0.02179640531539917, + 0.020606528967618942, + 0.022367186844348907, + 0.0074435509741306305, + 0.008854657411575317, + 0.008095640689134598, + 0.013283420354127884, + 0.015833985060453415, + 0.013412825763225555, + 0.010203748941421509, + 0.008726870641112328, + 0.012319939211010933, + 0.023505758494138718, + 0.006469354499131441, + 0.010591613128781319, + 0.009514685720205307, + 0.013347974978387356, + 0.011766643263399601, + 0.006575775798410177, + 0.007833638228476048, + 0.010388921946287155, + 0.010534246452152729, + 0.010563718155026436, + 0.009239437058568, + 0.011724988929927349, + 0.01381633710116148, + 0.013207264244556427, + 0.010492723435163498, + 0.007631246466189623, + 0.009100615046918392, + 0.006511244922876358, + 0.007867081090807915, + 0.01491591613739729, + 0.009265482425689697, + 0.012675722129642963, + 0.0072471085004508495, + 0.0157723780721426, + 0.008294407278299332, + 0.008343511261045933, + 0.006898446474224329, + 0.012971988879144192, + 0.010885098949074745, + 0.016252443194389343, + 0.005420723930001259, + 0.006693236529827118, + 0.007455240003764629, + 0.00636658538132906, + 0.008383159525692463, + 0.006926221773028374, + 0.004290990065783262, + 0.00797950942069292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.04274465888738632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903401345014572 + ], + "opposite_logprob_tokens": [ + -2.464034080505371 + ], + "logprob_reward": -0.08903401345014572, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903401345014572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11180014908313751, + 0.0764666423201561, + 0.3113712966442108, + 0.05368821322917938, + 0.11835341900587082, + 0.08379918336868286, + 0.06590862572193146, + 0.12539884448051453, + 0.026343993842601776, + 0.026869596913456917 + ], + "credit_map": [ + 0.017440496012568474, + 0.011317783035337925, + 0.020348358899354935, + 0.01784122735261917, + 0.007546440232545137, + 0.009585137479007244, + 0.010981457307934761, + 0.016739245504140854, + 0.007179480977356434, + 0.01127493567764759, + 0.011878318153321743, + 0.010212084278464317, + 0.007708784658461809, + 0.0058444500900805, + 0.009933797642588615, + 0.012434794567525387, + 0.0372588224709034, + 0.01727926917374134, + 0.04505041614174843, + 0.0325172133743763, + 0.02654605358839035, + 0.05988861992955208, + 0.03810010850429535, + 0.05473080649971962, + 0.003180132480338216, + 0.0036003151908516884, + 0.0034227019641548395, + 0.0034297960810363293, + 0.01151742972433567, + 0.008856423199176788, + 0.007389230187982321, + 0.012292183935642242, + 0.008873979561030865, + 0.009180781431496143, + 0.008968661539256573, + 0.012173112481832504, + 0.029435081407427788, + 0.02059095725417137, + 0.016119549050927162, + 0.013011297211050987, + 0.011246263980865479, + 0.021712305024266243, + 0.0072503709234297276, + 0.010755636729300022, + 0.007942145690321922, + 0.010024173185229301, + 0.008669056929647923, + 0.006199232302606106, + 0.004586981609463692, + 0.006457002833485603, + 0.007796792779117823, + 0.009410164318978786, + 0.007177034392952919, + 0.008568434044718742, + 0.012553741224110126, + 0.009358473122119904, + 0.012406058609485626, + 0.018428318202495575, + 0.013737168163061142, + 0.011932183988392353, + 0.023433558642864227, + 0.01759783737361431, + 0.014194735325872898, + 0.013668999075889587, + 0.00207750522531569, + 0.004056862555444241, + 0.0025699047837406397, + 0.002436971291899681, + 0.0024228948168456554, + 0.003753413911908865, + 0.0037825533654540777, + 0.005243887659162283, + 0.002495116088539362, + 0.0027265569660812616, + 0.0029519323725253344, + 0.003352248342707753, + 0.0022595140617340803, + 0.00418453011661768, + 0.0038062026724219322, + 0.005093495361506939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.08903401345014572 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979971095919609 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979971095919609, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979971095919609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05511011555790901, + 0.08046984672546387, + 0.07051951438188553, + 0.04359359294176102, + 0.3675200045108795, + 0.08449499309062958, + 0.09448160976171494, + 0.08496937900781631, + 0.05610052868723869, + 0.06274037063121796 + ], + "credit_map": [ + 0.005492045544087887, + 0.006292372010648251, + 0.007250962313264608, + 0.010372884571552277, + 0.005409577861428261, + 0.005340240430086851, + 0.008054130710661411, + 0.006897902116179466, + 0.007656547240912914, + 0.007945251651108265, + 0.012284147553145885, + 0.012205176055431366, + 0.006824191194027662, + 0.0051829032599925995, + 0.010052562691271305, + 0.01831907220184803, + 0.006007534451782703, + 0.006372188217937946, + 0.014040211215615273, + 0.008911038748919964, + 0.005278483498841524, + 0.008198868483304977, + 0.009726282209157944, + 0.011984902434051037, + 0.0045400517992675304, + 0.0035644620656967163, + 0.004701925907284021, + 0.003513449802994728, + 0.007038802374154329, + 0.0062439944595098495, + 0.004357225261628628, + 0.009633679874241352, + 0.021658139303326607, + 0.029201645404100418, + 0.030655084177851677, + 0.029193691909313202, + 0.1071561798453331, + 0.07482264190912247, + 0.043036092072725296, + 0.03179655969142914, + 0.005455540493130684, + 0.007988357916474342, + 0.005839281249791384, + 0.008778011426329613, + 0.012229173444211483, + 0.018759405240416527, + 0.0169784314930439, + 0.008466796018183231, + 0.00724114291369915, + 0.007566799875348806, + 0.009363455697894096, + 0.01088155061006546, + 0.00770950224250555, + 0.012956203892827034, + 0.021051904186606407, + 0.01771104522049427, + 0.008965219371020794, + 0.009250492788851261, + 0.012068833224475384, + 0.008982386440038681, + 0.012295068241655827, + 0.01237251702696085, + 0.0109190559014678, + 0.01011581253260374, + 0.005185638554394245, + 0.008067275397479534, + 0.00470525398850441, + 0.003944001160562038, + 0.004889984615147114, + 0.00878361240029335, + 0.007258906029164791, + 0.013265860266983509, + 0.00636517908424139, + 0.005677349865436554, + 0.00574317155405879, + 0.006677711848169565, + 0.007943094708025455, + 0.009561875835061073, + 0.011268429458141327, + 0.009503562934696674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.02979971095919609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306123033165932 + ], + "opposite_logprob_tokens": [ + -3.1680612564086914 + ], + "logprob_reward": -0.04306123033165932, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306123033165932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08693484216928482, + 0.0767747089266777, + 0.11844348162412643, + 0.11971684545278549, + 0.10140300542116165, + 0.06187981739640236, + 0.16762632131576538, + 0.16865073144435883, + 0.029180729761719704, + 0.06938954442739487 + ], + "credit_map": [ + 0.01064041256904602, + 0.011628037318587303, + 0.010612023063004017, + 0.017577871680259705, + 0.00865088403224945, + 0.008698178455233574, + 0.009800182655453682, + 0.009327252395451069, + 0.0062343361787498, + 0.012273556552827358, + 0.009118719026446342, + 0.014920398592948914, + 0.00660340441390872, + 0.004521168768405914, + 0.011665960773825645, + 0.01143716648221016, + 0.00760000990703702, + 0.009473830461502075, + 0.03387930616736412, + 0.010348450392484665, + 0.004308472853153944, + 0.008953159675002098, + 0.009611089713871479, + 0.03426916524767876, + 0.016201630234718323, + 0.010284967720508575, + 0.012368748895823956, + 0.00907803513109684, + 0.02197277918457985, + 0.01588835008442402, + 0.013163981027901173, + 0.020758356899023056, + 0.008152300491929054, + 0.006857545580714941, + 0.008053281344473362, + 0.006971060764044523, + 0.026547417044639587, + 0.018452994525432587, + 0.014287896454334259, + 0.012080512940883636, + 0.005379043985158205, + 0.009247236885130405, + 0.009253602474927902, + 0.00677309138700366, + 0.010365469381213188, + 0.008797110058367252, + 0.006991622969508171, + 0.005072639789432287, + 0.014288683421909809, + 0.01910381205379963, + 0.019533254206180573, + 0.026493221521377563, + 0.015641752630472183, + 0.025970997288823128, + 0.02604176662862301, + 0.02055283822119236, + 0.013566093519330025, + 0.022606024518609047, + 0.017100106924772263, + 0.017005659639835358, + 0.029386932030320168, + 0.023814503103494644, + 0.023474296554923058, + 0.021697107702493668, + 0.0031420201994478703, + 0.00564565509557724, + 0.0029489363078027964, + 0.0032656879629939795, + 0.0030061956495046616, + 0.0043070195242762566, + 0.003372235456481576, + 0.0034929784014821053, + 0.007666414603590965, + 0.008334876969456673, + 0.010839966125786304, + 0.008169976994395256, + 0.010820044204592705, + 0.008767710067331791, + 0.008451166562736034, + 0.006339391693472862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.04306123033165932 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return 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 queen bee with an ornate golden carapace?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the scene take place in a vast futuristic amphitheater with a massive circular overhead spotlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen bee lunge forward and bite down on a hideous beetle-like creature?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are emerald green bodily fluids shown bursting from the opponent's wounds in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a low-angle upward sweep and later pull back to focus on the queen bee?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233218669891357 + ], + "opposite_logprob_tokens": [ + -0.8983218669891357 + ], + "logprob_reward": -0.5233218669891357, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233218669891357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11131603270769119, + 0.1371929794549942, + 0.08894053101539612, + 0.17243769764900208, + 0.14014798402786255, + 0.1365758180618286, + 0.0923268124461174, + 0.08585208654403687, + 0.013579132035374641, + 0.02163085713982582 + ], + "credit_map": [ + 0.013911811634898186, + 0.015897776931524277, + 0.015782784670591354, + 0.01914231665432453, + 0.010394846089184284, + 0.0104912631213665, + 0.015394927933812141, + 0.010300315916538239, + 0.011226429603993893, + 0.010573502629995346, + 0.020634083077311516, + 0.027323756366968155, + 0.012872039340436459, + 0.008794983848929405, + 0.019337452948093414, + 0.02643074095249176, + 0.006847306154668331, + 0.011514098383486271, + 0.016182702034711838, + 0.011243771761655807, + 0.005833163857460022, + 0.008028093725442886, + 0.010897791013121605, + 0.018393613398075104, + 0.017540697008371353, + 0.01535121351480484, + 0.013758443295955658, + 0.01590277999639511, + 0.035176269710063934, + 0.023326169699430466, + 0.019706597551703453, + 0.0316755510866642, + 0.010222368873655796, + 0.008306085132062435, + 0.02019057795405388, + 0.011672849766910076, + 0.03388366475701332, + 0.02459917962551117, + 0.019685208797454834, + 0.01158806774765253, + 0.01814635656774044, + 0.036001820117235184, + 0.010157864540815353, + 0.01418273150920868, + 0.013994391076266766, + 0.020810186862945557, + 0.014313939958810806, + 0.008968541398644447, + 0.008176061324775219, + 0.013859800063073635, + 0.01118061225861311, + 0.010823152959346771, + 0.00876571610569954, + 0.018162501975893974, + 0.010870384983718395, + 0.010488592088222504, + 0.00789121724665165, + 0.01350946445018053, + 0.010246078483760357, + 0.008805373683571815, + 0.016211092472076416, + 0.01146242581307888, + 0.009656824171543121, + 0.008069616742432117, + 0.0012327143922448158, + 0.0018173935823142529, + 0.0018272585002705455, + 0.0020228982903063297, + 0.0013725647004321218, + 0.0018819279503077269, + 0.001433192053809762, + 0.001991183264181018, + 0.002555491868406534, + 0.002178905066102743, + 0.003698660060763359, + 0.0024801543913781643, + 0.0020207976922392845, + 0.0028910471592098475, + 0.002102176658809185, + 0.003703626338392496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.5233218669891357 + } + ], + "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/0000160/group_9/index_tensor([3568, 3568])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_tensor([3568, 3568])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..11095db30ade9a57c7308ac31f98083bfc87511b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_tensor([3568, 3568])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An underwater close-up, captured with a clean, cartoon-style high-frame-rate camera, presents a serene and captivating scene. The static camera focuses on a single, graceful stem of Mayaka fluviatilis rising from a soft, sandy lakebed. Its delicate, mossy leaves, rendered in vibrant shades of lime and emerald green with simple, clean outlines, gently sway in an invisible current. The plant's leaves exhibit an incredibly smooth, hypnotic undulation, rhythmically expanding and contracting beneath the glassy, calm surface of the water. Golden sunlight penetrates the crystal-clear water from above, casting ever-changing patterns of light and shadow (caustics) that dance across the plant and the lakebed. Light catches the edges of the soft leaves, making them appear to softly glow. The overall atmosphere is one of profound peace and tranquility, the plant's graceful, repetitive movement creating a deeply meditative and visually captivating experience.", + "seed": 21658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_3568_seed_21658_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": -0.3362687833607197, + "vlm_logprob": -0.3362687833607197, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.3362687833607197, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.605150580406189 + ], + "opposite_logprob_tokens": [ + -0.23015058040618896 + ], + "logprob_reward": -1.605150580406189, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.605150580406189, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3928975760936737, + 0.19424991309642792, + 0.06774603575468063, + 0.060553599148988724, + 0.06859856098890305, + 0.04671890288591385, + 0.04293053597211838, + 0.02362423576414585, + 0.07814060151576996, + 0.024540124461054802 + ], + "credit_map": [ + 0.06992227584123611, + 0.05456862971186638, + 0.06801185011863708, + 0.03927359730005264, + 0.04517601430416107, + 0.04470458999276161, + 0.027086082845926285, + 0.044154513627290726, + 0.030793044716119766, + 0.015676455572247505, + 0.03793766349554062, + 0.013934159651398659, + 0.015100141055881977, + 0.03931857645511627, + 0.010483031161129475, + 0.031006839126348495, + 0.01017800159752369, + 0.011176824569702148, + 0.01034288015216589, + 0.005930996499955654, + 0.007674257270991802, + 0.006322487723082304, + 0.004002118483185768, + 0.012118463404476643, + 0.006388086825609207, + 0.0077200839295983315, + 0.010926838964223862, + 0.006588201504200697, + 0.006494040135294199, + 0.004642431624233723, + 0.006628415547311306, + 0.011165499687194824, + 0.009072591550648212, + 0.007053831126540899, + 0.009366234764456749, + 0.011882907710969448, + 0.006594256963580847, + 0.006346988026052713, + 0.004629977513104677, + 0.013651768676936626, + 0.005189421121031046, + 0.006698021199554205, + 0.007713912986218929, + 0.006023557391017675, + 0.004087856970727444, + 0.004617517814040184, + 0.005189649295061827, + 0.007198967970907688, + 0.008112682029604912, + 0.004703615326434374, + 0.0059365760535001755, + 0.004085122607648373, + 0.0036883587017655373, + 0.007117223460227251, + 0.003246360458433628, + 0.006040598731487989, + 0.0028222715482115746, + 0.0025942011270672083, + 0.003780598985031247, + 0.0024299451615661383, + 0.0026713376864790916, + 0.002988143591210246, + 0.002510637044906616, + 0.003827099921181798, + 0.011399625800549984, + 0.007725704926997423, + 0.012388274073600769, + 0.010317866690456867, + 0.005850786343216896, + 0.009783531539142132, + 0.006013394333422184, + 0.014661422930657864, + 0.0031158411875367165, + 0.0021869377233088017, + 0.004023998510092497, + 0.002731277607381344, + 0.0024212158750742674, + 0.00311138853430748, + 0.0017962770070880651, + 0.005153187084943056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -1.605150580406189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00284576416015625 + ], + "opposite_logprob_tokens": [ + -5.877845764160156 + ], + "logprob_reward": -0.00284576416015625, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00284576416015625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3705666959285736, + 0.12290717661380768, + 0.07257398217916489, + 0.05242258310317993, + 0.10916344076395035, + 0.05434906482696533, + 0.048601843416690826, + 0.055644720792770386, + 0.0836925357580185, + 0.030077965930104256 + ], + "credit_map": [ + 0.0566055066883564, + 0.055558543652296066, + 0.05306624993681908, + 0.04089353233575821, + 0.042639292776584625, + 0.04389914497733116, + 0.030626408755779266, + 0.047278039157390594, + 0.01542902085930109, + 0.011238893494009972, + 0.024172449484467506, + 0.010142823681235313, + 0.011096770875155926, + 0.020204240456223488, + 0.008965727873146534, + 0.02165726199746132, + 0.008628174662590027, + 0.011120803654193878, + 0.009297240525484085, + 0.007701305206865072, + 0.014985494315624237, + 0.006903527304530144, + 0.004425270948559046, + 0.009512163698673248, + 0.008117050863802433, + 0.006091514136642218, + 0.010406777262687683, + 0.005885313265025616, + 0.005315367132425308, + 0.0041885655373334885, + 0.005194299854338169, + 0.007223693188279867, + 0.013593981973826885, + 0.009675802662968636, + 0.015621447004377842, + 0.018263548612594604, + 0.008477624505758286, + 0.014273660257458687, + 0.007890256121754646, + 0.021367112174630165, + 0.008783368393778801, + 0.005077183246612549, + 0.007818684913218021, + 0.004514340776950121, + 0.00457551097497344, + 0.009467346593737602, + 0.004704577382653952, + 0.009408052079379559, + 0.00777825340628624, + 0.005039020907133818, + 0.007079140283167362, + 0.005912741646170616, + 0.005360892042517662, + 0.005972480401396751, + 0.0045636738650500774, + 0.006895637605339289, + 0.012230044230818748, + 0.00421471381559968, + 0.010303928516805172, + 0.004243651404976845, + 0.003961987793445587, + 0.009722265414893627, + 0.003430539509281516, + 0.007537590339779854, + 0.012397975660860538, + 0.008500426076352596, + 0.012998506426811218, + 0.010010399855673313, + 0.0069432551972568035, + 0.011171797290444374, + 0.006599135231226683, + 0.015071039088070393, + 0.0037133414298295975, + 0.0030932279769331217, + 0.0045714969746768475, + 0.0027961793821305037, + 0.003694833256304264, + 0.0037378063425421715, + 0.0029548597522079945, + 0.0055162180215120316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00284576416015625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37032702565193176, + 0.1337655633687973, + 0.04966020956635475, + 0.03244001418352127, + 0.13664501905441284, + 0.12318282574415207, + 0.04542740806937218, + 0.050571274012327194, + 0.03467622771859169, + 0.023304427042603493 + ], + "credit_map": [ + 0.07474245131015778, + 0.03476179763674736, + 0.09911742061376572, + 0.03251808136701584, + 0.02755650132894516, + 0.03162401169538498, + 0.023659542202949524, + 0.04634721577167511, + 0.023284509778022766, + 0.009864543564617634, + 0.027166100218892097, + 0.010315770283341408, + 0.009021659381687641, + 0.027379661798477173, + 0.007601443212479353, + 0.01913187839090824, + 0.006217703688889742, + 0.006540979724377394, + 0.006968309637159109, + 0.00398845924064517, + 0.010286551900207996, + 0.005223165266215801, + 0.004233055282384157, + 0.006201986223459244, + 0.0032298255246132612, + 0.004375459160655737, + 0.004600889049470425, + 0.003934849519282579, + 0.004379668738692999, + 0.004198119509965181, + 0.0026906656567007303, + 0.005030535627156496, + 0.01717662438750267, + 0.012314713560044765, + 0.023337088525295258, + 0.019350608810782433, + 0.010134796611964703, + 0.02025427110493183, + 0.008903106674551964, + 0.02517380192875862, + 0.017870595678687096, + 0.011048858053982258, + 0.016900550574064255, + 0.01200169138610363, + 0.008840801194310188, + 0.01956600323319435, + 0.013006564229726791, + 0.023947766050696373, + 0.007046374026685953, + 0.004034298937767744, + 0.006881724111735821, + 0.006012244615703821, + 0.005190116818994284, + 0.004960463382303715, + 0.0041552078910171986, + 0.007146978750824928, + 0.00673903850838542, + 0.0049308533780276775, + 0.008443345315754414, + 0.006628979463130236, + 0.0031362995505332947, + 0.00677069928497076, + 0.0037592824082821608, + 0.010162774473428726, + 0.004713152535259724, + 0.0032901782542467117, + 0.005395007785409689, + 0.004634333308786154, + 0.0025723050348460674, + 0.0044965920969843864, + 0.0026443658862262964, + 0.006930291652679443, + 0.0037676971405744553, + 0.0025417518336325884, + 0.0035302513279020786, + 0.0022332100197672844, + 0.0020773555152118206, + 0.0028110183775424957, + 0.0020128225442022085, + 0.004330319818109274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486447848379612 + ], + "opposite_logprob_tokens": [ + -3.048644781112671 + ], + "logprob_reward": -0.0486447848379612, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486447848379612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22451788187026978, + 0.14457391202449799, + 0.06582094728946686, + 0.07248564809560776, + 0.20896509289741516, + 0.15800821781158447, + 0.04275558516383171, + 0.03664585202932358, + 0.025591500103473663, + 0.02063538320362568 + ], + "credit_map": [ + 0.04188627749681473, + 0.0331382192671299, + 0.03741416707634926, + 0.0255590733140707, + 0.02209562063217163, + 0.02213960513472557, + 0.01451632846146822, + 0.027768582105636597, + 0.021349672228097916, + 0.013057705946266651, + 0.028309732675552368, + 0.011470910161733627, + 0.013416235335171223, + 0.022203003987669945, + 0.009705385193228722, + 0.025061262771487236, + 0.007333442103117704, + 0.008734249509871006, + 0.010838741436600685, + 0.0062739052809774876, + 0.008782900869846344, + 0.0072814952582120895, + 0.004670318216085434, + 0.011905894614756107, + 0.010946798138320446, + 0.00820230133831501, + 0.010556979104876518, + 0.007631456945091486, + 0.007328487932682037, + 0.006603449583053589, + 0.006432975176721811, + 0.014783195219933987, + 0.028029561042785645, + 0.023206334561109543, + 0.0347432903945446, + 0.01971055381000042, + 0.02116776816546917, + 0.03207547590136528, + 0.020140133798122406, + 0.0298919677734375, + 0.019338427111506462, + 0.015214625746011734, + 0.024371717125177383, + 0.01995714008808136, + 0.017088092863559723, + 0.018858497962355614, + 0.017810611054301262, + 0.025369111448526382, + 0.005688543897122145, + 0.003547623986378312, + 0.007082080002874136, + 0.004326415713876486, + 0.006309136748313904, + 0.003351924940943718, + 0.0037696787621825933, + 0.00868018064647913, + 0.007800503168255091, + 0.0031637842766940594, + 0.005484006367623806, + 0.0024046830367296934, + 0.0035587174352258444, + 0.005906833801418543, + 0.0032263475004583597, + 0.0051009757444262505, + 0.0035626543685793877, + 0.002502736169844866, + 0.004012753255665302, + 0.0028290064074099064, + 0.002809774363413453, + 0.0031912317499518394, + 0.0023707274813205004, + 0.004312616307288408, + 0.0029731544200330973, + 0.003082975512370467, + 0.0026204942259937525, + 0.0018104520859196782, + 0.002757387002930045, + 0.002746852580457926, + 0.002024199580773711, + 0.0026198679115623236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0486447848379612 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062501817941666 + ], + "opposite_logprob_tokens": [ + -4.1410627365112305 + ], + "logprob_reward": -0.016062501817941666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062501817941666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19570417702198029, + 0.219746395945549, + 0.07338261604309082, + 0.053449470549821854, + 0.15491876006126404, + 0.12136685103178024, + 0.06954705715179443, + 0.05736641585826874, + 0.03196315839886665, + 0.022555062547326088 + ], + "credit_map": [ + 0.029210718348622322, + 0.027116091921925545, + 0.023434266448020935, + 0.024795232340693474, + 0.019082922488451004, + 0.026805374771356583, + 0.021376999095082283, + 0.023882588371634483, + 0.03663413226604462, + 0.020414583384990692, + 0.04423430934548378, + 0.016636459156870842, + 0.01707574352622032, + 0.04244517907500267, + 0.013461941853165627, + 0.028844039887189865, + 0.00756137166172266, + 0.010354312136769295, + 0.010691330768167973, + 0.010342423804104328, + 0.010840406641364098, + 0.007524559274315834, + 0.005440339911729097, + 0.010627870447933674, + 0.007792949676513672, + 0.0055497013963758945, + 0.009983799420297146, + 0.005024806130677462, + 0.003528841771185398, + 0.003267598571255803, + 0.003260093741118908, + 0.015041681006550789, + 0.017193473875522614, + 0.01318858377635479, + 0.02536877989768982, + 0.025446508079767227, + 0.010988038033246994, + 0.02103634923696518, + 0.00881816353648901, + 0.03287887945771217, + 0.019233092665672302, + 0.011785801500082016, + 0.016786517575383186, + 0.012920152395963669, + 0.009936181828379631, + 0.015944937244057655, + 0.00981499720364809, + 0.024945177137851715, + 0.00984441488981247, + 0.00954409223049879, + 0.009262517094612122, + 0.007268035784363747, + 0.009460887871682644, + 0.007728756871074438, + 0.006242250092327595, + 0.010196099057793617, + 0.007364120800048113, + 0.005152567755430937, + 0.010686401277780533, + 0.008268119767308235, + 0.00447540357708931, + 0.0076932343654334545, + 0.003470102557912469, + 0.010256467387080193, + 0.003781383391469717, + 0.0032619598787277937, + 0.004699463024735451, + 0.0037621434312313795, + 0.0035851977299898863, + 0.0046066236682236195, + 0.0023644750472158194, + 0.005901911295950413, + 0.002537090564146638, + 0.0025322691071778536, + 0.0035166400484740734, + 0.0022170771844685078, + 0.0023962296545505524, + 0.002811219310387969, + 0.0022520592901855707, + 0.0042924778535962105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.016062501817941666 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An underwater close-up, captured with a clean, cartoon-style high-frame-rate camera, presents a serene and captivating scene. The static camera focuses on a single, graceful stem of Mayaka fluviatilis rising from a soft, sandy lakebed. Its delicate, mossy leaves, rendered in vibrant shades of lime and emerald green with simple, clean outlines, gently sway in an invisible current. The plant's leaves exhibit an incredibly smooth, hypnotic undulation, rhythmically expanding and contracting beneath the glassy, calm surface of the water. Golden sunlight penetrates the crystal-clear water from above, casting ever-changing patterns of light and shadow (caustics) that dance across the plant and the lakebed. Light catches the edges of the soft leaves, making them appear to softly glow. The overall atmosphere is one of profound peace and tranquility, the plant's graceful, repetitive movement creating a deeply meditative and visually captivating experience.", + "seed": 21659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_3568_seed_21659_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": -0.01135333760175854, + "vlm_logprob": -0.01135333760175854, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.01135333760175854, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009360946714878082 + ], + "opposite_logprob_tokens": [ + -5.5093607902526855 + ], + "logprob_reward": -0.009360946714878082, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009360946714878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5408471822738647, + 0.056290220469236374, + 0.018016060814261436, + 0.03819727152585983, + 0.0221078060567379, + 0.18861331045627594, + 0.029572563245892525, + 0.04507758095860481, + 0.031068377196788788, + 0.030209625139832497 + ], + "credit_map": [ + 0.0434013232588768, + 0.024302683770656586, + 0.1912773996591568, + 0.02673608437180519, + 0.01779407449066639, + 0.03393871709704399, + 0.01941624842584133, + 0.18398061394691467, + 0.01112538855522871, + 0.0050270152278244495, + 0.012253208085894585, + 0.005582666955888271, + 0.003830696688964963, + 0.004720235243439674, + 0.0039286320097744465, + 0.009822377935051918, + 0.001963710179552436, + 0.0012132131960242987, + 0.0034828102216124535, + 0.0016456377925351262, + 0.0016856624279171228, + 0.0024556221906095743, + 0.0019296943210065365, + 0.0036397078074514866, + 0.006857289932668209, + 0.002801349852234125, + 0.00811079703271389, + 0.0024250890128314495, + 0.0024362229742109776, + 0.0033337220083922148, + 0.0019794353283941746, + 0.010253362357616425, + 0.004951638635247946, + 0.0017904681153595448, + 0.004158905241638422, + 0.0012234393507242203, + 0.0015478559071198106, + 0.003365175798535347, + 0.0012200231431052089, + 0.0038502984680235386, + 0.012798046693205833, + 0.0044035036116838455, + 0.07628656923770905, + 0.005439393687993288, + 0.002697058254852891, + 0.007032929919660091, + 0.002930133370682597, + 0.07702566683292389, + 0.0045325616374611855, + 0.0018963880138471723, + 0.0073560248129069805, + 0.001802107086405158, + 0.001954111736267805, + 0.0028226561844348907, + 0.001632987055927515, + 0.007575727067887783, + 0.0054137022234499454, + 0.004356276709586382, + 0.0067877438850700855, + 0.004437628667801619, + 0.00483033386990428, + 0.005939967464655638, + 0.004937406629323959, + 0.00837452057749033, + 0.007368593011051416, + 0.0025943282525986433, + 0.005701813846826553, + 0.0025349354837089777, + 0.002520371926948428, + 0.0029894865583628416, + 0.002046318491920829, + 0.005312525667250156, + 0.0038338357117027044, + 0.00304546020925045, + 0.005396803840994835, + 0.0019262036075815558, + 0.002256168285384774, + 0.0036300248466432095, + 0.002498374553397298, + 0.007622754201292992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.009360946714878082 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017215682892128825 + ], + "opposite_logprob_tokens": [ + -6.376721382141113 + ], + "logprob_reward": -0.0017215682892128825, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017215682892128825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4732317626476288, + 0.05034809187054634, + 0.015325766988098621, + 0.13363619148731232, + 0.01610645279288292, + 0.11141201108694077, + 0.05417995899915695, + 0.08222262561321259, + 0.02586377039551735, + 0.037673406302928925 + ], + "credit_map": [ + 0.03400520235300064, + 0.016497230157256126, + 0.1794246882200241, + 0.017599591985344887, + 0.012599621899425983, + 0.024699704721570015, + 0.012556470930576324, + 0.17584925889968872, + 0.010021586902439594, + 0.0038932415191084146, + 0.010866071097552776, + 0.0039956527762115, + 0.004095751792192459, + 0.004243090748786926, + 0.003895531641319394, + 0.00933716632425785, + 0.002666245447471738, + 0.0013014262076467276, + 0.0029544474091380835, + 0.0011584601597860456, + 0.0016171516617760062, + 0.0017135917441919446, + 0.001218595658428967, + 0.0026958498638123274, + 0.0077321212738752365, + 0.0028356232214719057, + 0.05590325593948364, + 0.004938940051943064, + 0.0014300517505034804, + 0.0040289657190442085, + 0.0027421677950769663, + 0.054025061428546906, + 0.002421322278678417, + 0.0015263273380696774, + 0.0029719718731939793, + 0.001303369994275272, + 0.0016792919486761093, + 0.001960797468200326, + 0.001485457643866539, + 0.0027579141315072775, + 0.007908374071121216, + 0.0033645739313215017, + 0.0440838448703289, + 0.003201210405677557, + 0.0023386457469314337, + 0.0038727379869669676, + 0.002168772043660283, + 0.04447385296225548, + 0.0059727635234594345, + 0.0017582423752173781, + 0.019242078065872192, + 0.0021387559827417135, + 0.0013597210636362433, + 0.0026992664206773043, + 0.0014667997602373362, + 0.019542334601283073, + 0.011303170584142208, + 0.006616382393985987, + 0.014648211188614368, + 0.007318759802728891, + 0.006733810529112816, + 0.009739775210618973, + 0.005933294538408518, + 0.01992921717464924, + 0.004915252793580294, + 0.0015891111688688397, + 0.004667617846280336, + 0.0026887794956564903, + 0.0014125083107501268, + 0.0023311597760766745, + 0.0019035880686715245, + 0.006355753634124994, + 0.008880233392119408, + 0.00328924716450274, + 0.0059983693063259125, + 0.002808645134791732, + 0.0030686953105032444, + 0.0034313397482037544, + 0.0027486365288496017, + 0.007448241114616394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0017215682892128825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001179118873551488 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001179118873551488, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001179118873551488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4383196532726288, + 0.04250497743487358, + 0.018427973613142967, + 0.05528881028294563, + 0.010879374109208584, + 0.2204563170671463, + 0.02421538718044758, + 0.04288117215037346, + 0.11867208778858185, + 0.028354261070489883 + ], + "credit_map": [ + 0.03339599072933197, + 0.01808815822005272, + 0.16130945086479187, + 0.021296411752700806, + 0.010483849793672562, + 0.021939855068922043, + 0.014574170112609863, + 0.15723174810409546, + 0.006499983835965395, + 0.003841073252260685, + 0.00789150595664978, + 0.006312806159257889, + 0.0023545396979898214, + 0.004096446093171835, + 0.004313160665333271, + 0.0071954610757529736, + 0.004076785873621702, + 0.0012984734494239092, + 0.004578302148729563, + 0.0012866401812061667, + 0.0011991315986961126, + 0.0018214236479252577, + 0.0012226339895278215, + 0.0029445826075971127, + 0.005560578778386116, + 0.0016952009173110127, + 0.021934598684310913, + 0.0020511182956397533, + 0.0010475802700966597, + 0.0019226730801165104, + 0.0013679154217243195, + 0.0197091456502676, + 0.0020887074060738087, + 0.0007480417261831462, + 0.0023072839248925447, + 0.0007058414630591869, + 0.001055861241184175, + 0.0012162859784439206, + 0.000579702842514962, + 0.002177649876102805, + 0.014731083996593952, + 0.005194628611207008, + 0.09085578471422195, + 0.005867818370461464, + 0.0019181460374966264, + 0.006546640768647194, + 0.0031694755889475346, + 0.09217274188995361, + 0.004235449247062206, + 0.0011648369254544377, + 0.00672317435964942, + 0.0012356393272057176, + 0.0008969682385213673, + 0.001596603775396943, + 0.0008784017991274595, + 0.007484314031898975, + 0.006613573059439659, + 0.003216751618310809, + 0.006952537223696709, + 0.0036851121112704277, + 0.003069542581215501, + 0.004441032186150551, + 0.002840953879058361, + 0.012061670422554016, + 0.02407795563340187, + 0.007879265584051609, + 0.022714437916874886, + 0.007260930258780718, + 0.005861205980181694, + 0.01167808286845684, + 0.006386488676071167, + 0.03281370922923088, + 0.005125811789184809, + 0.0019420352764427662, + 0.005488143768161535, + 0.002293481957167387, + 0.0019482464995235205, + 0.003153749741613865, + 0.002014709869399667, + 0.006388083565980196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.001179118873551488 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026331797242164612 + ], + "opposite_logprob_tokens": [ + -3.651331901550293 + ], + "logprob_reward": -0.026331797242164612, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026331797242164612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14736837148666382, + 0.021243935450911522, + 0.01430900115519762, + 0.1836237907409668, + 0.024380985647439957, + 0.44976383447647095, + 0.017793186008930206, + 0.0946008712053299, + 0.028875628486275673, + 0.018040355294942856 + ], + "credit_map": [ + 0.012399634346365929, + 0.010020948946475983, + 0.045193638652563095, + 0.010632157325744629, + 0.006719296332448721, + 0.012860290706157684, + 0.008310769684612751, + 0.04123165085911751, + 0.002822469687089324, + 0.001981585519388318, + 0.00528716342523694, + 0.0019241739064455032, + 0.001777808298356831, + 0.00235234503634274, + 0.0018424392910674214, + 0.0032559500541538, + 0.0021147432271391153, + 0.0011948285391554236, + 0.0026774301659315825, + 0.0016172919422388077, + 0.0013569273287430406, + 0.0015650616260245442, + 0.0015128391096368432, + 0.0022698782850056887, + 0.013662578538060188, + 0.0038779915776103735, + 0.0777801126241684, + 0.006215324159711599, + 0.00179110502358526, + 0.005151831544935703, + 0.0033744503743946552, + 0.07177039980888367, + 0.0044890278950333595, + 0.002713254652917385, + 0.004628532100468874, + 0.001948068616911769, + 0.002424669684842229, + 0.002320623490959406, + 0.00216550100594759, + 0.0036913082003593445, + 0.029226403683423996, + 0.009711732156574726, + 0.18689323961734772, + 0.011947532184422016, + 0.003823364619165659, + 0.013576623983681202, + 0.006055500824004412, + 0.18852944672107697, + 0.0029455069452524185, + 0.0015110603999346495, + 0.003915346693247557, + 0.0013406573561951518, + 0.001361002097837627, + 0.001604414894245565, + 0.001097112544812262, + 0.004018085077404976, + 0.013289287686347961, + 0.008157876320183277, + 0.015120355412364006, + 0.008045879192650318, + 0.009657418355345726, + 0.015155056491494179, + 0.007067548111081123, + 0.018107449635863304, + 0.005121349822729826, + 0.0015865662135183811, + 0.0053331018425524235, + 0.0032896765042096376, + 0.0017307013040408492, + 0.0029670880176126957, + 0.0022389478981494904, + 0.006608197465538979, + 0.0047673750668764114, + 0.001100298948585987, + 0.003067181445658207, + 0.0012142130872234702, + 0.0012057070853188634, + 0.0017315418226644397, + 0.0012146478984504938, + 0.003739389358088374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.026331797242164612 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173256888985634 + ], + "opposite_logprob_tokens": [ + -4.0181732177734375 + ], + "logprob_reward": -0.018173256888985634, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018173256888985634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20510883629322052, + 0.10389362275600433, + 0.03761817887425423, + 0.09022577852010727, + 0.025795098394155502, + 0.1275547444820404, + 0.08742490410804749, + 0.11337103694677353, + 0.05246586352586746, + 0.1565418839454651 + ], + "credit_map": [ + 0.023126820102334023, + 0.011538340710103512, + 0.0687413215637207, + 0.00970090925693512, + 0.008326149545609951, + 0.012559964321553707, + 0.005846107844263315, + 0.06526923179626465, + 0.022143865004181862, + 0.008855657652020454, + 0.02032792568206787, + 0.010562464594841003, + 0.008852776139974594, + 0.008893908001482487, + 0.007536733523011208, + 0.01672029122710228, + 0.005839090328663588, + 0.002857827115803957, + 0.007386511191725731, + 0.0034698056988418102, + 0.0038675207179039717, + 0.003383463714271784, + 0.0032869766000658274, + 0.007526984438300133, + 0.005760201718658209, + 0.0020694034174084663, + 0.03709208592772484, + 0.0028484510257840157, + 0.0018563176272436976, + 0.003213709918782115, + 0.0020597740076482296, + 0.035325825214385986, + 0.004996218718588352, + 0.002082415856420994, + 0.0051367138512432575, + 0.0018823033897206187, + 0.0023310696706175804, + 0.0025170131120830774, + 0.0020862105302512646, + 0.004763152450323105, + 0.017074206843972206, + 0.009592550806701183, + 0.03140197694301605, + 0.007637239061295986, + 0.011374196968972683, + 0.013303084298968315, + 0.006566083058714867, + 0.03060539998114109, + 0.008469954133033752, + 0.0029144289437681437, + 0.03111460991203785, + 0.004153758753091097, + 0.0024981440510600805, + 0.0044202860444784164, + 0.0024338492657989264, + 0.03141987323760986, + 0.014319921843707561, + 0.00991829950362444, + 0.016253985464572906, + 0.015282992273569107, + 0.010860666632652283, + 0.016367396339774132, + 0.00991182029247284, + 0.02045595645904541, + 0.009504856541752815, + 0.003923059906810522, + 0.009681533090770245, + 0.0039204880595207214, + 0.002922417363151908, + 0.005181979853659868, + 0.0037436271086335182, + 0.01358790509402752, + 0.028533080592751503, + 0.009184504859149456, + 0.027828073129057884, + 0.014433170668780804, + 0.007627294398844242, + 0.019562123343348503, + 0.009602095000445843, + 0.039771560579538345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.018173256888985634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An underwater close-up, captured with a clean, cartoon-style high-frame-rate camera, presents a serene and captivating scene. The static camera focuses on a single, graceful stem of Mayaka fluviatilis rising from a soft, sandy lakebed. Its delicate, mossy leaves, rendered in vibrant shades of lime and emerald green with simple, clean outlines, gently sway in an invisible current. The plant's leaves exhibit an incredibly smooth, hypnotic undulation, rhythmically expanding and contracting beneath the glassy, calm surface of the water. Golden sunlight penetrates the crystal-clear water from above, casting ever-changing patterns of light and shadow (caustics) that dance across the plant and the lakebed. Light catches the edges of the soft leaves, making them appear to softly glow. The overall atmosphere is one of profound peace and tranquility, the plant's graceful, repetitive movement creating a deeply meditative and visually captivating experience.", + "seed": 21660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_3568_seed_21660_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": -0.008720632712356746, + "vlm_logprob": -0.008720632712356746, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.008720632712356746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007137869950383902 + ], + "opposite_logprob_tokens": [ + -6.007137775421143 + ], + "logprob_reward": -0.007137869950383902, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.007137869950383902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21561770141124725, + 0.15937775373458862, + 0.030644254758954048, + 0.03496728092432022, + 0.23918099701404572, + 0.03205690532922745, + 0.03358672931790352, + 0.021767834201455116, + 0.021733472123742104, + 0.2110670953989029 + ], + "credit_map": [ + 0.040711503475904465, + 0.018658272922039032, + 0.04032669961452484, + 0.009686560370028019, + 0.011201283894479275, + 0.0687115490436554, + 0.010903920978307724, + 0.015417920425534248, + 0.01347629725933075, + 0.01798998937010765, + 0.02695261500775814, + 0.021456250920891762, + 0.023928940296173096, + 0.01795729249715805, + 0.02413460612297058, + 0.013481761328876019, + 0.0021765532437711954, + 0.004098259843885899, + 0.003941940609365702, + 0.005383642390370369, + 0.0043365065939724445, + 0.00183388392906636, + 0.006617562845349312, + 0.0022559065837413073, + 0.004383409861475229, + 0.00463072769343853, + 0.007961748167872429, + 0.004253442864865065, + 0.0028383382596075535, + 0.0033149519003927708, + 0.004644240718334913, + 0.0029404216911643744, + 0.027571000158786774, + 0.0200494471937418, + 0.10091541707515717, + 0.01411717850714922, + 0.011209307238459587, + 0.036744192242622375, + 0.010988680645823479, + 0.017585767433047295, + 0.0037834590766578913, + 0.0036839263048022985, + 0.004299798980355263, + 0.00455969013273716, + 0.005285661201924086, + 0.00265347003005445, + 0.004822521470487118, + 0.0029683776665478945, + 0.007810558192431927, + 0.0046158526092767715, + 0.004616076592355967, + 0.0031756130047142506, + 0.00343733886256814, + 0.0032798168249428272, + 0.0029300767928361893, + 0.00372139411047101, + 0.0020404679235070944, + 0.0022640260867774487, + 0.004859764128923416, + 0.003040172392502427, + 0.002471467014402151, + 0.002162776654586196, + 0.0030612570699304342, + 0.0018679023487493396, + 0.001543398480862379, + 0.002626564586535096, + 0.003577047260478139, + 0.004654752556234598, + 0.002219547750428319, + 0.001538604497909546, + 0.0032216799445450306, + 0.002351876348257065, + 0.02412448078393936, + 0.020235490053892136, + 0.04367682710289955, + 0.03001396358013153, + 0.02937205322086811, + 0.016690252348780632, + 0.028006529435515404, + 0.01894748955965042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.007137869950383902 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001050635357387364 + ], + "opposite_logprob_tokens": [ + -6.8760504722595215 + ], + "logprob_reward": -0.001050635357387364, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001050635357387364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5511727929115295, + 0.05862151458859444, + 0.03360886126756668, + 0.02986997924745083, + 0.07079583406448364, + 0.04492655768990517, + 0.028387442231178284, + 0.016230570152401924, + 0.05990935489535332, + 0.10647711902856827 + ], + "credit_map": [ + 0.11545517295598984, + 0.054633814841508865, + 0.09590668976306915, + 0.021182212978601456, + 0.0282574575394392, + 0.17620089650154114, + 0.024800721555948257, + 0.034735821187496185, + 0.00984922144562006, + 0.0104169100522995, + 0.009568557143211365, + 0.006167789921164513, + 0.00656198849901557, + 0.0045648920349776745, + 0.006359474267810583, + 0.00513268169015646, + 0.0059115095064044, + 0.008016678504645824, + 0.0068168337456882, + 0.0025821139570325613, + 0.002444262383505702, + 0.002071387367323041, + 0.002708065090700984, + 0.0030580132734030485, + 0.003556026378646493, + 0.003826872445642948, + 0.004992631264030933, + 0.004496726673096418, + 0.002774067921563983, + 0.002479720627889037, + 0.003986210096627474, + 0.0037577212788164616, + 0.007338763680309057, + 0.005880285520106554, + 0.02902957610785961, + 0.004822606686502695, + 0.0035770866088569164, + 0.011195346713066101, + 0.0033059946727007627, + 0.005646176170557737, + 0.0028261158149689436, + 0.006417154800146818, + 0.0073916553519666195, + 0.006294648163020611, + 0.008924296125769615, + 0.0027007893659174442, + 0.007302273064851761, + 0.0030696268659085035, + 0.0036566199269145727, + 0.003881192998960614, + 0.004466618876904249, + 0.0038467692211270332, + 0.0032185260206460953, + 0.002377812284976244, + 0.00324439350515604, + 0.003695508698001504, + 0.002575731836259365, + 0.001639676745980978, + 0.0033458115067332983, + 0.0012544557685032487, + 0.0017455921042710543, + 0.00135534699074924, + 0.0018894594395533204, + 0.002424494829028845, + 0.00840678345412016, + 0.0070976936258375645, + 0.008906567469239235, + 0.004922040738165379, + 0.007287086918950081, + 0.011204518377780914, + 0.005853862967342138, + 0.006230799946933985, + 0.013294558972120285, + 0.007790395990014076, + 0.031710997223854065, + 0.011790809221565723, + 0.011865166947245598, + 0.009719299152493477, + 0.009920152835547924, + 0.010385747067630291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.001050635357387364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011803095694631338 + ], + "opposite_logprob_tokens": [ + -6.751180171966553 + ], + "logprob_reward": -0.0011803095694631338, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011803095694631338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6912223100662231, + 0.033714473247528076, + 0.08649958670139313, + 0.015324159525334835, + 0.04338594898581505, + 0.02187599055469036, + 0.034596942365169525, + 0.012716144323348999, + 0.021040111780166626, + 0.039624348282814026 + ], + "credit_map": [ + 0.14964240789413452, + 0.062108416110277176, + 0.1288384348154068, + 0.024151332676410675, + 0.028453508391976357, + 0.22778715193271637, + 0.027521682903170586, + 0.04271939396858215, + 0.0063945637084543705, + 0.004547801334410906, + 0.004921848885715008, + 0.0032665408216416836, + 0.0030221303459256887, + 0.0042625595815479755, + 0.0031972534488886595, + 0.004101776517927647, + 0.018712395802140236, + 0.02475355751812458, + 0.008725764229893684, + 0.00532979890704155, + 0.004609361756592989, + 0.009799194522202015, + 0.007323168218135834, + 0.007246345281600952, + 0.0021854627411812544, + 0.0016216288786381483, + 0.0023856391198933125, + 0.002155616180971265, + 0.0015643362421542406, + 0.0013411118416115642, + 0.00225561554543674, + 0.0018147504888474941, + 0.0053601632826030254, + 0.003947692923247814, + 0.016325227916240692, + 0.0025720938574522734, + 0.0028634739574044943, + 0.005358465015888214, + 0.003514755517244339, + 0.003444080241024494, + 0.001702694222331047, + 0.00277822301723063, + 0.0034377295523881912, + 0.0033288041595369577, + 0.004592732060700655, + 0.0013419537572190166, + 0.003471463918685913, + 0.0012223913799971342, + 0.00915211346000433, + 0.00442680437117815, + 0.005704592447727919, + 0.0028016383294016123, + 0.0026501528918743134, + 0.003844187129288912, + 0.002143066842108965, + 0.0038743880577385426, + 0.0020620126742869616, + 0.0016164507251232862, + 0.0016816726420074701, + 0.0015010136412456632, + 0.0009441880974918604, + 0.001382013433612883, + 0.001739908242598176, + 0.0017888847505673766, + 0.003214603988453746, + 0.0024494491517543793, + 0.0028800733853131533, + 0.0019957865588366985, + 0.002211214741691947, + 0.0034820172004401684, + 0.0022195964120328426, + 0.002587371040135622, + 0.005109024234116077, + 0.003198361722752452, + 0.010948964394629002, + 0.004471931606531143, + 0.004609735682606697, + 0.0032442482188344, + 0.004080189857631922, + 0.003961896989494562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0011803095694631338 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059335321187973 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059335321187973, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059335321187973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28147265315055847, + 0.05131705850362778, + 0.03206470608711243, + 0.06372594833374023, + 0.3184548616409302, + 0.037227753549814224, + 0.09405825287103653, + 0.017133140936493874, + 0.05241268500685692, + 0.052132975310087204 + ], + "credit_map": [ + 0.023973654955625534, + 0.030783288180828094, + 0.04316919296979904, + 0.04198697581887245, + 0.04630016162991524, + 0.036710888147354126, + 0.035640183836221695, + 0.0229082852602005, + 0.006443745922297239, + 0.006857362575829029, + 0.0105993477627635, + 0.004793145228177309, + 0.004078627564013004, + 0.006505425553768873, + 0.005884289741516113, + 0.006155111826956272, + 0.005332200322300196, + 0.00822726171463728, + 0.003425430040806532, + 0.002365630818530917, + 0.0028737538959831, + 0.0034258575178682804, + 0.0036484524607658386, + 0.0027661165222525597, + 0.005110842175781727, + 0.005758482031524181, + 0.008420589379966259, + 0.01061218697577715, + 0.008300689980387688, + 0.005364004988223314, + 0.013848358765244484, + 0.00631078751757741, + 0.03569721058011055, + 0.02715292200446129, + 0.12001509219408035, + 0.02465841919183731, + 0.021388627588748932, + 0.04285868629813194, + 0.025753064081072807, + 0.020930815488100052, + 0.002375230425968766, + 0.005359582602977753, + 0.004392449744045734, + 0.005949185695499182, + 0.008032339625060558, + 0.002047252142801881, + 0.006122277583926916, + 0.002949431771412492, + 0.008656039834022522, + 0.011464950628578663, + 0.014657340943813324, + 0.014699705876410007, + 0.01302437111735344, + 0.00694480212405324, + 0.013831389136612415, + 0.01077965646982193, + 0.0014237657887861133, + 0.0019812502432614565, + 0.003225387306883931, + 0.0019088267581537366, + 0.0024098879657685757, + 0.0018520663725212216, + 0.0020561693236231804, + 0.0022757847327739, + 0.008058170787990093, + 0.005531155504286289, + 0.008930552750825882, + 0.004544137045741081, + 0.005764915142208338, + 0.008415108546614647, + 0.004823680035769939, + 0.006344963330775499, + 0.004894668236374855, + 0.0044286418706178665, + 0.011233089491724968, + 0.006059244740754366, + 0.00660861236974597, + 0.0058067492209374905, + 0.006381868850439787, + 0.006720095407217741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.016059335321187973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817501336336136 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.01817501336336136, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817501336336136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2816939651966095, + 0.06756704300642014, + 0.05283995345234871, + 0.04737161472439766, + 0.3596605956554413, + 0.0497613325715065, + 0.025830447673797607, + 0.016763901337981224, + 0.05076197162270546, + 0.047749195247888565 + ], + "credit_map": [ + 0.037298064678907394, + 0.022494694218039513, + 0.06473065912723541, + 0.02049601450562477, + 0.0202438086271286, + 0.07782001793384552, + 0.017819801345467567, + 0.02079090289771557, + 0.011797674000263214, + 0.01135858055204153, + 0.012247751466929913, + 0.006956877652555704, + 0.0060541327111423016, + 0.005045249126851559, + 0.007264583371579647, + 0.006842194590717554, + 0.01192498579621315, + 0.014998143538832664, + 0.006797134410589933, + 0.0034217664506286383, + 0.002924293512478471, + 0.005103244911879301, + 0.003679161425679922, + 0.003991224337369204, + 0.007631835527718067, + 0.0076723238453269005, + 0.0103540550917387, + 0.004942200146615505, + 0.003888694103807211, + 0.00380579917691648, + 0.004118694923818111, + 0.004958013538271189, + 0.04197142273187637, + 0.031041622161865234, + 0.15221068263053894, + 0.020542219281196594, + 0.016166649758815765, + 0.05464780330657959, + 0.01606013998389244, + 0.027020029723644257, + 0.0029639743734151125, + 0.006595829967409372, + 0.0069324602372944355, + 0.00752383004873991, + 0.011028996668756008, + 0.002437512157484889, + 0.008917823433876038, + 0.003360908245667815, + 0.0033211198169738054, + 0.00277930055744946, + 0.003964096773415804, + 0.0035893376916646957, + 0.003833749797195196, + 0.003686902578920126, + 0.0020524223800748587, + 0.002603519707918167, + 0.0010388765949755907, + 0.001808499451726675, + 0.003001466393470764, + 0.0021694768220186234, + 0.0014860862866044044, + 0.001633216510526836, + 0.003814484691247344, + 0.0018117941217496991, + 0.007233723532408476, + 0.005605215672403574, + 0.006265512201935053, + 0.005538571625947952, + 0.005595293361693621, + 0.008180470205843449, + 0.005829464178532362, + 0.0065137166529893875, + 0.004841442219913006, + 0.004167742095887661, + 0.010646003298461437, + 0.007040945813059807, + 0.005806272849440575, + 0.00421942351385951, + 0.006108044646680355, + 0.004919324070215225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.01817501336336136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An underwater close-up, captured with a clean, cartoon-style high-frame-rate camera, presents a serene and captivating scene. The static camera focuses on a single, graceful stem of Mayaka fluviatilis rising from a soft, sandy lakebed. Its delicate, mossy leaves, rendered in vibrant shades of lime and emerald green with simple, clean outlines, gently sway in an invisible current. The plant's leaves exhibit an incredibly smooth, hypnotic undulation, rhythmically expanding and contracting beneath the glassy, calm surface of the water. Golden sunlight penetrates the crystal-clear water from above, casting ever-changing patterns of light and shadow (caustics) that dance across the plant and the lakebed. Light catches the edges of the soft leaves, making them appear to softly glow. The overall atmosphere is one of profound peace and tranquility, the plant's graceful, repetitive movement creating a deeply meditative and visually captivating experience.", + "seed": 21661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000160/group_9/index_3568_seed_21661_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": -0.012298199324868619, + "vlm_logprob": -0.012298199324868619, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.012298199324868619, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018520403653383255 + ], + "opposite_logprob_tokens": [ + -4.393520355224609 + ], + "logprob_reward": -0.018520403653383255, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.018520403653383255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14771108329296112, + 0.02246677316725254, + 0.14544595777988434, + 0.020968906581401825, + 0.46845608949661255, + 0.03086988255381584, + 0.10042458027601242, + 0.022364269942045212, + 0.01569988764822483, + 0.025592580437660217 + ], + "credit_map": [ + 0.02955276146531105, + 0.008658667095005512, + 0.02828689105808735, + 0.007367027923464775, + 0.007397433742880821, + 0.041362982243299484, + 0.009018998593091965, + 0.01606631651520729, + 0.0040707336738705635, + 0.0033685415983200073, + 0.0024709440767765045, + 0.0021990197710692883, + 0.0024086367338895798, + 0.0034747712779790163, + 0.0021314923651516438, + 0.00234263320453465, + 0.029445050284266472, + 0.01973017491400242, + 0.01359705999493599, + 0.01822330616414547, + 0.016115164384245872, + 0.027896154671907425, + 0.010433695279061794, + 0.010005351155996323, + 0.0028882648330181837, + 0.002529572695493698, + 0.0028311689384281635, + 0.0015237852931022644, + 0.0019259097753092647, + 0.0034189836587756872, + 0.0028761911671608686, + 0.0029750282410532236, + 0.13266249001026154, + 0.01281870435923338, + 0.12432298064231873, + 0.02126413770020008, + 0.008749634958803654, + 0.08673490583896637, + 0.01783975027501583, + 0.06406349688768387, + 0.004871855024248362, + 0.0026936407666653395, + 0.0038928152061998844, + 0.0019853669218719006, + 0.003216224955394864, + 0.005721973720937967, + 0.004396810196340084, + 0.004091194830834866, + 0.012505285441875458, + 0.010376968421041965, + 0.019697045907378197, + 0.009412750601768494, + 0.010573208332061768, + 0.009774886071681976, + 0.010375860147178173, + 0.01770857721567154, + 0.002600934822112322, + 0.004218406975269318, + 0.0016525713726878166, + 0.004044516943395138, + 0.0034494760911911726, + 0.002077673329040408, + 0.0023603057488799095, + 0.0019603834953159094, + 0.002835163613781333, + 0.001558929099701345, + 0.001775071257725358, + 0.0009507345384918153, + 0.0019316715188324451, + 0.0027103133033961058, + 0.001445825444534421, + 0.0024921775329858065, + 0.008216887712478638, + 0.001471764757297933, + 0.0012104982743039727, + 0.0014964250149205327, + 0.0016290866769850254, + 0.008741237223148346, + 0.0014138532569631934, + 0.0014128295006230474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.018520403653383255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015257871709764004 + ], + "opposite_logprob_tokens": [ + -6.50152587890625 + ], + "logprob_reward": -0.0015257871709764004, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015257871709764004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09731470793485641, + 0.03201122209429741, + 0.0633290484547615, + 0.018897676840424538, + 0.46235108375549316, + 0.23384146392345428, + 0.027023551985621452, + 0.017736036330461502, + 0.02802187204360962, + 0.01947333849966526 + ], + "credit_map": [ + 0.02781207300722599, + 0.005526138469576836, + 0.016568977385759354, + 0.003656963584944606, + 0.004807448945939541, + 0.023585811257362366, + 0.004777249880135059, + 0.010580044239759445, + 0.003862211015075445, + 0.005345507524907589, + 0.004521434661000967, + 0.002850581193342805, + 0.004280112683773041, + 0.0033802927937358618, + 0.0033604225609451532, + 0.0044106608256697655, + 0.009687056764960289, + 0.005809525027871132, + 0.011739565059542656, + 0.0046828920021653175, + 0.0059441691264510155, + 0.014601297676563263, + 0.00477633299306035, + 0.006088208872824907, + 0.002413647249341011, + 0.0032267472706735134, + 0.002079832134768367, + 0.0013309359783306718, + 0.002294336911290884, + 0.0028859234880656004, + 0.0016810387605801225, + 0.0029852152802050114, + 0.03749721497297287, + 0.01438320055603981, + 0.17080967128276825, + 0.02667575515806675, + 0.0143273388966918, + 0.09044542163610458, + 0.02728927880525589, + 0.08092320710420609, + 0.07707619667053223, + 0.01246094424277544, + 0.012723262421786785, + 0.008204005658626556, + 0.01058498490601778, + 0.0780862420797348, + 0.018911385908722878, + 0.01579442247748375, + 0.0033818306401371956, + 0.0036366404965519905, + 0.0037376938853412867, + 0.00257856328971684, + 0.0020954173523932695, + 0.0036165129859000444, + 0.003344557248055935, + 0.004632336087524891, + 0.0024517022538930178, + 0.0017270389944314957, + 0.002227064222097397, + 0.0013489655684679747, + 0.0016575554618611932, + 0.0018065499607473612, + 0.0038542738184332848, + 0.002662885235622525, + 0.002279846463352442, + 0.003855425398796797, + 0.002334275282919407, + 0.002497071633115411, + 0.00439497223123908, + 0.0025137451011687517, + 0.006806027144193649, + 0.0033405092544853687, + 0.002410068642348051, + 0.0018906762124970555, + 0.0022384508047252893, + 0.0017770490376278758, + 0.0034765563905239105, + 0.0024467557668685913, + 0.0029107488226145506, + 0.00232303305529058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0015257871709764004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009210634743794799 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009210634743794799, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009210634743794799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035305969417095184, + 0.016336748376488686, + 0.08590304851531982, + 0.00850622821599245, + 0.52781081199646, + 0.25643855333328247, + 0.028120746836066246, + 0.018523117527365685, + 0.016217119991779327, + 0.006837588269263506 + ], + "credit_map": [ + 0.004314423073083162, + 0.003796616103500128, + 0.007862535305321217, + 0.0026871247682720423, + 0.0034106080420315266, + 0.004746768623590469, + 0.003313961438834667, + 0.005173936020582914, + 0.002694898983463645, + 0.0017555772792547941, + 0.00271900394000113, + 0.0016493344446644187, + 0.0013657775707542896, + 0.002119980985298753, + 0.0015451802173629403, + 0.0024869965855032206, + 0.018934696912765503, + 0.010443414561450481, + 0.0067520602606236935, + 0.00875921081751585, + 0.008889791555702686, + 0.020580144599080086, + 0.005528370849788189, + 0.006015364546328783, + 0.0012359048705548048, + 0.0007815092103555799, + 0.0012910185614600778, + 0.0007801155443303287, + 0.0006019721040502191, + 0.0013063261285424232, + 0.0008785761892795563, + 0.0016308058984577656, + 0.15333561599254608, + 0.014016441069543362, + 0.13468441367149353, + 0.021808717399835587, + 0.010246510617434978, + 0.10451077669858932, + 0.018882181495428085, + 0.07032618671655655, + 0.09102316200733185, + 0.012548963539302349, + 0.018682386726140976, + 0.006411057431250811, + 0.008498996496200562, + 0.0901869535446167, + 0.013205543160438538, + 0.015881508588790894, + 0.0040395683608949184, + 0.0034824798349291086, + 0.004139266908168793, + 0.002828063443303108, + 0.002958576427772641, + 0.0032125392463058233, + 0.0030130527447909117, + 0.004447201732546091, + 0.002545425668358803, + 0.0026167172472923994, + 0.0019609821029007435, + 0.0022574516478925943, + 0.002683388302102685, + 0.0020782570354640484, + 0.002059944672510028, + 0.0023209531791508198, + 0.002220164518803358, + 0.0023091244511306286, + 0.0019068400142714381, + 0.0013946215622127056, + 0.0016282785218209028, + 0.0016823179321363568, + 0.0029249449726194143, + 0.002150829415768385, + 0.0007061351207084954, + 0.0008298491593450308, + 0.0007322636665776372, + 0.0007856063893996179, + 0.0012560837203636765, + 0.0009494586265645921, + 0.000786191551014781, + 0.0007920010248199105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.0009210634743794799 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337021961808205 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.026337021961808205, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337021961808205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18562458455562592, + 0.03198365867137909, + 0.04037361592054367, + 0.02743091620504856, + 0.4990125000476837, + 0.09860294312238693, + 0.043609362095594406, + 0.010184410959482193, + 0.04316738247871399, + 0.02001061849296093 + ], + "credit_map": [ + 0.03650834038853645, + 0.009624406695365906, + 0.039969153702259064, + 0.010426616296172142, + 0.006949917878955603, + 0.05322907119989395, + 0.009165795519948006, + 0.019751306623220444, + 0.0044233547523617744, + 0.0052794283255934715, + 0.004726188722997904, + 0.0024643971119076014, + 0.004237708635628223, + 0.0034818819258362055, + 0.0037541042547672987, + 0.003616597503423691, + 0.004385243635624647, + 0.005800486542284489, + 0.00723179429769516, + 0.005759420804679394, + 0.005478702951222658, + 0.004343475215137005, + 0.0041099186055362225, + 0.003264580387622118, + 0.0039038253016769886, + 0.004009651951491833, + 0.0035751007962971926, + 0.00233720731921494, + 0.0032354407012462616, + 0.00478295749053359, + 0.003292301669716835, + 0.002294431673362851, + 0.0855165347456932, + 0.015245785936713219, + 0.19575455784797668, + 0.029639074578881264, + 0.013602785766124725, + 0.033416569232940674, + 0.02947162836790085, + 0.09636559337377548, + 0.025763332843780518, + 0.005841078236699104, + 0.00827004574239254, + 0.004465023055672646, + 0.006004002410918474, + 0.02814764343202114, + 0.011361541226506233, + 0.008750290609896183, + 0.005010321736335754, + 0.005901927128434181, + 0.007537344470620155, + 0.0047470941208302975, + 0.005612242966890335, + 0.004369752481579781, + 0.004555816296488047, + 0.005874865222722292, + 0.0013290601782500744, + 0.0012366328155621886, + 0.0010483122896403074, + 0.0010079977801069617, + 0.001221303129568696, + 0.001112152705900371, + 0.001549256150610745, + 0.001679696375504136, + 0.004320717416703701, + 0.005922054406255484, + 0.003717019921168685, + 0.0035429648123681545, + 0.007264201063662767, + 0.005016484763473272, + 0.008755693212151527, + 0.004628250375390053, + 0.006786778103560209, + 0.0008284267387352884, + 0.0008489807951264083, + 0.0008006114512681961, + 0.0013521210057660937, + 0.007223609834909439, + 0.0010829109232872725, + 0.0010871824342757463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.026337021961808205 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater with a static camera focused on a single stem of Mayaka fluviatilis?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the plant's leaves rendered in vibrant shades of lime and emerald green with simple, clean outlines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do golden sunlight patterns (caustics) dance across the plant and the sandy lakebed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the leaves exhibit a smooth, rhythmic undulation of expanding and contracting motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the edges of the leaves appear to softly glow due to the light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186720363795757 + ], + "opposite_logprob_tokens": [ + -4.264186859130859 + ], + "logprob_reward": -0.014186720363795757, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186720363795757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714654803276062, + 0.03884506598114967, + 0.06535813212394714, + 0.010974914766848087, + 0.22798669338226318, + 0.314730703830719, + 0.030266372486948967, + 0.014167793095111847, + 0.009115752764046192, + 0.017089031636714935 + ], + "credit_map": [ + 0.055676162242889404, + 0.013289938680827618, + 0.05738645792007446, + 0.01498190313577652, + 0.00946581456810236, + 0.07846958935260773, + 0.0125160226598382, + 0.02967962995171547, + 0.005468911491334438, + 0.002981100929901004, + 0.006170888431370258, + 0.0034584151580929756, + 0.003690393641591072, + 0.006940423045307398, + 0.004078930243849754, + 0.006056004669517279, + 0.0070637790486216545, + 0.00926909502595663, + 0.01346649695187807, + 0.008625617250800133, + 0.007180715911090374, + 0.0076661244966089725, + 0.005562053993344307, + 0.006524260621517897, + 0.0008249482489190996, + 0.001206322223879397, + 0.0026219841092824936, + 0.0007079860079102218, + 0.0011010271264240146, + 0.0017837421037256718, + 0.0013182139955461025, + 0.0014106916496530175, + 0.05381617322564125, + 0.00824478454887867, + 0.07234141230583191, + 0.01278414111584425, + 0.006079722195863724, + 0.027643170207738876, + 0.011510686948895454, + 0.035566624253988266, + 0.09820210933685303, + 0.017071109265089035, + 0.02106570638716221, + 0.01228342391550541, + 0.014189064502716064, + 0.09814077615737915, + 0.028700238093733788, + 0.025078294798731804, + 0.004448864609003067, + 0.002711888402700424, + 0.005244819447398186, + 0.0025050209369510412, + 0.004051166586577892, + 0.0035882219672203064, + 0.0035092830657958984, + 0.004207110032439232, + 0.00218105036765337, + 0.0013727183686569333, + 0.002130244392901659, + 0.0010783959878608584, + 0.0012815529480576515, + 0.0021321061067283154, + 0.0018670454155653715, + 0.002124680671840906, + 0.0013734663370996714, + 0.0012545317877084017, + 0.0007874324801377952, + 0.0006780702387914062, + 0.0016619149828329682, + 0.001284757163375616, + 0.001086401753127575, + 0.0009891780791804194, + 0.0033176918514072895, + 0.0013928258558735251, + 0.0019673684146255255, + 0.0014279126189649105, + 0.0016952273435890675, + 0.0033015788067132235, + 0.0022178981453180313, + 0.0017685313941910863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.014186720363795757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file